Split (1)

train · 2.34k rows

title stringlengths 28 135	abstract stringlengths 0 12k	introduction stringlengths 0 12k
Wang_RODIN_A_Generative_Model_for_Sculpting_3D_Digital_Avatars_Using_CVPR_2023	Abstract This paper presents a 3D diffusion model that automat- ically generates 3D digital avatars represented as neural radiance fields (NeRFs). A significant challenge for 3D diffusion is that the memory and processing costs are pro- hibitive for producing high-quality results with rich details. To tackle this problem, we propose the roll-out diffusion net- work (RODIN), which takes a 3D NeRF model represented as multiple 2D feature maps and rolls out them onto a sin- gle 2D feature plane within which we perform 3D-aware diffusion. The RODIN model brings much-needed compu- tational efficiency while preserving the integrity of 3D dif- fusion by using 3D-aware convolution that attends to pro- jected features in the 2D plane according to their origi- nal relationships in 3D. We also use latent conditioning to orchestrate the feature generation with global coherence, leading to high-fidelity avatars and enabling semantic edit- ing based on text prompts. Finally, we use hierarchical syn- thesis to further enhance details. The 3D avatars generated by our model compare favorably with those produced by ex- isting techniques. We can generate highly detailed avatars with realistic hairstyles and facial hair. We also demon- strate 3D avatar generation from image or text, as well as text-guided editability. †Intern at Microsoft Research.‡Corresponding author.∗Equal contri- bution. Project Webpage: https://3d-avatar-diffusion.microsoft.com	1. Introduction Generative models [2, 34] are one of the most promising ways to analyze and synthesize visual data including 2D images and 3D models. At the forefront of generative mod- eling is the diffusion model [14, 24, 61], which has shown phenomenal generative power for images [19,47,50,52] and videos [23,59]. Indeed, we are witnessing a 2D content cre- ation revolution driven by the rapid advances of diffusion and generative modeling. In this paper, we aim to expand the applicability of diffusion to 3D digital avatars. We use “digital avatars” to refer to the traditional avatars manually created by 3D artists, as opposed to the recently emerging photorealistic avatars [8, 43]. The reason for focusing on digital avatars is twofold. On the one hand, digital avatars are widely used in movies, games, the metaverse, and the 3D industry in general. On the other hand, the available digital avatar data is very scarce as each avatar must be painstakingly created by a specialized 3D artist using a so- phisticated creation pipeline [20, 35], especially for mod- eling hair and facial hair. All this leads to a compelling scenario for generative modeling. We present a diffusion model for automatically produc- ing digital avatars represented as NeRFs [38], with each point describing its color radiance and density within the 3D volume. The core challenge in generating these avatars This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4563 is the prohibitive memory and computational cost required by high-quality avatars with rich details. Without rich de- tails, an avatar will always be somewhat “toy-like”. To tackle this challenge, we develop RODIN, the roll-out dif- fusion network. We take a NeRF model represented as mul- tiple 2D feature maps and roll out these maps onto a single 2D feature plane and perform 3D-aware diffusion within this plane. Specifically, we use the tri-plane representa- tion [9], which represents a volume by three orthogonal fea- ture planes. By simply rolling out the feature maps, RODIN can perform 3D-aware diffusion using an efficient 2D ar- chitecture and by drawing power from RODIN’s three key ingredients. The first is the 3D-aware convolution. The 2D CNN processing used in conventional 2D diffusion cannot ef- fectively handle feature maps originating from orthogonal planes. Rather than treating the features as plain 2D input, the 3D-aware convolution explicitly accounts for the fact that a 2D feature in one plane (of the tri-plane) is a projec- tion from a piece of 3D data and is hence intrinsically as- sociated with the same data’s projected features in the other two planes. To encourage cross-plane communication, we involve all these associated features in the convolution and synchronize their detail synthesis according to their 3D re- lationship. The second ingredient is latent conditioning. We use a latent vector to orchestrate the feature generation so that it is globally coherent in 3D, leading to better quality avatars and enabling semantic editing. We do this by using the avatars in the training dataset to train an additional image encoder which extracts a semantic latent vector serving as the condi- tional input to the diffusion model. This latent conditioning essentially acts as an autoencoder in orchestrating the fea- ture generation. For semantic editability, we adopt a frozen CLIP image encoder [46] that shares the latent space with text prompts. The final ingredient is hierarchical synthesis. We start by generating at low resolution ( 64×64), followed by a diffusion-based upsampling that yields a higher resolution (256×256). When training the diffusion upsampler, it is instrumental to penalize the image-level loss that we com- pute in a patch-wise manner. RODIN supports several application scenarios. We can use the model to generate an unlimited number of avatars from scratch, each different from the others as well as those in the training data. As shown in Figure 1, we can generate highly detailed avatars with realistic hairstyles and facial hair styled as beards, mustaches, goatees, and sideburns. Hairstyle and facial hair are essential parts of personal iden- tity yet have been notoriously difficult to model well with existing approaches. The RODIN model also allows avatar customization, with the resulting avatar capturing the visual characteristics of a person portrayed in an image or a textualdescription. Finally, our framework supports text-guided semantic editing. Our work shows that 3D diffusion holds great model- ing power, and this power can be effectively unleashed by rolling out the feature maps onto a 2D plane, leading to rich details, including those highly desirable but extremely dif- ficult to produce with existing techniques. It is worth not- ing that while this paper focuses on RODIN’s application to avatars, the design of RODIN is not avatar specific. Indeed, we believe RODIN is applicable to general 3D scenes.
Wu_Co-Salient_Object_Detection_With_Uncertainty-Aware_Group_Exchange-Masking_CVPR_2023	Abstract The traditional definition of co-salient object detection (CoSOD) task is to segment the common salient objects in a group of relevant images. Existing CoSOD models by- default adopt the group consensus assumption. This brings about model robustness defect under the condition of ir- relevant images in the testing image group, which hinders the use of CoSOD models in real-world applications. To address this issue, this paper presents a group exchange- masking (GEM) strategy for robust CoSOD model learn- ing. With two group of image containing different types of salient object as input, the GEM first selects a set of images from each group by the proposed learning based strategy, then these images are exchanged. The proposed feature ex- traction module considers both the uncertainty caused by the irrelevant images and group consensus in the remain- ing relevant images. We design a latent variable genera- tor branch which is made of conditional variational autoen- coder to generate uncertainly-based global stochastic fea- tures. A CoSOD transformer branch is devised to capture the correlation-based local features that contain the group consistency information. At last, the output of two branches are concatenated and fed into a transformer-based decoder, producing robust co-saliency prediction. Extensive evalua- tions on co-saliency detection with and without irrelevant images demonstrate the superiority of our method over a variety of state-of-the-art methods.	1. Introduction Co-salient object detection (CoSOD) is to segment the common salient objects in a group of relevant images. By detecting the co-salient object in a group of images, the images’ background and redundant content are re- ∗Corresponding author. This work is supported in part by Na- tional Key Research and Development Program of China under Grant No. 2018AAA0100400, in part by the NSFC under Grant Nos. 62276141, 61872189. Rearrange (b) Group exchange -masking Learn to select top-k noisy images RearrangeTrains Dogs Inputs GT OURS DCFM (a) Comparison results GCoNet Figure 1. (a) When there exists an irrelevant image in the test group, the state-of-the-art CoSOD models such as DCFM [39] and GCoNet [9] tend to generate false positive predictions for the noisy image, yet our method can achieve an accurate prediction due to the use of group exchange-masking strategy in (b). moved, which helps the downstream tasks such as ob- ject tracking [38], co-segmentation [48] and video co- localization [14], to name a few. Group consensus assumption is widely used by exist- ing CoSOD models, i.e., these models presume that all im- ages in the same group contain the common salient tar- gets. The widely-used CoSOD benchmark datasets, such as COCO-SEG [33], CoCA [50], and CoSOD3k [8] organize the training and testing images that contain the same ob- ject as a group. Many existing CoSOD models consider the group consensus characteristic in modeling. For example, early works such as [2, 10, 13] extract hand-crafted features for inter-image co-object correspondence discovery. The deep learning models proposed in [15, 20, 36, 39, 44, 46, 48] use one group of relevant images as training data input for consensus representation learning. Among them, a va- riety of novel model design techniques have been devel- oped to make full use of the group consistency characteris- tic, such as the low-rank feature learning [46], co-attention model [20] and intra-saliency correlation learning [15]. The issues of the group consensus assumption are partially stud- ied in recent literature [9], which further models the inter- 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19639 group separability for more discriminative feature learning by a group collaborating module. In this paper, we find that the group consensus assump- tion also restricts the CoSOD model’s robustness against the images without common object. As illustrated by Fig- ure 1(a), state-of-the-art CoSOD models tend to output false positive predictions for the irrelevant image. This issue hinders the use of CoSOD models in real-world applica- tions where the testing inputs are likely to contain irrele- vant images. To enhance the model’s robustness, we pro- pose a learning framework called group exchange-masking (GEM). The GEM is illustrated by Figure 1(b). Given two image groups that contain different types of co-salient ob- jects, we exchange several images between one group and the other. Those exchanged images are called noisy images. The number of noisy images is chosen to be less than the number of remaining relevant images in the group, so that the co-salient object in the noisy images forms a negative object but not the dominant co-salient object. The “mask- ing” strategy refers to the label regeneration of the noisy images. Because there is no co-salient object, in the regen- erated label, the original ground-truth object is masked. The learning objective is to correctly predict both the co-salient objects in the original relevant images and the added noisy images. Adding noisy images to the training image group brings about uncertainty to the CoSOD model learning since there is some probability of no expected common object in each image. We design a dual-path image feature extrac- tion module to model the group uncertainty in addition to the group consensus property. Specifically, we design a latent variable generator branch (LVGB) to extract the uncertainty-based global image features. The LVGB mod- ule is motivated by the conditional variational autoencoder (CV AE) [32] that is widely used to address the uncertainty in vision applications including image background model- ing [19], RGB-D saliency detection [43] and image recon- struction [41]. In parallel with LVGB, we feed the im- age group into a CoSOD Transformer Branch (CoSOD- TB). The CoSOD-TB partitions each image group into lo- cal patches, and the attention mechanism in the transformer enables this branch to model patch-wise correlation-based local features. As a result, the group consistency informa- tion can be captured by this branch. The outputs of the two branches are concatenated and fed into a transformer-based decoder for co-saliency prediction. The proposed model has the following technical contributions. • A robust CoSOD model learning mechanism, called group exchange-masking is proposed. By exchanging images between two groups, we augment the training data containing irrelevant images as noise to enhance model’s robustness. This is different from the tra- ditional CoSOD model learning frameworks that usegroups of relevant images as training data. • We propose a dual-path feature extraction module composed of the LVGB and the CoSOD-TB. The LVGB is designed to model the uncertainty of co- salient object existence. The CoSOD-TB is for the consensus feature extraction of the salient object in the relevant images. • Extensive evaluations on three benchmark datasets, including CoSal2015 [42], CoCA [15], and CoSOD3k [7] show that the superiority of the proposed model to the state-of-the-art methods in terms of all evaluation metrics. Besides, the proposed model demonstrates good robustness for dealing with noisy data without co-salient objects.
Wang_Gradient-Based_Uncertainty_Attribution_for_Explainable_Bayesian_Deep_Learning_CVPR_2023	Abstract Predictions made by deep learning models are prone to data perturbations, adversarial attacks, and out-of- distribution inputs. To build a trusted AI system, it is there- fore critical to accurately quantify the prediction uncertain- ties. While current efforts focus on improving uncertainty quantification accuracy and efficiency, there is a need to identify uncertainty sources and take actions to mitigate their effects on predictions. Therefore, we propose to de- velop explainable and actionable Bayesian deep learning methods to not only perform accurate uncertainty quan- tification but also explain the uncertainties, identify their sources, and propose strategies to mitigate the uncertainty impacts. Specifically, we introduce a gradient-based uncer- tainty attribution method to identify the most problematic regions of the input that contribute to the prediction uncer- tainty. Compared to existing methods, the proposed UA- Backprop has competitive accuracy, relaxed assumptions, and high efficiency. Moreover, we propose an uncertainty mitigation strategy that leverages the attribution results as attention to further improve the model performance. Both qualitative and quantitative evaluations are conducted to demonstrate the effectiveness of our proposed methods.	1. Introduction Despite significant progress in many fields, conventional deep learning models cannot effectively quantify their pre- diction uncertainties, resulting in overconfidence in un- known areas and the inability to detect attacks caused by data perturbations and out-of-distribution inputs. Left unad- dressed, this may cause disastrous consequences for safety- critical applications, and lead to untrustworthy AI models. The predictive uncertainty can be divided into epistemic uncertainty and aleatoric uncertainty [16]. Epistemic un-certainty reflects the model’s lack of knowledge about the input. High epistemic uncertainty arises in regions, where there are few or no observations. Aleatoric uncertainty mea- sures the inherent stochasticity in the data. Inputs with high noise are expected to have high aleatoric uncertainty. Con- ventional deep learning models, such as deterministic clas- sification models that output softmax probabilities, can only estimate the aleatoric uncertainty. Bayesian deep learning (BDL) offers a principled frame- work for estimating both aleatoric and epistemic uncertain- ties. Unlike the traditional point-estimated models, BDL constructs the posterior distribution of model parameters. By sampling predictions from various models derived from the parameter posterior, BDL avoids overfitting and al- lows for systematic quantification of predictive uncertain- ties. However, current BDL methods primarily concentrate on enhancing the accuracy and efficiency of uncertainty quantification, while failing to explicate the precise loca- tions of the input data that cause predictive uncertainties and take suitable measures to reduce the effects of uncertainties on model predictions. Uncertainty attribution (UA) aims to generate an uncer- tainty map of the input data to identify the most problem- atic regions that contribute to the prediction uncertainty. It evaluates the contribution of each pixel to the uncertainty, thereby increasing the transparency and interpretability of BDL models. Previous attribution methods are mainly de- veloped for classification attribution (CA) with determinis- tic neural networks (NNs) to find the contribution of im- age pixels to the classification score. Unlike UA, directly leveraging the gradient-based CA methods for detecting problematic regions is unreliable. While CA explains the model’s classification process, assuming its predictions are confident, UA intends to identify the sources of input im- perfections that contribute to the high predictive uncertain- ties. Moreover, CA methods are often class-discriminative This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12044 since the classification score depends on the predicted class. As a result, they often fail to explain the inputs which have wrong predictions
Wang_Images_Speak_in_Images_A_Generalist_Painter_for_In-Context_Visual_CVPR_2023	Abstract In-context learning, as a new paradigm in NLP , allows the model to rapidly adapt to various tasks with only a hand- ful of prompts and examples. But in computer vision, the difficulties for in-context learning lie in that tasks vary sig- nificantly in the output representations, thus it is unclear how to define the general-purpose task prompts that the vi- *Equal contribution. Correspondence to [email protected] . This work is done when Wen Wang is an intern at BAAI.sion model can understand and transfer to out-of-domain tasks. In this work, we present Painter, a generalist model which addresses these obstacles with an “image”-centric solution, that is, to redefine the output of core vision tasks as images, and specify task prompts as also images. With this idea, our training process is extremely simple, which performs standard masked image modeling on the stitch of input and output image pairs. This makes the model capable of performing tasks conditioned on visible image patches. Thus, during inference, we can adopt a pair of input and This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6830 output images from the same task as the input condition, to indicate which task to perform. Without bells and whistles, our generalist Painter can achieve competitive performance compared to well-established task-specific models, on seven representative vision tasks ranging from high-level visual understanding to low-level image processing. In addition, Painter significantly outperforms recent generalist models on several challenging tasks.	1. Introduction Training one generalist model that can execute diverse tasks simultaneously, and even can perform a new task given a prompt and very few examples, is one important step closer to artificial general intelligence. In NLP, the emergence of in-context learning [2, 7, 18] presents a new path in this direction, which uses language sequences as the general interface and allows the model to rapidly adapt to various language-centric tasks with only a handful of prompts and examples. Thus far, in computer vision, in-context learning is rarely explored and remains unclear how to achieve that. This can be attributed to the differences between two modalities. One difference is that NLP tasks mainly consist of language un- derstanding and generation, so their output spaces can be unified as sequences of discrete language tokens. But vision tasks are the abstractions of raw visual input to varied gran- ularity and angles. Thus, vision tasks vary significantly in output representations , leading to various task-specific loss functions and architecture designs. The second difference is that the output space of NLP tasks is even the same as the input. Thus, the task instruction and the example’s input/out- put, which are all language-token sequences, can be directly used as the input condition (also denoted as the task prompt), which can be processed straightforwardly by the large lan- guage model. However, in computer vision, it is unclear how to define general-purpose task prompts or instructions that the vision model can understand and transfer to out-of- domain tasks. Several recent attempts [6, 9, 10, 27, 35, 41] tackle these difficulties by following the solutions in NLP. They more-or-less convert vision problems into NLP ones via discretizing the continuous output spaces of vision tasks, and using the language or specially-designed discrete tokens as the task prompts. However, we believe that images speak in images, i.e., image itself is a natural interface for general-purpose visual perception. In this work, we address the above obstacles with a vision-centric solution. The core observation is that most dense-prediction vision problems can be formulated as image inpainting ,i.e.: Given an input image, prediction is to inpaint the desired but missing output “image”. Thus, we need a representation of 3-channel tensor thatappears as an “image” for the output of the vision tasks, and specify the task prompts using a pair of images. Here we showcase several representative vision tasks for train- ing, including depth estimation, human keypoint detection, semantic segmentation, instance segmentation, image de- noising, image deraining, and image enhancement, and unify their output spaces using a 3-channel tensor, a.k.a. “out- put image”. We carefully design the data format for each task, such as instance mask and per-pixel discrete labels of panoptic segmentation, per-pixel continuous values of depth estimation, and high-precision coordinates of pose estimation. Including more tasks is very straightforward, as we only need to construct new data pairs and add them to the training set, without modifications to either the model architecture or loss function. Based on this unification, we train a generalist Painter model with an extremely simple training process. During training, we stitch two images from the same task into a larger image, and so do their corresponding output images. Then we apply masked image modeling (MIM) on pixels of the output image, with the input image being the condition. With such a learning process, we enable the model to per- form tasks conditioned on visible image patches, that is, the capability of in-context prediction with the visual signal as context. Thus the trained model is capable of the in-context infer- ence. That is, we directly use the input/output paired images from the same task as the input condition to indicate which task to perform. Examples of in-context inference are illus- trated in Figure 1, consisting of seven in-domain examples (seven rows at top) and three out-of-domain examples (three rows at bottom). This definition of task prompts does not require deep understanding of language instructions as need by almost all previous approaches, and makes it very flexible for performing both in-domain and out-of-domain vision tasks. Without bells and whistles, our model can achieve com- petitive performance compared to well-established task- specific models, on several fundamental vision tasks across high-level visual understanding to low-level image process- ing, namely, depth estimation on NYUv2 [37], semantic segmentation on ADE-20K [54], human keypoint detection on COCO [31], panoptic segmentation on COCO [31], and three low-level image restoration tasks. Notably, on depth estimation of NYUv2, our model achieves state-of-the-art performance, outperforming previous best results by large margins which have heavy and specialized designs on archi- tectures and loss functions. Compared to other generalist models, Painter yields significant improvements on several challenging tasks. 6831
Wang_SunStage_Portrait_Reconstruction_and_Relighting_Using_the_Sun_as_a_CVPR_2023	Abstract A light stage uses a series of calibrated cameras and lights to capture a subject’s facial appearance under vary- ing illumination and viewpoint. This captured information is crucial for facial reconstruction and relighting. Unfortu- nately, light stages are often inaccessible: they are expen- sive and require significant technical expertise for construc- tion and operation. In this paper, we present SunStage: a lightweight alternative to a light stage that captures com- parable data using only a smartphone camera and the sun. Our method only requires the user to capture a selfie video outdoors, rotating in place, and uses the varying angles between the sun and the face as guidance in joint recon- struction of facial geometry, reflectance, camera pose, and lighting parameters. Despite the in-the-wild un-calibrated setting, our approach is able to reconstruct detailed facial appearance and geometry, enabling compelling effects such as relighting, novel view synthesis, and reflectance editing.	1. Introduction A light stage [11] acquires the shape and material prop- erties of a face in high detail using a series of images cap- tured under synchronized cameras and lights. This captured information can be used to synthesize novel images of the subject under arbitrary lighting conditions or from arbitrary viewpoints. This process enables a number of visual effects, such as creating digital replicas of actors that can be used in movies [1] or high-quality postproduction relighting [46]. In many cases, however, it is often infeasible to get ac- cess to a light stage for capturing a particular subject, be- cause light stages are not easy to find: they are expensive and require significant technical expertise (often teams of people) to build and operate. In these cases, hope is not lost — one can turn to methods that are trained on light stage data, with the intention of generalizing to new sub- jects. These methods do not require the subject to be cap- tured by a light stage but instead use a machine learning This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20792 model trained on a collection of previously acquired light stage captures to enable the same applications as a light stage, but from only one or several images of a new sub- ject [6, 25, 30,38,40,50,52]. Unfortunately, these methods have difficulty faithfully reproducing and editing the ap- pearance of new subjects, as they lack much of the signal necessary to resolve the ambiguities of single-view recon- struction, i.e., a single image of a face can be reasonably explained by different combinations of geometry, illumina- tion, and reflectance. In this paper, we propose an intermediate solution — one that allows for personalized, high-quality capture of a given subject, but without the need for expensive, calibrated cap- ture equipment. Our method, which we dub SunStage, uses only a handheld smartphone camera and the sun to simu- late a minimalist light stage, enabling the reconstruction of individually-tailored geometry and reflectance without spe- cialized equipment. Our capture setup only requires the user to hold the camera at arm’s length and rotate in place, al- lowing the face to be observed under varying angles of inci- dent sunlight, which causes specular highlights to move and shadows to swing across the face. This provides strong sig- nals for the reconstruction of facial geometry and spatially- varying reflectance properties. The reconstructed face and scene parameters estimated by our system can be used to realistically render the subject in new, unseen lighting con- ditions — even with complex details like self-occluding cast shadows, which are typically missing in purely image-based relighting techniques, i.e., those that do not explicitly model geometry. In addition to relighting, we also show applica- tions in view synthesis, correcting facial perspective distor- tion, and editing skin reflectance. Our contributions include: (1) a novel capture technique for personalized facial scanning without custom equip- ment, (2) a system for optimization and disentanglement of scene parameters (geometry, materials, lighting, and camera poses) from an unaligned, handheld video, and (3) multiple portrait editing applications that produce photorealistic re- sults, using as input only a single selfie video.
Xie_GP-VTON_Towards_General_Purpose_Virtual_Try-On_via_Collaborative_Local-Flow_Global-Parsing_CVPR_2023	Abstract Image-based Virtual Try-ON aims to transfer an in-shop garment onto a specific person. Existing methods employ a global warping module to model the anisotropic defor- mation for different garment parts, which fails to preserve the semantic information of different parts when receiving challenging inputs (e.g, intricate human poses, difficult gar- ments). Moreover, most of them directly warp the input garment to align with the boundary of the preserved re- gion, which usually requires texture squeezing to meet the boundary shape constraint and thus leads to texture dis- tortion. The above inferior performance hinders existing methods from real-world applications. To address these problems and take a step towards real-world virtual try-on, we propose a General-Purpose Virtual Try-ON framework, named GP-VTON, by developing an innovative Local-Flow Global-Parsing (LFGP) warping module and a DynamicGradient Truncation (DGT) training strategy. Specifically, compared with the previous global warping mechanism, LFGP employs local flows to warp garments parts individ- ually, and assembles the local warped results via the global garment parsing, resulting in reasonable warped parts and a semantic-correct intact garment even with challenging in- puts.On the other hand, our DGT training strategy dynam- ically truncates the gradient in the overlap area and the warped garment is no more required to meet the bound- ary constraint, which effectively avoids the texture squeez- ing problem. Furthermore, our GP-VTON can be easily extended to multi-category scenario and jointly trained by using data from different garment categories. Extensive ex- periments on two high-resolution benchmarks demonstrate our superiority over the existing state-of-the-art methods.1 1Corresponding author is Xiaodan Liang. Code is available at gp-vton. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23550	1. Introduction The problem of Virtual Try-ON (VTON), aiming to transfer a garment onto a specific person, is of particu- lar importance for the nowadays e-commerce and the fu- ture metaverse. Compared with the 3D-based solutions [2, 14, 16, 28, 34] which rely upon 3D scanning equipment or labor-intensive 3D annotations, the 2D image-based meth- ods [1, 9, 12, 17, 19, 20, 29, 30, 32, 36, 38–40, 43], which directly manipulate on the images, are more practical for the real world scenarios and thus have been intensively ex- plored in the past few years. Although the pioneering 2D image-based VTON meth- ods [12, 19, 29] can synthesize compelling results on the widely used benchmarks [6, 8, 18], there still exist some deficiencies preventing them from the real-world scenarios, which we argue mainly contain three-folds. First, existing methods have strict constraints on the input images, and are prone to generate artifacts when receiving challenging in- puts. To be specific, as shown in the 1st row of Fig. 1(A), when the pose of the input person is intricate, existing methods [12, 19] fail to preserve the semantic information of different garment parts, resulting in the indistinguish- able warped sleeves. Besides, as shown in the 2nd row of Fig. 1(A), if the input garment is a long sleeve without ob- vious seam between the sleeve and torso, existing methods will generate adhesive artifact between the sleeve and torso. Second, most of the existing methods directly squeeze the input garment to make it align with the preserved region, leading to the distorted texture around the preserved region (e.g., the 3rd row of Fig. 1(A)). Third, most of the exist- ing works only focus on the upper-body try-on and neglect other garment categories (i.e, lower-body, dresses), which further limits their scalability for real-world scenarios. To relieve the input constraint for VTON systems and fully exploit their application potential, in this paper, we take a step forwards and propose a unified framework, named GP-VTON, for the General- Purposed Virtual Try- ON, which can generate realistic try-on results even for the challenging scenario (Fig. 1(A)) (e.g., intricate human poses, difficult garment inputs, etc.), and can be easily ex- tended to the multi-category scenario (Fig. 1(B)). The innovations of our GP-VTON lie in a novel Local- Flow Global-Parsing (LFGP) warping module and a Dy- namic Gradient Truncation (DGT) training strategy for the warping network, which enable the network to generate high fidelity deformed garments, and further facilitate our GP-VTON to generate photo-realistic try-on results. Specifically, most of the existing methods employ neural network to model garment deformation by introducing the Thin Plate Splines (TPS) transformation [3] or the appear- ance flow [44] into the network, and training the network in a weakly supervised manner (i.e., without ground truth for the deformation function). However, both of the TPS-based methods [6, 18, 30, 36, 40] and the flow-based meth- ods [1, 12, 17, 19, 29] directly learn a global deformation field, therefore fail to represent complicated non-rigid gar- ment deformation that requires diverse transformation for different garment parts. Taking the intricate pose case in Fig. 1(A) as an example, existing methods [12, 19] can not simultaneously guarantee accurate deformation for the torso region and sleeve region, and lead to exceeding distorted sleeves. In contrast, our LFGP warping module chooses to learn diverse local deformation fields for different garment parts, which is capable of individually warping each gar- ment part, and generating semantic-correct warped garment even for intricate pose case. Besides, since each local defor- mation field merely affects one corresponding garment part, garment texture from other parts is agnostic to the current deformation field and will not appear in the current local warped result. Therefore, the garment adhesion problem in the complex garment scenario can be completely addressed (as demonstrated in the 2nd row of Fig. 1(A)). However, directly assembling the local warped parts together can not obtain realistic warped garments, because there would be overlap among different warped parts. To deal with this, our LFGP warping module collaboratively estimates a global garment parsing to fuse different local warped parts, result- ing in a complete and unambiguous warped garment. On the other hand, the warping network in existing meth- ods [12,19,29] takes as inputs the flat garment and the mask of the preserved region (i.e., region to be preserved during the try-on procedure, such as the lower garment for upper- body VTON), and force the input garment to align with the boundary of the preserved region (e.g., the junction of the upper and lower garment), which usually require garment squeezing to meet the shape constraint and lead to texture distortion around the garment junction (please refer to the 3rd row of Fig. 1(A)). An effective solution to this problem is exploiting the gradient truncation strategy for the network training, in which the warped garment will be processed by the preserved mask before calculating the warping loss and the gradient in the preserved region will not be back- propagated. By using such a strategy, the warped garment is no longer required to strictly align with the preserved boundary, which largely avoids the garment squeezing and texture distortion. However, due to the poor supervision of warped garments in the preserved region, directly em- ploying the gradient truncation for all training data will lead to excessive freedom for the deformation field, which usu- ally results in texture stretching in the warped results. To tackle this problem, we proposed a Dynamic Gradient Trun- cation (DGT) training strategy which dynamically conducts gradient truncation for different training samples accord- ing to the disparity of height-width ratio between the flat garment and the warped garment. By introducing the dy- namic mechanism, our LFGP warping module can alleviate 23551 LFGP Warping Module Warp �� ’��’ �� ’��’ {��}�=13{��}�=13{�’�}�=13 Skin Color Map �� Figure 2. Overview of GP-VTON. The LFGP warping module aims to estimate the local flows {fk}3 k=1and global garment parsing S′, which is used to warp different garment parts {Gk}3 k=1and assembles warped parts {G′k}3 k=1into the intact garment G′, respectively. The generator Gtakes as inputs G′and the other person-related conditions to generate the final try-on result I′. the texture stretching problem and obtain realistic warped garments with better texture preservation. Overall, our contributions can be summarized as fol- lows: (1) We propose a unified try-on framework, named GP-VTON, to generate photo-realistic results for diverse scenarios. (2) We propose a novel LFGP warping mod- ule to generate semantic-correct deformed garments even with challenging inputs. (3) We introduce a simple, yet ef- fective DGT training strategy for the warping network to obtain distortion-free deformed garments. (4) Extensive ex- periments on two challenging high-resolution benchmarks show the superiority of GP-VTON over existing SOTAs.
Xu_Grid-Guided_Neural_Radiance_Fields_for_Large_Urban_Scenes_CVPR_2023	Abstract Purely MLP-based neural radiance fields (NeRF-based methods) often suffer from underfitting with blurred ren- derings on large-scale scenes due to limited model capac- ity. Recent approaches propose to geographically divide the scene and adopt multiple sub-NeRFs to model each region individually, leading to linear scale-up in training costs and the number of sub-NeRFs as the scene expands. An alterna- tive solution is to use a feature grid representation, which is computationally efficient and can naturally scale to a large scene with increased grid resolutions. However, the feature grid tends to be less constrained and often reaches subop- timal solutions, producing noisy artifacts in renderings, es- pecially in regions with complex geometry and texture. In this work, we present a new framework that realizes high- fidelity rendering on large urban scenes while being com- putationally efficient. We propose to use a compact multi- resolution ground feature plane representation to coarsely capture the scene, and complement it with positional encod- ing inputs through another NeRF branch for rendering in a joint learning fashion. We show that such an integrationcan utilize the advantages of two alternative solutions: a light-weighted NeRF is sufficient, under the guidance of the feature grid representation, to render photorealistic novel views with fine details; and the jointly optimized ground fea- ture planes, can meanwhile gain further refinements, form- ing a more accurate and compact feature space and output much more natural rendering results.	1. Introduction Large urban scene modeling has been drawing lots of research attention with the recent emergence of neural radi- ance fields (NeRF) due to its photorealistic rendering and model compactness [3, 34, 56, 59, 62, 64]. Such model- ing can enable a variety of practical applications, including autonomous vehicle simulation [23, 39, 67], aerial survey- ing [6,15], and embodied AI [35,61]. NeRF-based methods have shown impressive results on object-level scenes with their continuity prior benefited from the MLP architecture andhigh-frequency details with the globally shared posi- tional encodings. However, they often fail to model large and complex scenes. These methods suffer from underfit- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8296 ting due to limited model capacity and only produce blurred renderings without fine details [56,62,64]. BlockNeRF [58] and MegaNeRF [62] propose to geographically divide ur- ban scenes and assign each region a different sub-NeRF to learn in parallel. Subsequently, when the scale and com- plexity of the target scene increases, they inevitably suffer from a trade-off between the number of sub-NeRFs and the capacity required by each sub-NeRF to fully capture all the fine details in each region. Another stream of grid-based representations represents the target scene using a grid of features [27, 27, 30, 36, 55, 68, 69]. These methods are gen- erally much faster during rendering and more efficient when the scene scales up. However, as each cell of the feature grid is individually optimized in a locally encoded manner, the resulting feature grids tend to be less continuous across the scene compared to NeRF-based methods. Although more flexibility is intuitively beneficial for capturing fine details, the lack of inherent continuity makes this representation vulnerable to suboptimal solutions with noisy artifacts, as demonstrated in Fig. 1. To effectively reconstruct large urban scenes with im- plicit neural representations, in this work, we propose a two- branch model architecture that takes a unified scene repre- sentation that integrates both grid-based andNeRF-based approaches under a joint learning scheme. Our key insight is that these two types of representations can be used com- plementary to each other: while feature grids can easily fit local scene content with explicit and independently learned features, NeRF introduces an inherent global continuity on the learned scene content with its shareable MLP weights across all 3D coordinate inputs. NeRF can also encourage capturing high-frequency scene details by matching the po- sitional encodings as Fourier features with the bandwidth of details. However, unlike feature grid representation, NeRF is less effective in compacting large scene contents into its globally shared latent coordinate space. Concretely, we firstly model the target scene with a fea- ture grid in a pre-train stage, which coarsely captures scene geometry and appearance. The coarse feature grid is then used to 1) guide NeRF’s point sampling to let it concentrate around the scene surface; and 2) supply NeRF’s positional encodings with extra features about the scene geometry and appearance at sampled positions. Under such guidance, NeRF can effectively and efficiently pick up finer details in a drastically compressed sampling space. Moreover, as coarse-level geometry and appearance information are ex- plicitly provided to NeRF, a light-weight MLP is sufficient to learn a mapping from global coordinates to volume den- sities and color values. The coarse feature grids get further optimized with gradients from the NeRF branch in the sec- ond joint-learning stage, which regularizes them to produce more accurate and natural rendering results when applied in isolation. To further reduce memory footprint and learn areliable feature grid for large urban scenes, we adopt a com- pact factorization of the 3D feature grid to approximate it without losing representation capacity. Based on the obser- vation that essential semantics such as the urban layouts are mainly distributed on the ground ( i.e.,xy-plane), we pro- pose to factorize the 3D feature grid into 2D ground feature planes spanning the scene and a vertically shared feature vector along the z-axis. The benefits are manifold: 1) The memory is reduced from O(N3)toO(N2). 2) The learned feature grid is enforced to be disentangled into highly com- pact ground feature plans, offering explicit and informative scene layouts. Extensive experiments show the effective- ness of our unified model and scene representation. When rendering novel views in practice, users are allowed to use either the grid branch at a faster rendering speed, or the NeRF branch, with more high-frequency details and spatial smoothness, yet at the cost of a relatively slower rendering.
Wang_Scene-Aware_Egocentric_3D_Human_Pose_Estimation_CVPR_2023	Abstract Egocentric 3D human pose estimation with a single head-mounted fisheye camera has recently attracted atten- tion due to its numerous applications in virtual and aug- mented reality. Existing methods still struggle in challeng- ing poses where the human body is highly occluded or is closely interacting with the scene. To address this issue, we propose a scene-aware egocentric pose estimation method that guides the prediction of the egocentric pose with scene constraints. To this end, we propose an egocentric depth estimation network to predict the scene depth map from a wide-view egocentric fisheye camera while mitigating the occlusion of the human body with a depth-inpainting net- work. Next, we propose a scene-aware pose estimation network that projects the 2D image features and estimated depth map of the scene into a voxel space and regresses the 3D pose with a V2V network. The voxel-based fea- ture representation provides the direct geometric connec- tion between 2D image features and scene geometry, and further facilitates the V2V network to constrain the pre- dicted pose based on the estimated scene geometry. To en- able the training of the aforementioned networks, we also generated a synthetic dataset, called EgoGTA, and an in- the-wild dataset based on EgoPW, called EgoPW-Scene. The experimental results of our new evaluation sequences show that the predicted 3D egocentric poses are accurate and physically plausible in terms of human-scene interac- tion, demonstrating that our method outperforms the state- of-the-art methods both quantitatively and qualitatively.	1. Introduction Egocentric 3D human pose estimation with head- or body-mounted cameras is extensively researched recently because it allows capturing the person moving around in a large space, while the traditional pose estimation methods can only record in a fixed volume. With this advantage, the egocentric pose estimation methods show great potential in various applications, including the xR technologies and mo- Image EgoPW OursFigure 1. Previous egocentric pose estimation methods like EgoPW predict body poses that may suffer from body floating is- sue (the first row) or body-environment penetration issue (the sec- ond row). Our method predicts accurate and plausible poses com- plying with the scene constraints. The red skeletons are the ground truth poses and the green skeletons are the predicted poses. bile interaction applications. In this work, we estimated the full 3D body pose from a single head-mounted fisheye camera. A number of works have been proposed, including Mo2Cap2[39], xR- egopose [32], Global-EgoMocap [36], and EgoPW [35]. These methods have made significant progress in estimat- ing egocentric poses. However, when taking account of the interaction between the human body and the surrounding environment, they still suffer from artifacts that contrast the physics plausibility, including body-environment penetra- tions or body floating (see the EgoPW results in Fig. 1), which is mostly ascribed to the ambiguity caused by the self-occluded and highly distorted human body in the ego- centric view. This problem will render restrictions on sub- sequent applications including action recognition, human- object interaction recognition, and motion forecasting. To address this issue, we propose a scene-aware pose estimation framework that leverages the scene context to constrain the prediction of an egocentric pose. This frame- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13031 work produces accurate and physically plausible 3D human body poses from a single egocentric image, as illustrated in Fig. 1. Thanks to the wide-view fisheye camera mounted on the head, the scene context can be easily obtained even with only one egocentric image. To this end, we train an egocentric depth estimator to predict the depth map of the surrounding scene. In order to mitigate the occlusion caused by the human body, we predict the depth map including the visible human and leverage a depth-inpainting network to recover the depth behind the human body. Next, we combine the projected 2D pose features and scene depth in a common voxel space and regress the 3D body pose heatmaps with a V2V network [22]. The 3D voxel representation projects the 2D poses and depth in- formation from the distorted fisheye camera space to the canonical space, and further provides direct geometric con- nection between 2D image features and 3D scene geome- try. This aggregation of 2D image features and 3D scene geometry facilitates the V2V network to learn the rela- tive position and potential interactions between the human body joints and the surrounding environment and further enables the prediction of plausible poses under the scene constraints. Since no available dataset can be used for train these net- works, we proposed EgoGTA, a synthetic dataset based on the motion sequences of GTA-IM [3], and EgoPW-Scene, an in-the-wild dataset based on EgoPW [35]. Both of the datasets contain body pose labels and scene depth map la- bels for each egocentric frame. To better evaluate the relationship between estimated egocentric pose and scene geometry, we collected a new test dataset containing ground truth joint positions in the egocentric view. The evaluation results on the new dataset, along with results on datasets in Wang et al . [36] and Mo2Cap2[39] demonstrate that our method significantly outperforms existing methods both quantitatively and qual- itatively. We also qualitatively evaluate our method on in- the-wild images. The predicted 3D poses are accurate and plausible even in challenging real-world scenes. To summa- rize, our contributions are listed as follows: • The first scene-aware egocentric human pose estima- tion framework that predicts accurate and plausible egocentric pose with the awareness of scene context; • Synthetic and in-the-wild egocentric datasets contain- ing egocentric pose labels and scene geometry labels;1 • A new depth estimation and inpainting networks to predict the scene depth map behind the human body; • By leveraging a voxel-based representation of body pose features and scene geometry jointly, our method 1Datasets are released in our project page. Meta did not access or pro- cess the data and is not involved in the dataset release.outperforms the previous approaches and generates plausible poses considering the scene context.
Wu_Spatiotemporal_Self-Supervised_Learning_for_Point_Clouds_in_the_Wild_CVPR_2023	Abstract Self-supervised learning (SSL) has the potential to ben- efit many applications, particularly those where manually annotating data is cumbersome. One such situation is the semantic segmentation of point clouds. In this context, ex- isting methods employ contrastive learning strategies and define positive pairs by performing various augmentation of point clusters in a single frame. As such, these meth- ods do not exploit the temporal nature of LiDAR data. In this paper, we introduce an SSL strategy that leverages pos- itive pairs in both the spatial and temporal domain. To this end, we design (i) a point-to-cluster learning strategy that aggregates spatial information to distinguish objects; and (ii) a cluster-to-cluster learning strategy based on unsu- pervised object tracking that exploits temporal correspon- dences. We demonstrate the benefits of our approach via extensive experiments performed by self-supervised train- ing on two large-scale LiDAR datasets and transferring the resulting models to other point cloud segmentation bench- marks. Our results evidence that our method outperforms the state-of-the-art point cloud SSL methods.1	1. Introduction Semantic segmentation from LiDAR point clouds can be highly beneficial in practical applications, e.g., for self- driving vehicles to safely interact with their surroundings. Nowadays, state-of-the-art methods [13, 36, 46] achieve this with deep neural networks. While effective, the training of such semantic segmentation networks requires large amounts of annotated data, which is prohibitively costly to acquire, particularly for point-level LiDAR anno- tations [45]. By contrast, with the rapid proliferation of self- driving vehicles, large amounts of unlabeled LiDAR data are generated. Here, we develop a method to exploit such unlabeled data in a self-supervised learning framework. Self-supervised learning (SSL) aims to learn features without any human annotations [1, 2, 22, 26, 33, 35, 40, 45] but so that they can be effectively used for fine-tuning on a 1Our code and pretrained models will be found at https://github.com/YanhaoWu/STSSL. Correspondence to Ke Wei. Temporal𝑃𝑖1𝑃𝑖2 𝜏1Previous Methods Our Method 𝑃𝑚1𝑃𝑛1 𝑃𝑚2𝑃𝑛2𝜏2 𝜏1 𝜏2𝜏1 𝜏2 :Spatial features aggregation :Temporal features aggregation :Point -wise feature :Cluster -wise featureSpatial Spatial Spatial Temporal :Cluster/Scene -wise featureFigure 1. Our method vs existing ones. (Top) Previous methods create positive pairs for SSL by applying different augmentations, τ1andτ2(e.g., random flipping, clipping), to a single frame. (Bot- tom) By contrast, we leverage both spatial and temporal informa- tion via a point-to-cluster and an inter-frame SSL strategy. Points in the same color are from the same cluster in the latent space. downstream task with a small number of labeled samples. This is achieved by defining a pre-task that does not re- quire annotations. While many pre-tasks have been pro- posed [27], contrastive learning has nowadays become a highly popular choice [30,33,40,41,45]. In general, it aims to maximize the similarity of positive pairs while potentially minimizing that of negative ones. In this context, most of This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5251 Figure 2. Cars in the same frame but under different illumina- tion angles . Note that the main source of difference between the two instance point clouds arises from the different illumination an- gles. the point cloud SSL literature focuses on indoor scenes, for which relatively dense point clouds are available. Unfor- tunately, for outdoor scenes, such as the ones we consider here, the data is more complex and much sparser, and cre- ating effective pairs remains a challenge. Several approaches [33, 45] have nonetheless been pro- posed to perform SSL on outdoor LiDAR point cloud data. As illustrated in the top portion of Fig. 1, they construct positive pairs of point clusters or scenes by applying aug- mentations to a single frame. As such, they neglect the tem- poral information of the LiDAR data. By contrast, in this paper, we introduce an SSL approach to LiDAR point cloud segmentation based on extracting effective positive pairs in both the spatial and temporal domain. To achieve this without requiring any pose sensor as in [24, 40], we introduce (i) a point-to-cluster (P2C) SSL strategy that maximizes the similarity between the features encoding a cluster and those of its individual points, thus encouraging the points belonging to the same object to be close in feature space; (ii) a cluster-level inter-frame self- supervised learning strategy that tracks an object across consecutive frames in an unsupervised manner and encour- ages feature similarity between the different frames. These two strategies are depicted in the bottom portion of Fig. 1. Note that the illumination angle of one object seen in two different frames typically differs. As shown in Fig. 2, this is also the main source of difference between two objects of the same class in the same frame. Therefore, our inter- frame SSL strategy lets us encode not only temporal infor- mation, but also the fact that points from different objects from the same class should be close to each other in feature space. As simulating different illumination angles via data augmentation is challenging, our approach yields positive pairs that better reflects the intra-class variations in LiDAR point clouds than existing single-frame methods [33, 45]. Our contribution can be summarized as follows:• We introduce an SSL strategy for point cloud segmen- tation based only on positive pairs. It does not require any external information, such as pose, GPS, and IMU. • We propose a novel Point-to-Cluster (P2C) training paradigm that combines the advantages of point-level and cluster-level representations to learn a structured point-level embedding space. • We introduce the use of cluster-level inter-frame self- supervised leaning on point clouds generated by a Li- DAR sensor, which introduces a new way to integrate temporal information into SSL. Our experiments on several datasets, including KITTI [17], nuScene [5], SemanticKITTI [4] and SemanticPOSS [34], evidence that our method outperforms the state-of-the-art SSL techniques for point cloud data.
VS_Instance_Relation_Graph_Guided_Source-Free_Domain_Adaptive_Object_Detection_CVPR_2023	Abstract Unsupervised Domain Adaptation (UDA) is an effective approach to tackle the issue of domain shift. Speciﬁcally, UDA methods try to align the source and target representa- tions to improve generalization on the target domain. Fur- ther, UDA methods work under the assumption that the source data is accessible during the adaptation process. However, in real-world scenarios, the labelled source data is often restricted due to privacy regulations, data transmis- sion constraints, or proprietary data concerns. The Source- Free Domain Adaptation (SFDA) setting aims to alleviate these concerns by adapting a source-trained model for the target domain without requiring access to the source data. In this paper, we explore the SFDA setting for the task of adaptive object detection. To this end, we propose a novel training strategy for adapting a source-trained object de- tector to the target domain without source data. More pre- cisely, we design a novel contrastive loss to enhance the target representations by exploiting the objects relations for a given target domain input. These object instance rela- tions are modelled using an Instance Relation Graph (IRG) network, which are then used to guide the contrastive repre- sentation learning. In addition, we utilize a student-teacher to effectively distill knowledge from source-trained model to target domain. Extensive experiments on multiple ob- ject detection benchmark datasets show that the proposed approach is able to efﬁciently adapt source-trained object detectors to the target domain, outperforming state-of-the- art domain adaptive detection methods. Code and models are provided in https://viudomain.github.io/irg-sfda-web/ .	1. Introduction In recent years, object detection has seen tremendous advancements due to the rise of deep networks [ 12,42, 44,45,53,79]. The major contributor to this success is the availability of large-scale annotated detection datasets [10,13,15,43,73], as it enables the supervised training of deep object detector models. However, these models Instance Relation Graph Labeled source domainUnlabeled target domain GraphCon Loss Pseudo-label Loss Source-free domain adaptationSource model trainingFigure 1. Left: Supervised training of detection model on the source domain. Right: Source-Free Domain Adaptation where a source-trained model is adapted to the target domain in the ab- sence of source data with pseudo-label self-training and proposed Instance Relation Graph (IRG) network guided contrastive loss. often have poor generalization when deployed in visual domains not encountered during training. In such cases, most works in the literature follow the Unsupervised Do- main Adaptation (UDA) setting to improve generalization [7,14,23,24,57,62]. Speciﬁcally, UDA methods aim to minimize the domain discrepancy by aligning the feature distribution of the detector model between source and tar- get domain [ 9,19,28,56,59]. To perform feature align- ment, UDA methods require simultaneous access to the la- beled source and unlabeled target data. However in practi- cal scenarios, the access to source data is often restricted due to concerns related to privacy/safety, data transmis- sion, data proprietary etc. For example, consider a detec- tion model trained on large-scale source data, that performs poorly when deployed in new devices having data with dif- ferent visual domains. In such cases, it is far more efﬁ- cient to transmit the source-trained detector model ( ∼500- 1000MB) for adaptation rather than transmitting the source data (∼10-100GB) to these new devices [ 27,37]. Moreover, transmitting only source-trained model alleviates many pri- vacy/safety, data proprietary concerns as well [ 41,47,70]. Hence, adapting the source-trained model to the target do- main without having access to source data is essential in the case of practical deployment of detection models. This mo- tivates us to study Source-Free Domain Adaptation (SFDA) setting for adapting object detectors (illustrated in Fig. 1). The SFDA is a more challenging setting than UDA. Speciﬁcally, on top of having no labels for the target data, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3520 (a) Prediction (b) Ground truth Figure 2. (a) Object predictions by Cityscapes-trained model on the FoggyCityscapes image. (b) Corresponding ground truth. Here, the proposals around the bus instance have inconsistent pre- dictions, indicating that instance features are prone to large shift in the feature space, for a small shift in the proposal location. the source data is not accessible during adaptation. There- fore, most SFDA methods for detection consider train- ing with pseudo-labels generated by source-trained model [27,40]. During our initial SFDA training experiments, we identiﬁed two key challenges. Firstly, noisy pseudo- labels generated by the source-trained model due to domain shift can result in suboptimal distillation of target domain information into the source-trained model [ 11,46]. Sec- ondly, Fig. 2shows object proposals for an image from FoggyCityscapes, predicted by a detector model trained on Cityscapes. Here, all the proposals have Intersection-over- Union>0.9 with respective ground-truth bounding boxes and each proposal is assigned a prediction with a conﬁdence score. Noticeably, the proposals around the bus instance have different predictions, e.g., car with 18 %, truck with 93%, and bus with 29 %conﬁdence. This indicates that the pooled features are prone to a large shift in the feature space for a small shift in the proposal location. This is because, the source-trained model representations would tend to be biassed towards source data, resulting in weak representa- tion for the target data. Therefore, we consider two major challenges in SFDA training: 1) Effectively distill target do- main information into source-trained model 2) Enhancing the target domain feature representations . Motivated by [ 46], we utilize mean-teacher [ 61] frame- work to effectively distill of target domain knowledge into source-trained model. However, the key challenge of en- hancing the target domain feature representations remained. To address this, we turned to contrastive representation learning (CRL) methods, has been shown to learn high- quality representations from images in an unsupervised manner [ 5,6,69]. CRL methods achieve this by forcing representations to be similar under multiple views (or aug- mentations) of an anchor image and dissimilar to all other images. In classiﬁcation, the CRL methods assume that each image contains only one object. On the contrary, for object detection, each image is highly likely to have multiple object instances. Furthermore, the CRL train- ing also requires large batch sizes and multiple views to Source only model (a) RPN ProposalMultiple views cropped from RPN proposals PushPull Pull (b) RPN-view Contrastive LearningBus Car Figure 3. (a) Class agnostic object proposals generated by Re- gion Proposal Network (RPN). (b) Cropping out RPN propos- als will provide multiple contrastive views of an object instance. We utilize this to improve target domain feature representations through RPN-view contrastive learning. However as RPN pro- posals are class agnostic, it is challenging to form positive (same class)/negative pairs (different class), which is essential for CRL. learn high-quality representations, which incurs a very high GPU/memory cost, as detection models are computation- ally expensive. To circumvent these issues, we propose an alternative strategy which exploits the architecture of the detection model like Faster-RCNN [ 54]. Interestingly, the proposals generated by the Region Proposal Network (RPN) of a Faster-RCNN essentially provide multiple views for any object instance as shown in Fig. 3(a). In other words, the RPN module provides instance augmentation for free , which could be exploited for CRL, as shown in Fig. 3(b). However, RPN predictions are class agnos- tic and without the ground-truth annotations for target do- main, it is impossible to know which of these proposals would form positive (same class)/negative pairs (different class), which is essential for CRL. To this end, we propose a Graph Convolution Network (GCN) based network that models the inter-instance relations for generated RPN pro- posals. Speciﬁcally, each node corresponds to a proposal and the edges represent the similarity relations between the proposals. This learned similarity relations are utilized to extract information regarding which proposals would form positive/negative pairs and are used to guide CRL. By doing so, we show that such graph-guided contrastive representa- tion learning is able to enhance representations for the target data. Our contributions are summarized as follows: • We investigate the problem of source-free domain adap- tation for object detection and identify some of the major challenges that need to be addressed. • We introduced an Instance Relation Graph (IRG) frame- work to model the relationship between proposals gener- ated by the region proposal network. • We propose a novel contrastive loss which is guided by the IRG network to improve the feature representations for the target data. • The effectiveness of the proposed method is evaluated on multiple object detection benchmarks comprising of visu- ally distinct domains. Our method outperforms existing source-free domain adaptation methods and many unsu- pervised domain adaptation methods. 3521
Wen_Highly_Confident_Local_Structure_Based_Consensus_Graph_Learning_for_Incomplete_CVPR_2023	Abstract Graph-based multi-view clustering has attracted exten- sive attention because of the powerful clustering-structure representation ability and noise robustness. Considering the reality of a large amount of incomplete data, in this pa- per, we propose a simple but effective method for incomplete multi-view clustering based on consensus graph learning, termed as HCLS CGL. Unlike existing methods that uti- lize graph constructed from raw data to aid in the learn- ing of consistent representation, our method directly learns a consensus graph across views for clustering. Speciﬁ- cally, we design a novel conﬁdence graph and embed it to form a conﬁdence structure driven consensus graph learn- ing model. Our conﬁdence graph is based on an intu- itive similar-nearest-neighbor hypothesis, which does not require any additional information and can help the model to obtain a high-quality consensus graph for better cluster- ing. Numerous experiments are performed to conﬁrm the effectiveness of our method.	1. Introduction In today’s world, graphs are common tools to mine the intrinsic structure of data. Graph learning, as a power- ful data analysis approach, has attracted increasing atten- tion over the past few years [38, 42]. More and more ma- chine learning and data mining tasks attempt to adopt graph learning to enhance the ability of structural representation and performance [35, 41, 43]. As an emerging data analy- sis and representation task, multi-view clustering also needs to deeply mine the geometric structure information of data samples [15,36,44]. That is also the core focus of this paper. As we all know, multi-view data is collected from differ- ent sources or perspectives, which is a diverse description of the target, and this diversity endows the data with stronger *Corresponding author.†Co-ﬁrst authors. ？Intersectionof local nearest neighborConfidence graphComputeweightFigure 1. Our key motivation: A novel conﬁdence graph whose edges are computed by the nodes’ shared nearest neighbors. discriminative and expressive power [19, 48]. For example, in the webpage recommendation tasks, we can assess the attributes of webpages from different elements such as au- dio, video, image, and text; in the autonomous driving sce- nario, ultrasonic radar, optical camera, and infrared radar together provide data basis for road trafﬁc analysis. These different styles of multi-view data play an extremely impor- tant role in real-world applications and provide a decision- making basis for plentiful downstream tasks [4, 18]. To better mine and utilize multi-view data, researchers have proposed abundant multi-view learning methods. These methods can be broadly divided into the following cate- gories according to different technical routes: multiple ker- nel kmeans based methods, canonical correlation analysis (CCA) based methods, subspace learning based methods, spectral clustering based methods, nonnegative matrix fac- torization (NMF) based methods, and deep learning based methods [2, 3, 5]. However, these conventional multi-view clustering methods are built on the premise that each sam- ple has all complete views. Obviously, this idealized as- sumption is often violated in practice. Missing data is al- most inevitable due to human oversight or the reasons from the application scenario itself [17]. Numerous solutions have been proposed for this incomplete multi-view cluster- ing (IMC) problem. The classic PMVC (partial multi-view clustering) [17] seeks a common space to connect two in- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15712 complete views. Liu et al . proposed a matrix factoriza- tion based model dubbed as LSIMVC [20], which directly learns a consistent low-dimensional representation of multi- view feature data and introduces prior indicator information to he
Wang_Out-of-Distributed_Semantic_Pruning_for_Robust_Semi-Supervised_Learning_CVPR_2023	Abstract Recent advances in robust semi-supervised learning (SSL) typically ﬁlter out-of-distribution (OOD) information at the sample level. We argue that an overlooked problem ofrobust SSL is its corrupted information on semantic level, practically limiting the development of the ﬁeld. In this pa- per , we take an initial step to explore and propose a uniﬁed framework termed OOD Semantic Pruning (OSP), whichaims at pruning OOD semantics out from in-distribution (ID) features. Speciﬁcally, (i) we propose an aliasing OOD matching module to pair each ID sample with an OOD sam-ple with semantic overlap. (ii) We design a soft orthogonal- ity regularization, which ﬁrst transforms each ID feature bysuppressing its semantic component that is collinear with paired OOD sample. It then forces the predictions beforeand after soft orthogonality decomposition to be consistent. Being practically simple, our method shows a strong per-formance in OOD detection and ID classiﬁcation on chal-lenging benchmarks. In particular , OSP surpasses the pre- vious state-of-the-art by 13.7% on accuracy for ID classiﬁ- cation and 5.9% on AUROC for OOD detection on TinyIm- ageNet dataset. The source codes are publicly available athttps://github.com/rain305f/OSP .	1. Introduction Deep neural networks have obtained impressive per- formance on various tasks [ 30,44,46]. Their success is partially dependent on a large amount of labeled train-ing data, of which the acquisition is expensive and time-consuming [ 20,25,40,45]. A prevailing way to reduce the dependency on human annotation is semi-supervised learn-ing (SSL). It learns informative semantics using annotation- *Equal contribution. †Corresponding author. (b) (c)OOD Samples ID Samples OOD Semantic OOD Semantic Pruning Predict : Beetles Predict : Butterfly Anchor ID FeaturesPruned ID FeatureID FeatureOOD Semantic Pruning OOD Semantic (a) Figure 1. (a) Intuitive diagram of OOD Semantic Pruning (OSP), that pruning OOD semantics out from ID features. (b) t-SNE vi- sualization [ 52] from the baseline [ 23]. (c) t-SNE visualization from our OSP model. The colorful dots donate ID features, whilethe black dots mark OOD features. The dots with the same colorrepresent the features of the same class. Here, our OSP and thebaseline are trained on CIFAR100 with 100 labeled data per classand 60% OOD in unlabeled data. free and acquisition-easy unlabeled data to extend the la- bel information from limited labeled data and has achievedpromising results in various tasks [ 38,50,51,54]. Unfortunately, classical SSL relies on a basic assump- tion that the labeled and unlabeled data are collected from the same distribution, which is difﬁcult to hold in real-worldapplications. In most practical cases, unlabeled data usually contains classes that are not seen in the labeled data. Exist- ing works [ 12,17,21,40] have shown that training the SSL model with these OOD samples in unlabeled data leads toa large degradation in performance. To solve this problem,robust semi-supervised learning (Robust SSL) has been in- vestigated to train a classiﬁcation model that performs sta- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23849 bly when the unlabeled set is corrupted by OOD samples. Typical methods focus on discarding OOD information at the sample level, that is, detecting and ﬁltering OOD sam-ples to purify the unlabeled set [ 12,17,21,57]. However, these methods ignore the semantic-level pollution caused by the classiﬁcation-useless semantics from OOD samples, which improperly disturbs the feature distribution learnedfrom ID samples, eventually resulting in weak ID and OODdiscrimination and low classiﬁcation performance. We pro-vide an example to explain such a problem in Fig. 1.A sw e can see, due to the semantics of Orchid in OOD examples, the model pays too much attention to the background andmisclassiﬁes the Butterﬂy asBeetle . In this paper, we propose Out-of-distributed Semantic Pruning (OSP) method to solve the problem mentionedabove and achieve effective robust SSL. Concretely, our OSP approach consists of two main modules. We ﬁrst develop an aliasing OOD matching mod-ule to pair each ID sample with an OOD sample with which it has feature aliasing. Secondly, we propose a soft orthogo-nality regularization, which constrains the predictions of ID samples to keep consistent before and after soft-orthogonal decomposition according to their matching OOD samples. We evaluate the effectiveness of our OSP in extensive robust semi-supervised image recognition benchmarks in- cluding MNIST [ 53], CIFAR10 [ 29], CIFAR100 [ 29] and TinyImageNet [ 15]. We show that our OSP obtains signiﬁ- cant improvements compared to state-of-the-art alternatives (e.g., 13.7% and 15.0% on TinyImagetNet with an OOD ratio of 0.3 and 0.6 respectively). Besides, we also empiri-cally demonstrate that OSP indeed increases the feature dis-crimination between ID and OOD samples. To summarize,the contributions of this work are as follows: • To the best of our knowledge, we are the ﬁrst to exploit the OOD effects at the semantic level by regularizationID features to be orthogonal to OOD features. • We develop an aliasing OOD matching module that adaptively pairs each ID sample with an OOD sample. In addition, we propose a soft orthogonality regulariza- tion to restrict ID and OOD features to be orthogonal. • We conduct extensive experiments on four datasets, i.e., MNIST, CIFAR10, CIFAR100, and TIN200, and achieve new SOTA performance. Moreover, we ana-lyze that the superiority of OSP lies in the enhanced discrimination between ID and OOD features.
Wang_CF-Font_Content_Fusion_for_Few-Shot_Font_Generation_CVPR_2023	Abstract Content and style disentanglement is an effective way to achieve few-shot font generation. It allows to transfer the style of the font image in a source domain to the style de- fined with a few reference images in a target domain. How- ever, the content feature extracted using a representative font might not be optimal. In light of this, we propose a con- tent fusion module (CFM) to project the content feature into a linear space defined by the content features of basis fonts, which can take the variation of content features caused by different fonts into consideration. Our method also al- lows to optimize the style representation vector of reference images through a lightweight iterative style-vector refine- ment (ISR) strategy. Moreover, we treat the 1D projection of a character image as a probability distribution and leverage the distance between two distributions as the reconstruc- tion loss (namely projected character loss, PCL). Compared to L2 or L1 reconstruction loss, the distribution distance pays more attention to the global shape of characters. We *This work was done during an internship at Alibaba Group. †Corresponding author.have evaluated our method on a dataset of 300 fonts with 6.5k characters each. Experimental results verify that our method outperforms existing state-of-the-art few-shot font generation methods by a large margin. The source code can be found at https://github.com/wangchi95/CF-Font.	1. Introduction Few-shot font generation aims to produce characters of a new font by transforming font images from a source do- main to a target domain according to just a few reference images. It can greatly reduce the labor of expert design- ers to create a new style of fonts, especially for logographic languages that contain multiple characters, such as Chinese (over 60K characters), Japanese (over 50K characters), and Korean (over 11K characters), since only several reference images need to be manually designed. Therefore, font gen- eration has wide applications in font completion for ancient books and monuments, personal font generation, etc. Recently, with the rapid development of convolu- tional neural networks [22] and generative adversarial net- works [9] (GAN), pioneers have made great progress in This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1858 generating gratifying logographic fonts. Zi2zi [38] intro- duces pix2pix [14] method to generate complex charac- ters of logographic languages with high quality, but it can- not handle those fonts that do not appear in training (un- seen fonts). For the few-shot font generation, many meth- ods [3, 7, 31, 32, 34, 42, 47] verify that content and style dis- entanglement is effective to convert the style of a character in the source domain, denoted as source character , to the target style embodied with reference images of seen or un- seen fonts. The neural networks in these methods usually have two branches to learn content and style features respec- tively, and the content features are usually obtained with the character image from a manually-chosen font, denoted as source font . However, since it’s a difficult task to achieve a complete disentanglement between content and style fea- tures [17, 21], the choice of the font for content-feature en- coding influences the font generation results substantially. For instance, Song andKaiare commonly selected as the source font [20, 28, 31, 42, 43, 47]. While such choices are effective in many cases, the generated images sometimes contain artifacts, such as incomplete and unwanted strokes. The main contribution of this paper is a novel content feature fusion scheme to mitigate the influence of incom- plete disentanglement by exploring the synchronization of content and style features, which significantly enhances the quality of few-shot font generation. Specifically, we design a content fusion module (CFM) to take the content features of different fonts into consideration during training and in- ference. It is realized by computing the content feature of a character of a target font through linearly blending con- tent features of the corresponding characters in the auto- matically determined basis fonts, and the blending weights are determined through a carefully designed font-level dis- tance measure. In this way, we can form a linear cluster for the content feature of a semantic character, and explore how to leverage the font-level similarity to seek for an opti- mized content feature in this cluster to improve the quality of generated characters. In addition, we introduce an iterative style-vector refine- ment (ISR) strategy to find a better style feature vector for font-level style representation. For each font, we average the style vectors of reference images and treat it as a learn- able parameter. Afterward, we fine-tune the style vector with a reconstruction loss, which further improves the qual- ity of the generated fonts. Most font-generation algorithms [3, 20, 31, 32, 38, 42] choose L1 loss as the character image reconstruction loss. However, L1 or L2 loss mainly supervises per-pixel accu- racy and is easily disturbed by the local misalignment of details. Hence, we employ a distribution-based projected character loss (PCL) to measure the shape difference be- tween characters. Specifically, by treating the 1D projec- tion of 2D character images as a 1D probability distribution,PCL computes the distribution distance to pay more atten- tion to the global properties of character shapes, resulting in the large improvement of skeleton topology transfer results. The CFM can be embedded into the few-shot font gen- eration task to enhance the quality of generated results. Ex- tensive experiments verify that our method, referred to as CF-Font, remarkably outperforms state-of-the-art methods on both seen and unseen fonts. Fig. 1 reveals that our method can generate high-quality fonts of various styles.
Wang_Multi-Agent_Automated_Machine_Learning_CVPR_2023	Abstract In this paper, we propose multi-agent automated ma- chine learning (MA2ML) with the aim to effectively han- dle joint optimization of modules in automated machine learning (AutoML). MA2ML takes each machine learning module, such as data augmentation (AUG), neural archi- tecture search (NAS), or hyper-parameters (HPO), as an agent and the final performance as the reward, to formu- late a multi-agent reinforcement learning problem. MA2ML explicitly assigns credit to each agent according to its marginal contribution to enhance cooperation among mod- ules, and incorporates off-policy learning to improve search efficiency. Theoretically, MA2ML guarantees monotonic improvement of joint optimization. Extensive experiments show that MA2ML yields the state-of-the-art top-1 accuracy on ImageNet under constraints of computational cost, e.g., 79.7%/80.5%with FLOPs fewer than 600M/800M. Exten- sive ablation studies verify the benefits of credit assignment and off-policy learning of MA2ML.	1. Introduction Automated machine learning (AutoML) aims to find high-performance machine learning (ML) pipelines without human effort involvement. The main challenge of AutoML lies in finding optimal solutions in huge search spaces. In recent years, reinforcement learning (RL) has been validated to be effective to optimize individual AutoML modules, such as data augmentation (AUG) [4], neural ar- chitecture search (NAS) [25, 53, 54], and hyper-parameter optimization (HPO) [38]. However, when facing the huge search space (Figure 1 left) and the joint optimization of these modules, the efficiency and performance challenges remain. Through experiments, we observed that among AutoML modules there exists a cooperative relationship that facilities the joint optimization of modules. For example, a small net- work (ResNet-34) with specified data augmentation and op- †The work was done during his Master program at Peking University. ‡Correspondence to [email protected] HPO AUG HPO AUG NAS NAS𝒌ା𝟏 𝒌 𝒑𝒌ା𝟏 𝒑𝒌Huge search space of ML pipelines Searched pipelineFigure 1. Search spaces of machine learning pipelines. Left: sin- gle agent controls all modules, and the huge search space makes it ineffective to learn. Mid: each agent controls one module, and the learning difficulty is reduced by introducing MA2ML. Right : MA2ML guarantees monotonic improvement of the searched pipeline, where pkandR(pk)denote the k-th searched pipeline and its expected performance, respectively. timized hyper-parameters significantly outperforms a large one (ResNet-50) with default training settings (76.8% vs. 76.1%). In other words, good AUG and HPO alleviate the need for NAS to some extent. Accordingly, we pro- pose multi-agent automated machine learning (MA2ML), which explores the cooperative relationship towards joint optimization of ML pipelines. In MA2ML, ML modules are defined as RL agents (Figure 1 mid), which take ac- tions to jointly maximize the reward, so that the training ef- ficiency and test accuracy are significantly improved. Spe- cially, we introduce credit assignment to differentiate the contribution of each module, such that all modules can be simultaneously updated. To handle both continuous ( e.g., learning rate) and discrete ( e.g., architecture) action spaces, MA2ML employs a multi-agent actor-critic method, where a centralized Q-function is learned to evaluate the joint ac- tion. Besides, to further improve search efficiency, MA2ML adopts off-policy learning to exploit historical samples for policy updates. MA2ML is justified theoretically and experimentally. Theoretically, we prove that MA2ML guarantees mono- tonic policy improvement (Figure 1 right ),i.e., the per- formance of the searched pipeline monotonically improves in expectation. This enables MA2ML to fit the joint op- timization problem and be adaptive to all modules in the ML pipeline, potentially achieving full automation. Exper- imentally, we take the combination of individual RL-based modules to form MA2ML-Lite, and compare their perfor- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11960 mance on ImageNet [31] and CIFAR-10/100 [20] datasets. To better balance performance and computational cost, we add constraints of FLOPs in the experiment on ImageNet. Experiments show that MA2ML substantially outperforms MA2ML-Lite w.r.t. both accuracy and sample efficiency, and MA2ML achieves remarkable accuracy compared with recent methods. Our contributions are summarized as follows: • We propose MA2ML, which utilizes credit assignment to differentiate the contributions of ML modules, pro- viding a systematic solution for the joint optimization of AutoML modules. • We prove the monotonic improvement of module poli- cies, which enables to MA2ML fit the joint optimiza- tion problem and be adaptive to various modules in the ML pipeline. • MA2ML yields the state-of-the-art performance under constraints of computational cost, e.g.,79.7%/80.5% on ImageNet, with FLOPs fewer than 600M/800M, validating the superiority of the joint optimization of MA2ML.
Xue_Egocentric_Video_Task_Translation_CVPR_2023	Abstract Different video understanding tasks are typically treated in isolation, and even with distinct types of curated data (e.g., classifying sports in one dataset, tracking animals in another). However, in wearable cameras, the immer- sive egocentric perspective of a person engaging with the world around them presents an interconnected web of video understanding tasks—hand-object manipulations, naviga- tion in the space, or human-human interactions—that un- fold continuously, driven by the person’s goals. We argue that this calls for a much more uniﬁed approach. We pro- pose EgoTask Translation (EgoT2), which takes a collec- tion of models optimized on separate tasks and learns to translate their outputs for improved performance on any or all of them at once. Unlike traditional transfer or multi- task learning, EgoT2’s “ﬂipped design” entails separate task-speciﬁc backbones and a task translator shared across all tasks, which captures synergies between even heteroge- neous tasks and mitigates task competition. Demonstrat- ing our model on a wide array of video tasks from Ego4D, we show its advantages over existing transfer paradigms and achieve top-ranked results on four of the Ego4D 2022 benchmark challenges.1	1. Introduction In recent years, the introduction of large-scale video datasets ( e.g., Kinetics [ 6,33] and Something- Something [ 22]) have enabled the application of powerful deep learning models to video understanding and have led to dramatic advances. These third-person datasets, however, have overwhelmingly focused on the single task of action recognition in trimmed clips [ 12,36,47,64]. Unlike curated third-person videos, our daily life involves frequent and heterogeneous interactions with other hu- mans, objects, and environments in the wild. First-person videos from wearable cameras capture the observer’s perspective and attention as a continuous stream. As such, *Work done during an internship at FAIR, Meta AI. 1Project webpage: https : / / vision . cs . utexas . edu / projects/egot2/ . Task of InterestHolistic Egocentric PerceptionEgoT2Task Relatednesssmall large [Task 1][Task 2][Task 3][Task 4] ...Is she looking at me? Did the object gothrough a state change?Who is speaking?Is she talking to me? What am I doing? What will I do next?When did I knead dough? Auxiliary TasksWhat will I do next?Figure 1. Given a set of diverse egocentric video tasks, the pro- posed EgoT2 leverages synergies among the tasks to improve each individual task performance. The attention maps produced by EgoT2 offer good interpretability on inherent task relations. they are better equipped to reveal these multi-faceted, spontaneous interactions. Indeed egocentric datasets, such as EPIC-Kitchens [ 9] and Ego4D [ 23], provide suites of tasks associated with varied interactions. However, while these benchmarks have promoted a broader and more heterogeneous view of video understanding, they risk perpetuating the fragmented development of models specialized for each individual task. In this work, we argue that the egocentric perspective offers an opportunity for holistic perception that can ben- eﬁcially leverage synergies among video tasks to solve all problems in a uniﬁed manner. See Figure 1. Imagine a cooking scenario where the camera wearer ac- tively interacts with objects and other people in an environ- ment while preparing dinner. These interactions relate to each other: a hand grasping a knife suggests the upcoming action of cutting; the view of a tomato on a cutting board suggests that the object is likely to undergo a state transi- tion from whole to chopped; the conversation may further reveal the camera wearer’s ongoing and planned actions. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2310 Apart from the natural relation among these tasks, egocen- tric video’s partial observability (i.e., the camera wearer is largely out of the ﬁeld of view) further motivates us to seek synergistic, comprehensive video understanding to leverage complementary cues among multiple tasks. Our goal presents several technical challenges for con- ventional transfer learning (TL) [ 65] and multi-task learn- ing (MTL) [ 63]. First, MTL requires training sets where each sample includes annotations for all tasks [ 15,24,48, 53,55,62], which is often impractical. Second, egocen- tric video tasks are heterogeneous in nature, requiring dif- ferent modalities (audio, visual, motion), diverse labels (e.g., temporal, spatial or semantic), and different tem- poral granularities ( e.g., action anticipation requires long- term observations, but object state recognition operates at a few sparsely sampled frames)—all of which makes a uniﬁed model design problematic and fosters specializa- tion. Finally, while existing work advocates the use of a shared encoder across tasks to learn general representa- tions [ 3,18,26,32,39,44,45,51], the diverse span of ego- centric tasks poses a hazard to parameter sharing which can lead to negative transfer [ 21,24,38,53]. To address the above limitations, we propose EgoTask Translation (EgoT2), a uniﬁed learning framework to ad- dress a diverse set of egocentric video tasks together. EgoT2 is ﬂexible and general in that it can handle separate datasets for the different tasks; it takes video heterogeneity into ac- count; and it mitigates negative transfer when tasks are not strongly related. To be speciﬁc, EgoT2 consists of special- ized models developed for individual tasks and a task trans- lator that explicitly models inter-task and inter-frame rela- tions. We propose two distinct designs: (1) task-speciﬁc EgoT2 (EgoT2-s) optimizes a given primary task with the assistance of auxiliary tasks (Figure 2(c)) while (2) task- general EgoT2 (EgoT2-g) supports task translation for mul- tiple tasks at the same time (Figure 2(d)). Compared with a uniﬁed backbone across tasks [ 62], adopting task-speciﬁc backbones preserves peculiarities of each task ( e.g. different temporal granularities) and miti- gates negative transfer since each backbone is optimized on one task. Furthermore, unlike traditional parameter shar- ing [ 51], the proposed task translator learns to “translate” all task features into predictions for the target task by se- lectively activating useful features and discarding irrelevant ones. The task translator also facilitates interpretability by explicitly revealing which temporal segments and which subsets of tasks contribute to improving a given task. We evaluate EgoT2 on a diverse set of 7 egocentric perception tasks from the world’s largest egocentric video benchmark, Ego4D [ 23]. Its heterogeneous tasks extend beyond mere action recognition to speaker/listener identi- ﬁcation, keyframe localization, object state change classiﬁ- cation, long-term action anticipation, and others, and pro-Shared BackboneHead AHead BHead CTask ATask BTask CBackbonepretrained on Task ATransfer LayersTask B Backbone A Shared TaskTranslatorTask ATask BTask C Backbone B Backbone C (c) EgoT2-s(a) conventional TL (d) EgoT2-g Backbone A Task Translator Task B Backbone B Backbone C (b) conventional MTL Figure 2. (a) Conventional TL uses a backbone pretrained on the source task followed by a head transferring supervision to the tar- get task; (b) Traditional MTL consists of a shared backbone and several task-speciﬁc heads; (c) EgoT2-s adopts task-speciﬁc back- bones and optimizes the task translator for a given primary task; (d) EgoT2-g jointly optimizes the task translator for all tasks. vide a perfect ﬁt for our study. Our results reveal inher- ent task synergies, demonstrate consistent performance im- provement across tasks, and offer good interpretability in task translation. Among all four Ego4D challenges involved in our task setup, EgoT2 outperforms all submissions to three Ego4D-CVPR’22 challenges and achieves state-of- the-art performance in one Ego4D-ECCV’22 challenge.
Xie_Toward_Stable_Interpretable_and_Lightweight_Hyperspectral_Super-Resolution_CVPR_2023	Abstract For real applications, existing HSI-SR methods are not only limited to unstable performance under unknown sce- narios but also suffer from high computation consumption. In this paper, we develop a new coordination optimiza- tion framework for stable, interpretable, and lightweight HSI-SR. Specifically, we create a positive cycle between fu- sion and degradation estimation under a new probabilis- tic framework. The estimated degradation is applied to fu- sion as guidance for a degradation-aware HSI-SR. Under the framework, we establish an explicit degradation estima- tion method to tackle the indeterminacy and unstable per- formance caused by the black-box simulation in previous methods. Considering the interpretability in fusion, we in- tegrate spectral mixing prior into the fusion process, which can be easily realized by a tiny autoencoder, leading to a dramatic release of the computation burden. Based on the spectral mixing prior, we then develop a partial fine- tune strategy to reduce the computation cost further. Com- prehensive experiments demonstrate the superiority of our method against the state-of-the-arts under synthetic and real datasets. For instance, we achieve a 2.3dB promo- tion on PSNR with 120×model size reduction and 4300× FLOPs reduction under the CAVE dataset. Code is avail- able in https://github.com/WenjinGuo/DAEM .	1. Introduction Different from traditional optical images with a few channels, hyperspectral images (HSIs) with tens to hun- dreds of bands hold discriminative information about ma- terials, leading to a great advantage in a wide range of ap- plications, e.g., the monitoring and management of ecosys- tems, biodiversity, and disasters [1–8]. However, the physi- cal limitation in imaging causes a trade-off between spatial *Equal contribution. †Cooresponding author. CUCaNetMHFNet UAL Ours 0.1 1103060 120Size (M) 13090027000 10 100 1000 10000FLOPs (G) Time (s)(a) (b) Figure 1. Comparison among recent state-of-the-art (SOTA) meth- ods and our method. We report the computational efficiency (Size, FLOPs, and Time) in (a) and the distribution of two measurement metrics (PSNR, SAM) under various degradations in (b). Obvi- ously, our method is remarkably superior to others in lightweight, fidelity, and stability. and spectral resolution. HSIs are suffered from low spa- tial resolution in real applications. Therefore, hyperspectral super-resolution (HSI-SR), which aims to promote the spa- tial resolution of HSIs, has become a significant task, and always performs as a necessary pre-processing of HSI ap- plications, i.e. detection and classification [9–16]. Due to the common optical platforms which are equipped with both HSI sensor and multispectral sensor (such as satellites, airborne platforms), HSIs and multispec- tral images (MSIs) imaged in the same scene are easy to access. Fusion-based HSI-SR aims to estimate the desired high resolution HSI (HR-HSI) from the corresponding low resolution HSI (LR-HSI) and high resolution MSI (HR- MSI). Naturally, HSI-SR can be modeled as a maximum a posterior (MAP) estimation: p(Z\|X,Y)∝p(X,Y\|Z,θ)p(Z\|ϕ), (1) where Z ∈RH×W×Bis the target HR-HSI with H,Wand Bas its height, width and number of bands, respectively. X ∈Rh×w×Bis the observed LR-HSI with handwas This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22272 its height and width ( h < H, w < W ).Y ∈RH×W×b is the HR-MSI with bbands ( b < B ).θandϕare pa- rameters in the likelihood and the prior, respectively. As a pre-processing task, HSI-SR faces two challenges, high fidelity recovery and efficient processing (i.e., low compu- tational burden and fast processing). Referred to Eq. (1), we analyse these two ingredients from three perspectives: the likelihood term, the prior term, and coordination between them. Likelihood Term The likelihood term reflects the degradation process, and the widely accepted degradation model can be formulated as: X= (Z ∗C)↓s+NX, Y=Z × 3R+NY,(2) where C∈Rs×s,R∈RB×brepresent the PSF (point spread function) and SRF (spectral response function), re- spectively. ↓srepresents spatial downsampling with s times. ×3is matrix multiple on the third dimension. Early methods demand precise degradation parameters to opti- mize the likelihood [17–19]. However, PSF and SRF are various and unknown in real scenarios, hence the high de- mand for blind HSI-SR. In [20], two convolution blocks are built to simulate the degradation process, which learn the degradation in training sets and apply it to testing samples. Unsupervised blind HSI-SR methods estimate degradation in each image-pair individually [21–23]. Yao et al. [22] pro- pose a spatial-spectral consistency to further constrain the degradation estimation. Despite better applicability, recent blind HSI-SR methods are limited by the DL-based estima- tion. On one hand, the implicit modeling of degradation imports a black box in optimization and draws uncertainty, especially in volatile degradation. On the other hand, degra- dation estimation is an inverse problem with multiple candi- dates, the common over-fitting in neural networks will lead to inaccurate and unstable performance. Prior Term If only considering the likelihood term, HSI-SR is an ill-posed problem with infinite solutions. The prior term shrinks the solution space as a regularization. Earlier works make assumptions on prior through manual induction of data characteristics, such as low rank [24, 25], and spar- sity [26–30]. Recent supervised deep learning (DL)-based methods replace this process through data characterization by neural networks. These inductive methods will drop a lot on performance when testing samples differ training samples, hence a critical need for general prior assumption. As an inherent feature in HSIs, spectral mixing prior sim- ulates the common spectral mixing phenomenon in imag- ing. Unmixing-based methods make a breakthrough in fi- delity [17,18,31,32]. Moreover, to promote the generaliza- tion, coupled autoencoder is proposed to simulate the spec-tral mixing with deep learning toolkit [21, 22]. Despite su- perior fidelity, recent unsupervised DL-based methods over- focus on network architecture and underestimate the effort of prior knowledge in HSIs, resulted in two drawbacks. On one hand, the over-designed network structures weaken the role of the prior assumption, which harms the interpretabil- ity. On the other hand, the complicated network structures pose a heavy burden on power and memory. Coordination between Likelihood and Prior The most recent blind HSI-SR methods generally con- tain two modules, the fusion module and the degradation estimator [22,23,33]. From the viewpoint of MAP problem, the former aims to recover the HR-HSI through the obser- vations under the regularization of the prior. The later mini- mizes the likelihood term by estimating degradation param- eters. The unknown HR-HSI and degradation determine the observations. Naturally, the estimated degradation can be fed to the fusion module as a guidance. However, in recent methods, the degradation estimator and fusion module only interacts in backward stage. The learned degradation pa- rameters are not involved in the updating of fusion module or prior parameters, which brings two weaknesses. Firstly, the nearly independent optimization of likelihood and prior causes more updating steps and slows the convergence. Sec- ondly, optimization with less interaction inducts conflict op- timization directions and results in local optimum with un- real recovery. In this paper, we develop a novel coordination optimiza- tion framework for stable, interpretable, and lightweight HSI-SR. Through integrating the Wald protocol in the MAP problem, we construct a postive feedback loop between prior and likelihood. Based on the framework, we explore an explicit degradation estimation to remedy the unstable performance of black-box estimation. As for the prior, we establish a lightweight autoencoder to simulate the spectral mixing prior in HSIs for a interpretable fusion. Our contri- butions are summarized as follows: 1. We explore a coordination optimization framework for HSI-SR with fast convergence and stable performance. Under the framework, the degradation estimation and fusion process promote each other concordantly. To the best of our knowledge, it is the first work to explore the cooperative relationship between prior and likelihood in HSI-SR. 2. An explicit estimation method is established in HSI- SR firstly. Through modeling PSF and SRF with anisotropic Gaussian kernel and Gaussian mixture, tiny parameters re- alize stable and precise estimation. 3. We simulate the general spectral mixing prior with only one interpretable autoencoder. Compared with recent complicated models, the network makes a breakthrough in lightweight. Based on the fusion netwrok, we explore a par- tial optimization strategy in test stage, which only updates the decoder that handles the individual spectral feature of 22273 images, reducing large computaion and time consumption.
Wu_OmniObject3D_Large-Vocabulary_3D_Object_Dataset_for_Realistic_Perception_Reconstruction_and_CVPR_2023	Abstract Recent advances in modeling 3D objects mostly rely on synthetic datasets due to the lack of large-scale real- scanned 3D databases. To facilitate the development of 3D perception, reconstruction, and generation in the real world, we propose OmniObject3D , a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects. OmniObject3D has several appealing properties: 1) Large Vocabulary: It comprises 6,000 scanned objects in 190 daily categories, sharing common classes with pop- ular 2D datasets (e.g., ImageNet and LVIS), benefiting the pursuit of generalizable 3D representations. 2) Rich An- notations: Each 3D object is captured with both 2D and BCorresponding authors. https://omniobject3d.github.io/3D sensors, providing textured meshes, point clouds, multi- view rendered images, and multiple real-captured videos. 3) Realistic Scans: The professional scanners support high- quality object scans with precise shapes and realistic ap- pearances. With the vast exploration space offered by Om- niObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b)novel-view synthesis, c)neural sur- face reconstruction, and d)3D object generation. Extensive studies are performed on these four benchmarks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision.	1. Introduction Sensing, understanding, and synthesizing realistic 3D objects is a long-standing problem in computer vision, with This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 803 rapid progress emerging in recent years. However, a major- ity of the technical approaches rely on unrealistic synthetic datasets [6, 19, 64] due to the absence of a large-scale real- world 3D object database. However, the appearance and distribution gaps between synthetic and real data cannot be compensated for trivially, hindering their real-life applica- tions. Therefore, it is imperative to equip the community with a large-scale and high-quality 3D object dataset from the real world, which can facilitate a variety of 3D vision tasks and downstream applications. Recent advances partially fulfill the requirements while still being unsatisfactory. As shown in Table 1, CO3D [48] contains 19k videos capturing objects from 50 MS-COCO categories, while only 20% of the videos are annotated with accurate point clouds reconstructed by COLMAP [50]. Moreover, they do not provide textured meshes. GSO [16] has 1k scanned objects while covering only 17 household classes. AKB-48 [33] focuses on robotics manipulation with 2k articulated object scans in 48 categories, but the focus on articulation leads to a relatively narrow semantic distribution, failing to support general 3D object research. To boost the research on general 3D object understand- ing and modeling, we present OmniObject3D : a large- vocabulary 3D object dataset with massive high-quality, real-scanned 3D objects. Our dataset has several appealing properties: 1) Large Vocabulary: It contains 6,000 high- quality textured meshes scanned from real-world objects, which, to the best of our knowledge, is the largest among real-world 3D object datasets with accurate 3D meshes. It comprises 190 daily categories, sharing common classes with popular 2D and 3D datasets ( e.g., ImageNet [15], LVIS [25], and ShapeNet [6]), incorporating most daily object realms (See Figure 1 and Figure 2). 2) Rich An- notations: Each 3D object is captured with both 2D and 3D sensors, providing textured 3D meshes, sampled point clouds, posed multi-view images rendered by Blender [13], and real-captured video frames with foreground masks and COLMAP camera poses. 3) Realistic Scans: The object scans are of high fidelity thanks to the professional scan- ners, bearing precise shapes with geometric details and re- alistic appearance with high-frequency textures. Taking advantage of the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a)robust 3D perception, b)novel-view synthesis, c)neural surface reconstruction, and d)3D object genera- tion. Extensive studies are performed on these benchmarks: First, the high-quality, real-world point clouds in OmniOb- ject3D allow us to perform robust 3D perception analysis on both out-of-distribution (OOD) styles and corruptions, two major challenges in point cloud OOD generalization. Furthermore, we provide massive 3D models with multi- view images and precise 3D meshes for novel-view syn- thesis andneural surface reconstruction . The broad diver-Table 1. A comparison between OmniObject3D and other commonly-used 3D object datasets. Rlvisdenotes the ratio of the 1.2k LVIS [25] categories being covered. Dataset Real Full Mesh Video # Objs # Cats Rlvis(%) ShapeNet [6] ✓ 51k 55 4.1 ModelNet [64] ✓ 12k 40 2.4 3D-Future [19] ✓ 16k 34 1.3 ABO [12] ✓ 8k 63 3.5 Toys4K [53] ✓ 4k 105 7.7 CO3D V1 / V2 [48] ✓ ✓ 19 / 40k 50 4.2 DTU [1] ✓ ✓ 124 NA 0 ScanObjectNN [56] ✓ 15k 15 1.3 GSO [16] ✓ ✓ 1k 17 0.9 AKB-48 [33] ✓ ✓ 2k 48 1.8 Ours ✓ ✓ ✓ 6k 190 10.8 sity in shapes and textures offers a comprehensive training and evaluation source for both scene-specific and general- izable algorithms. Finally, we equip the community with a database for large vocabulary and realistic 3D object gener- ation , which pushes the boundary of existing state-of-the- art generation methods to real-world 3D objects. The four benchmarks reveal new observations, challenges, and op- portunities for future research in realistic 3D vision.
Xie_Visibility_Aware_Human-Object_Interaction_Tracking_From_Single_RGB_Camera_CVPR_2023	Abstract Capturing the interactions between humans and their environment in 3D is important for many applications in robotics, graphics, and vision. Recent works to reconstruct the 3D human and object from a single RGB image do not have consistent relative translation across frames because they assume a fixed depth. Moreover, their performance drops significantly when the object is occluded. In this work, we propose a novel method to track the 3D human, object, contacts, and relative translation across frames from a single RGB camera, while being robust to heavy occlu- sions. Our method is built on two key insights. First, we condition our neural field reconstructions for human and object on per-frame SMPL model estimates obtained by pre-fitting SMPL to a video sequence. This improves neu- ral reconstruction accuracy and produces coherent relative translation across frames. Second, human and object mo- tion from visible frames provides valuable information to infer the occluded object. We propose a novel transformer- based neural network that explicitly uses object visibil- ity and human motion to leverage neighboring frames to make predictions for the occluded frames. Building on these insights, our method is able to track both human and object robustly even under occlusions. Experiments on two datasets show that our method significantly im- proves over the state-of-the-art methods. Our code and pre- trained models are available at: https://virtualhumans.mpi- inf.mpg.de/VisTracker.	1. Introduction Perceiving and understanding human as well as their in- teraction with the surroundings has lots of applications in robotics, gaming, animation and virtual reality etc. Ac- curate interaction capture is however very hard. Early works employ high-end systems such as dense camera ar- rays [8, 17, 37] that allow accurate capture but are expen- sive to deploy. Recent works [6, 34, 36] reduce the require- Teaser figure Figure 1. From a monocular RGB video, our method tracks the human, object and contacts between them even under occlusions. ment to multi-view RGBD cameras but it is still compli- cated to setup the full capture system hence is not friendly for consumer-level usage. This calls for methods that can capture human-object interaction from a single RGB cam- era, which is more convenient and user-friendly. However, reasoning about the 3D human and object from monocular RGB images is very challenging. The lack of depth information makes the predictions suscepti- ble to depth-scale ambiguity, leading to temporally inco- herent tracking. Furthermore, the object or human can get heavily occluded, making inference very hard. Prior work PHOSA [89] relies on hand-crafted heuristics to reduce the ambiguity but such heuristic-based method is neither very accurate nor scalable. More recently, CHORE [79] com- bines neural field reconstructions with model based fitting obtaining promising results. However, CHORE, assumes humans are at a fixed depth from the camera and predicts scale alone, thereby losing the important relative transla- tionacross frames. Another limitation of CHORE is that it is not robust under occlusions as little information is avail- able from single-frame when the object is barely visible. Hence CHORE often fails in these cases, see Fig. 3. In this work, we propose the first method that can track This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4757 both human and object accurately from monocular RGB videos. Our approach combines neural field predictions and model fitting, which has been consistently shown to be more effective than directly regressing pose [4–6, 79]. In contrast to existing neural field based reconstruction meth- ods [60,79], we can do tracking including inference of rela- tive translation. Instead of assuming a fixed depth, we con- dition the neural field reconstructions (for object and hu- man) on per frame SMPL estimates (SMPL-T) including translation in camera space obtained by pre-fitting SMPL to the video sequence. This results in coherent translation and improved neural reconstruction. In addition, we argue that during human-object interaction, the object motion is highly correlated with the human motion, which provides us valu- able information to recover the object pose even when it is occluded (see Fig. 1 column 3-4). To this end, we propose a novel transformer based network that leverages the hu- man motion and object motion from nearby visible frames to predict the object pose under heavy occlusions. We evaluate our method on the BEHA VE [6] and Inter- Cap dataset [35]. Experiments show that our method can robustly track human, object and realistic contacts between them even under heavy occlusions and significantly outper- forms the currrent state of the art method, CHORE [79]. We further ablate the proposed SMPL-T conditioning and human and visibility aware object pose prediction network and demonstrate that they are key for accurate human-object interaction tracking. In summary, our key contributions include: • We propose the first method that can jointly track full- body human interacting with a movable object from a monocular RGB camera. • We propose SMPL-T conditioned interaction fields , predicted by a neural network that allows consistent 4D tracking of human and object. • We introduce a novel human and visibility aware ob- ject pose prediction network along with an object visi- bility prediction network that can recover object poses even under heavy occlusions. • Our code and pretrained models are publicly available to foster future research in this direction.
Xia_SCPNet_Semantic_Scene_Completion_on_Point_Cloud_CVPR_2023	Abstract Training deep models for semantic scene completion (SSC) is challenging due to the sparse and incomplete in- put, a large quantity of objects of diverse scales as well asthe inherent label noise for moving objects. To address theabove-mentioned problems, we propose the following threesolutions: 1) Redesigning the completion sub-network. We design a novel completion sub-network, which consists ofseveral Multi-Path Blocks (MPBs) to aggregate multi-scalefeatures and is free from the lossy downsampling opera-tions. 2) Distilling rich knowledge from the multi-frame model. We design a novel knowledge distillation objective, dubbed Dense-to-Sparse Knowledge Distillation (DSKD).It transfers the dense, relation-based semantic knowledgefrom the multi-frame teacher to the single-frame student,signiﬁcantly improving the representation learning of thesingle-frame model. 3) Completion label rectiﬁcation. We propose a simple yet effective label rectiﬁcation strategy,which uses off-the-shelf panoptic segmentation labels to re-move the traces of dynamic objects in completion labels,greatly improving the performance of deep models espe-cially for those moving objects. Extensive experimentsare conducted in two public SSC benchmarks, i.e., Se- manticKITTI and SemanticPOSS. Our SCPNet ranks 1st on SemanticKITTI semantic scene completion challenge and surpasses the competitive S3CNet [ 3]b y 7.2 mIoU. SCP- Net also outperforms previous completion algorithms on theSemanticPOSS dataset. Besides, our method also achievescompetitive results on SemanticKITTI semantic segmenta-tion tasks, showing that knowledge learned in the scenecompletion is beneﬁcial to the segmentation task.	1. Introduction Semantic scene completion (SSC) [ 20] aims at inferring †: Corresponding author.both geometry and semantics of the scene from an incom- plete and sparse observation, which is a crucial componentin 3D scene understanding. Performing semantic scenecompletion in the outdoor scenarios is challenging due tothe sparse and incomplete input, a large quantity of objectsof diverse scales as well as the inherent label noise for thosemoving objects (See Fig. 1(a)). Recent years have witnessed an explosion of methods in the outdoor scene completion ﬁeld [ 3,19,22,25,29,32]. For example, S3CNet [ 3] performs 2D and 3D completion tasks jointly and achieves impressive performance on Se-manticKITTI [ 1]. JS3C-Net [ 29] performs semantic seg- mentation ﬁrst and then feeds the segmentation features tothe completion sub-network. The coarse-to-ﬁne reﬁnementmodule is further put forward to improve the completionquality. Although signiﬁcant progress has been achievedin this area, these methods heavily rely on the voxelwise completion labels and show unsatisfactory completion per- formance on the small, distant objects and crowded scenes.Moreover, the long traces of dynamic objects in originalcompletion labels will hamper the learning of completionmodels, which is overlooked in the previous literature [ 20]. To address the preceding problems, we propose three solutions from the aspects of the completion sub-network redesign, distillation of multi-frame knowledge as well ascompletion label rectiﬁcation. Speciﬁcally, we ﬁrst makea comprehensive overhaul of the completion sub-network.We adopt the completion-ﬁrst principle and make the com- pletion module directly process the raw voxel features. Be- sides, we avoid the use of downsampling operations sincethey inevitably introduce information loss and cause se-vere misclassiﬁcation for those small objects and crowdedscenes. To improve the completion quality on objects of di-verse scales, we design Multi-Path Blocks (MPBs) with var- ied kernel sizes, which aggregate multi-scale features andfully utilize the rich contextual information. Second, to combat against the sparse and incomplete in- put signals, we make the single-scan student model distilknowledge from the multi-frame teacher model. However, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17642 1520253035 2016 2017 2018 2019 2020 2021 2022 2023SSCNet-Full (16.1)TS3D+DNet+SATNet (17.7)ESSCNet (17.5) LMSCNet-SS (17.6)Local-DIFs (22.7)JS3CNet (23.8)S3CNet (29.5) UDNet (19.5)SSA-SC (23.5)SCPNet (36.7)mIoU (%) (a) (b) (c) Figure 1. Left: challenges of semantic scene completion, i.e., (a) sparse and incomplete input, varying completion difﬁculty for objects of diverse scales and (b) long traces of dynamic objects in completion labels. Traces of cars and persons are highlighted by yellow ellipses.Right: (c) performance comparison of various completion algorithms on SemanticKITTI semantic scene completion challenge. mimicking the probabilistic knowledge of each point/voxel brings marginal gains. Instead, we propose to distill thepairwise similarity information. Considering the sparsityand unorderness of features, we align the features usingtheir indices and then force the consistency between the pairwise similarity maps of student features and those of teacher features, make the student beneﬁt from the rela-tional knowledge of the teacher. The resulting Dense-to-Sparse Knowledge Distillation objective is termed DSKD,which is speciﬁcally designed for the scene completion task. Finally, to address the long traces of those dynamic ob- jects in the completion labels, we propose a simple yeteffective label rectiﬁcation strategy. The core idea is touse off-the-shelf panoptic segmentation labels to removethe traces of dynamic objects in completion labels. Therectiﬁed completion labels are more accurate and reliable,greatly improving the completion qualities of deep modelson those moving objects. We conduct experiments on two large-scale outdoor scene completion benchmarks, i.e., SemanticKITTI [ 1] and SemanticPOSS [ 16]. Our SCPNet ranks 1st on Se- manticKITTI semantic scene completion challenge 1and outperforms the S3CNet [ 3]b y 7.2 mIoU. SCPNet also achieves better performance than other completion algo-rithms on the SemanticPOSS dataset. The learned knowl-edge from the completion task also beneﬁts the segmenta-tion task, making our SCPNet achieve superior performanceon the SemanticKITTI semantic segmentation task. Our contributions are summarized as follows. • The comprehensive redesign of the completion sub- 1https://codalab.lisn.upsaclay.fr/competitions/7170#results till 2022- 11-12 00:00 Paciﬁc Time, and our method is termed SCPNet.network. We unveil several key factors to building strong completion networks. • To cope with the sparsity and incompleteness of the input, we propose to distill the dense relation-basedknowledge from the multi-frame model. Note that weare the ﬁrst to apply knowledge distillation to the se- mantic scene completion task. • To address the long traces of moving objects in com- pletion labels, we present the completion label rectiﬁ-cation strategy. • Our SCPNet ranks 1st on SemanticKITTI semantic scene completion challenge, outperforming the previ-ous SOT A S3CNet [ 3]b y 7.2 mIoU. Competitive per- formance is also shown in SemanticPOSS completiontask and SemanticKITTI semantic segmentation task.
Wang_Masked_Video_Distillation_Rethinking_Masked_Feature_Modeling_for_Self-Supervised_Video_CVPR_2023	Abstract Benefiting from masked visual modeling, self-supervised video representation learning has achieved remarkable progress. However, existing methods focus on learning rep- resentations from scratch through reconstructing low-level features like raw pixel values. In this paper, we propose masked video distillation (MVD), a simple yet effective two- stage masked feature modeling framework for video repre- sentation learning: firstly we pretrain an image (or video) model by recovering low-level features of masked patches, then we use the resulting features as targets for masked fea- ture modeling. For the choice of teacher models, we ob- serve that students taught by video teachers perform bet- ter on temporally-heavy video tasks, while image teachers transfer stronger spatial representations for spatially-heavy video tasks. Visualization analysis also indicates different teachers produce different learned patterns for students. To leverage the advantage of different teachers, we design a spatial-temporal co-teaching method for MVD. Specifically, we distill student models from both video teachers and im- age teachers by masked feature modeling. Extensive ex- perimental results demonstrate that video transformers pre- trained with spatial-temporal co-teaching outperform mod- els distilled with a single teacher on a multitude of video datasets. Our MVD with vanilla ViT achieves state-of-the- art performance compared with previous methods on sev- eral challenging video downstream tasks. For example, with the ViT-Large model, our MVD achieves 86.4% and 76.7% Top-1 accuracy on Kinetics-400 and Something-Something- v2, outperforming VideoMAE by 1.2% and 2.4% respec- tively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with 77.3% Top-1 accuracy on Something-Something-v2. Code will be avail- able at https://github.com/ruiwang2021/mvd . †Corresponding authors 0 2000 4000 6000 8000 GFLOPs/Video676971737577SSv2 T op-1 Acc %MVD(ours) VideoMAE ST-MAE OmniMAE BEVTMaskFeat MViTv1 MViTv2 VideoSwin MformerFigure 1. Comparisons of MVD with previous supervised or self-supervised methods on Something-Something v2. Each line represents the corresponding model of different sizes.	1. Introduction For self-supervised visual representation learning, recent masked image modeling (MIM) methods like MAE [31] and BEiT [2] achieve promising results with vision trans- formers [17] on various vision downstream tasks. Such a pretraining paradigm has also been adapted to the video domain and boosts video transformers by clear margins compared with supervised pretraining on several video downstream tasks. Representative masked video modeling (MVM) works include BEVT [63], VideoMAE [57] and ST-MAE [21]. Following MAE [31] and BEiT [2], existing masked video modeling methods [21, 57, 63] pretrain video trans- formers through reconstructing low-level features, e.g., raw pixel values or low-level VQV AE tokens. However, us- ing low-level features as reconstruction targets often incur much noise. And due to the high redundancy in video data, it is easy for masked video modeling to learn shortcuts, thus resulting in limited transfer performance on downstream This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6312 tasks. To alleviate this issue, masked video modeling [57] often uses larger masking ratios. In this paper, we observe that much better performance on video downstream tasks can be achieved by conducting masked feature prediction by using the high-level features of pretrained MIM and MVM models as masked predic- tion targets. This can be viewed as two-stage masked video modeling, where MIM pretrained image models ( i.e., an image teacher) or MVM pretrained video models ( i.e., an video teacher) are obtained in the first stage, and they fur- ther act as teachers in the second stage for the student model via providing the high-level feature targets. Therefore, we call this method Masked Video Distillation (MVD). More interestingly, we find that student models distilled with different teachers in MVD exhibit different proper- ties on different video downstream tasks. Specifically, stu- dents distilled from the image teacher perform better on video tasks that mainly rely on spatial clues, while students distilled from the video teacher model perform better on the video downstream tasks where temporal dynamics are more necessary. We think during the pretraining process of masked video modeling in the first stage, video teach- ers have learned spatial-temporal context in their high-level features. Therefore, when employing such high-level repre- sentations as prediction targets of masked feature modeling, it will help encouraging the student model to learn stronger temporal dynamics. By analogy, image teachers provide high-level features as targets that include more spatial in- formation, which can help the student model learn more spatially meaningful representations. We further analyze the feature targets provided by image teachers and video teachers, and calculate the cross-frame feature similarity. It shows that the features provided by the video teachers con- tain more temporal dynamics. Motivated by the above observation, to leverage the ad- vantages of video teachers and image teachers, we propose a simple yet effective spatial-temporal co-teaching strategy for MVD. In detail, the student model is designed to re- construct the features coming from both the image teacher and video teacher with two different decoders, so as to learn stronger spatial representation and temporal dynam- ics at the same time. Experiments demonstrate that MVD with co-teaching from both the image teacher and the video teacher significantly outperforms MVD only using one sin- gle teacher on several challenging downstream tasks. Despite the simplicity, our MVD is super effective and achieves very strong performance on multiple standard video recognition benchmarks. For example, on Kinectics- 400 and Something-Something-v2 datasets, compared to the baseline without distillation, MVD with 400 epochs us- ing a teacher model of the same size achieves 1.2% ,2.8% Top-1 accuracy gain on ViT-B. If a larger teacher model ViT-L is used, more significant performance gains ( i.e.,1.9% ,4.0% ) can be obtained. When ViT-Large is the target student model, our method can achieves 86.4% and76.7% Top-1 accuracy on these two datasets, surpassing existing state-of-the-art method VideoMAE [57] by 1.2% and2.4% respectively. When a larger ViT-Huge model is adopted, MVD achieves the state-of-the-art performance with 77.3% Top-1 accuracy on Something-Something-v2. Our contributions can be summarized as below: • We find that using MIM pretrained image models and MVM pretrained video models as teachers to provide the high-level features for continued masked feature prediction can learn better video representation. And representations learned with image teachers and video teachers show different properties on different down- stream video datasets. • We propose masked video distillation together with a simple yet effective co-teaching strategy, which enjoys the synergy of image and video teachers. • We demonstrate strong performance on multiple stan- dard video recognition benchmarks, surpassing both the baseline without MVD and prior state-of-the-art methods by clear margins.
Wu_STMixer_A_One-Stage_Sparse_Action_Detector_CVPR_2023	Abstract Traditional video action detectors typically adopt the two-stage pipeline, where a person detector is first em- ployed to generate actor boxes and then 3D RoIAlign is used to extract actor-specific features for classification. This detection paradigm requires multi-stage training and inference, and cannot capture context information outside the bounding box. Recently, a few query-based action de- tectors are proposed to predict action instances in an end- to-end manner. However, they still lack adaptability in fea- ture sampling and decoding, thus suffering from the issues of inferior performance or slower convergence. In this pa- per, we propose a new one-stage sparse action detector, termed STMixer. STMixer is based on two core designs. First, we present a query-based adaptive feature sampling module, which endows our STMixer with the flexibility of mining a set of discriminative features from the entire spa- tiotemporal domain. Second, we devise a dual-branch fea- ture mixing module, which allows our STMixer to dynami- cally attend to and mix video features along the spatial and the temporal dimension respectively for better feature de- coding. Coupling these two designs with a video backbone yields an efficient end-to-end action detector. Without bells and whistles, our STMixer obtains the state-of-the-art re- sults on the datasets of AVA, UCF101-24, and JHMDB.	1. Introduction Video action detection [14,18,20,30,32,44,46] is an im- portant problem in video understanding, which aims to rec- ognize all action instances present in a video and also local- ize them in both space and time. It has drawn a significant amount of research attention, due to its wide applications in many areas like security and sports analysis. Since the proposal of large-scale action detection bench- marks [16, 22], action detection has made remarkable progress. This progress is partially due to the advances of video representation learning such as video convolution *: Equal contribution. : Corresponding author. 150 200 250 300 350 400 GFLOPs2830323436mAP SlowFast,SF-R101-NL WOO,SF-R101-NLSTMixer,SF-R101-NL CSN,CSN-152TubeR,CSN-152STMixer,CSN-152 VideoMAE,ViT-BSTMixer,ViT-B (from VideoMAE)STMixer,ViT-B (from VideoMAE V2)Figure 1. Comparion of mAP versus GFLOPs. We report detection mAP on A V A v2.2. The GFLOPs of CSN, SlowFast, and VideoMAE are the sum of Faster RCNN-R101-FPN detector GFLOPs and classifier GFLOPs. Different methods are marked by different makers and models with the same backbone are marked in the same color. The results of CSN are from [53]. Our STMixer achieves the best effectiveness and efficiency balance. neural networks [5, 11, 39–41, 45, 50] and video transform- ers [1, 3, 9, 27, 38, 43, 52]. Most current action detectors adopt the two-stage Faster R-CNN-alike detection paradigm [31]. They share two ba- sic designs. First, they use an auxiliary human detector to generate actor bounding boxes in advance. The training of the human detector is decoupled from the action classifica- tion network. Second, in order to predict the action category for each actor box, the RoIAlign [17] operation is applied on video feature maps to extract actor-specific features. How- ever, these two-stage action detection pipeline has several critical issues. First, it requires multi-stage training of per- son detector and action classifier, which requires large com- puting resources. Furthermore, the RoIAlign [17] opera- tion constrains the video feature sampling inside the actor bounding box and lacks the flexibility of capturing context information in its surroundings. To enhance RoI features, recent works use an extra heavy module that introduces in- teraction features of context or other actors [29, 36]. Recently sparse query-based object detector [4, 35, 54] has brought a new perspective on detection tasks. Several This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14720 query-based sparse action detectors [6, 53] are proposed. The key idea is that action instances can be represented as a set of learnable queries, and detection can be formulated as a set prediction task, which could be trained by a match- ing loss. These query-based methods detect action instances in an end-to-end manner, thus saving computing resources. However, the current sparse action detectors still lack adapt- ability in feature sampling or feature decoding, thus suffer- ing from inferior accuracy or slow convergence issues. For example, building on the DETR [4] framework, TubeR [53] adaptively attends action-specific features from single-scale feature maps but perform feature transformation in a static mannner. On the contrary, though decoding sampled fea- tures with dynamic interaction heads, WOO [6] still uses the 3D RoIAlign [17] operator for feature sampling, which constrains feature sampling inside the actor bounding box and fails to take advantage of other useful information in the entire spatiotemporal feature space. Following the success of adaptive sparse object detector AdaMixer [12] in images, we present a new query-based one-stage sparse action detector, named STMixer. Our goal is to create a simple action detection framework that can sample and decode features from the complete spatiotem- poral video domain in a more flexible manner, while re- taining the benefits of sparse action detectors, such as end- to-end training and reduced computational cost. Specifi- cally, we come up with two core designs. First, to over- come the aforementioned fixed feature sampling issue, we present a query-guided adaptive feature sampling module. This new sampling mechanism endows our STMixer with the flexibility of mining a set of discriminative features from the entire spatiotemporal domain and capturing context and interaction information. Second, we devise a dual-branch feature mixing module to extract discriminative represen- tations for action detection. It is composed of an adaptive spatial mixer and an adaptive temporal mixer in parallel to focus on appearance and motion information, respectively. Coupling these two designs with a video backbone yields a simple, neat, and efficient end-to-end actor detector, which obtains a new state-of-the-art performance on the datasets of A V A [16], UCF101-24 [33], and JHMDB [19]. In sum- mary, our contribution is threefold: • We present a new one-stage sparse action detection framework in videos (STMixer). Our STMixer is easy to train in an end-to-end manner and efficient to deploy for action detection in a single stage. • We devise two flexible designs to yield a powerful ac- tion detector. The adaptive sampling can select the discriminative feature points and the adaptive feature mixing can enhance spatiotemporal representations. • STMixer achieves a new state-of-the-art performance on three challenging action detection benchmarks.
Xiao_CutMIB_Boosting_Light_Field_Super-Resolution_via_Multi-View_Image_Blending_CVPR_2023	Abstract Data augmentation (DA) is an efficient strategy for im- proving the performance of deep neural networks. Recent DA strategies have demonstrated utility in single image super-resolution (SR). Little research has, however, focused on the DA strategy for light field SR, in which multi-view information utilization is required. For the first time in light field SR, we propose a potent DA strategy called CutMIB to improve the performance of existing light field SR networks while keeping their structures unchanged. Specifically, Cut- MIB first cuts low-resolution (LR) patches from each view at the same location. Then CutMIB blends all LR patches to generate the blended patch and finally pastes the blended patch to the corresponding regions of high-resolution light field views, and vice versa. By doing so, CutMIB en- ables light field SR networks to learn from implicit geo- metric information during the training stage. Experimen- tal results demonstrate that CutMIB can improve the re- construction performance and the angular consistency of existing light field SR networks. We further verify the ef- fectiveness of CutMIB on real-world light field SR and light field denoising. The implementation code is available at https://github.com/zeyuxiao1997/CutMIB .	1. Introduction Light field cameras, which can record spatial and angular information of light rays, have rapidly become prominent imaging devices in virtual and augmented reality. Light fields are suitable for various applications, such as post- capture refocusing [35, 55], disparity estimation [52], and foreground occlusion removal [54, 69], thanks to the abun- dance of 4D spatial-angular information they contain. Com- mercialized light field cameras generally adopt micro-lens- array in front of the sensor, which poses an essential trade- off between the angular and spatial resolutions [29, 35]. Therefore, light field super-resolution (SR) has been an im- portant and popular topic. Convolutional neural network *Corresponding author. Figure 1. Comparisons on the reconstruction fidelity (PSNR, ↑) and the angular consistency (MSE, ↓) between light fields super- resolved through different methods. Following [9], we super- resolve the whole light field of the scene Bicycle from the HCI dataset to analyze the angular consistency of the super-resolved results in terms of disparity estimation using SPO [70]. Note that, CutMIB improves the values of PSNR and lowers the values of MSE by a large margin as compared to na ¨ıve light field SR meth- ods ( e.g., ATO [27], InterNet [53], IINet [33], and DPT [47]). (CNN) based and Transformer based methods have recently shown promising performance for light field SR [7, 8, 10, 26, 31, 33, 47, 52, 53, 56], outperforming traditional non- learning based methods [1, 38] with noticeable gains. This performance boost is obtained by training deep methods on external datasets. Few works have investigated data aug- mentation (DA) strategies for light field SR, which can im- prove the model performance without the need for addi- tional training datasets given that obtaining these light field data is often time-consuming and expensive [19,23,36,48]. DA has been well studied in high-level vision tasks ( e.g., image recognition, image classification, and semantic seg- mentation) for achieving better network performance and alleviating the overfitting problem [14,44,49,64,66,71]. For example, as one of the pioneering strategies, Mixup [66] This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1672 (a) CutBlur (b) CutMIB (ours) Cut Cut Blend Blend Paste Paste LR light field HR light field (input) HR light field LR light field (input) LR patches LR light field Blended patch Blend Paste HR light field BP BP BP BP BP BP BP BP BP HR HR HR HR HR HR HR HR HR Cut HR HR HR LR LR LR LR LR LR LR LR LR LR LR LR LR LR LR LR LR LR LR view 1 LR view 𝑛 HR view 1 …… …… HR view 𝑛 Cut Cut Paste Paste HR HR LR LR LR LR LR LRFigure 2. Illustrative examples of (a) CutBlur and (b) our proposed CutMIB. CutBlur generates augmented SAIs view-by-view via the “cutting-pasting” operation. CutMIB generates the augmented light field via the “cutting-blending-pasting” operation. The implicit geo- metric information can be utilized during the training stage. (a) (d) (b) (c) (e) (f) Figure 3. Analyzing CutBlur and CutMIB from a phase spectrum perspective. (a) The center view image in a 5×5light field. The red rectangle denotes the area for the cutting and pasting operation. (b) The phase spectrum of the original LR center view image. (c) is the calculated residual map between (b) and (d). (d) The phase spectrum of the LR center view image with the pasted LR patch us- ing CutBlur. We cut an HR patch from the HR center view image, and paste it to the LR center view image. (e) The phase spectrum of the LR center view image with the pasted blended patch using CutMIB. We cut all HR patches from the HR light field, blend them, and then paste the blended patch to the LR center view im- age. (f) is the calculated residual map between (b) and (e). blends two images to generate an unseen training sample. The effectiveness of the DA strategy on light field SR has received very little attention. Instead, only geometric trans- formation strategies such as flipping and rotating are used in light field SR. Recently, Yoo et al. [60] propose CutBlur, a DA strategy for training a stronger single image SR model, in which a low-resolution (LR) patch is cutandpasted to the corresponding high-resolution (HR) image region, and vice versa. A straightforward way to utilize the DA strategy on light field SR is to perform CutBlur on each view in a light field and train single image SR networks view by view, asshown in Figure 2(a). However, the ignorance of the inher- ent correlation in the spatial-angular domain makes it sub- optimal. We provide a visual observation using the phase spectrum since it contains rich texture information [46, 62] in Figure 3. Specifically, we use CutBlur on the LR center view (Figure 3(a)) in a 5 ×5 light field, cut an HR patch, and then paste it to the original LR image, and analyze the phase spectrum of the processed LR image. We can directly ob- serve from the calculated residual map in Figure 3(c) that there is little additional information from the pasted HR patch using the CutBlur strategy. This encourages us to re- alize the need for a more effective strategy to exploit patches from multiple views. Based on the aforementioned observation, we propose CutMIB, a novel DA strategy specifically designed for light field SR, as shown in Figure 2(b). Our CutMIB, which is inspired by CutBlur [60], first cuts LR patches from differ- ent views in an LR light field at the same position. The cut LR patches are then blended to generate the blended LR patch, which is then pasted to the corresponding areas of various HR light field views, and vice versa. There- fore, each augmented light field pair has partially blended LR and blended HR pixel distributions with a random ra- tio. By feeding the augmented training pairs into light field SR networks, these networks can not only learn “how” and “where” to super-resolve the LR light field ( i.e., benefit from the cutting-blending operation [60]), but also utilize the implicit geometric information in multi-view images, resulting in better performance and higher angular consis- tency among super-resolved light field views ( i.e., benefit from the blending operation [2, 5, 17]). Figure 3(f) illus- trates that pasting the blended HR patch to the LR center view (Figure 3(a)) results in more additional details in the pasted area. This demonstrates that our CutMIB can more effectively use multi-view information in a light field. Thanks to CutMIB, we can improve both the reconstruc- tion quality and the angular consistency of light field SR 1673 results while maintaining the network structures unchanged (see Figure 1). Additionally, we verify the effectiveness of the proposed CutMIB on real-world light field SR and light field denoising tasks. Contributions of this paper are summarized as follows: (1) We propose a novel DA strategy, CutMIB, to improve the performance of existing light field SR networks. To our best knowledge, it is the first DA strategy for light field SR. Through the “cutting-blending-pasting” operation, CutMIB is designed to efficiently explore the geometric information in light fields during the training stage. (2) Extensive experiments demonstrate CutMIB can boost the reconstruction fidelity and the angular consistency of existing typical light field SR methods. (3) We verify the effectiveness of CutMIB on real-world light field SR and light field denoising tasks.
Wang_F2-NeRF_Fast_Neural_Radiance_Field_Training_With_Free_Camera_Trajectories_CVPR_2023	Abstract This paper presents a novel grid-based NeRF called F2- NeRF (Fast-Free-NeRF) for novel view synthesis, which enables arbitrary input camera trajectories and only costs a few minutes for training. Existing fast grid-based NeRF training frameworks, like Instant-NGP , Plenoxels, DVGO, or TensoRF , are mainly designed for bounded scenes and rely on space warping to handle unbounded scenes. Existing two widely-used space-warping methods are only designed for the forward-facing trajectory or the 360◦object-centric trajectory but cannot process arbitrary trajectories. In this paper, we delve deep into the mechanism of space warping to handle unbounded scenes. Based on our analysis, we further propose a novel space-warping method called perspective warping, which allows us to handle arbitrary trajectories in the grid-based NeRF framework. Extensive experiments demonstrate that F2-NeRF is able to use the same perspec- tive warping to render high-quality images on two standard datasets and a new free trajectory dataset collected by us. Project page: totoro97.github.io/projects/f2-nerf .	1. Introduction The research progress of novel view synthesis has ad- vanced drastically in recent years since the emergence of the Neural Radiance Field (NeRF) [24, 42]. Once the training is done, NeRF is able to render high-quality images from novel camera poses. The key idea of NeRF is to represent the scene as a density field and a radiance field encoded by Multi-layer Perceptron (MLP) networks, and optimize the MLP networks with the differentiable volume rendering technique. Though NeRF is able to achieve photo-realistic rendering results, training a NeRF takes hours or days due to the slow optimization of deep neural networks, which limits its application scopes. Recent works demonstrate that grid-based methods, such as Plenoxels [57], DVGO [38], TensoRF [6], and Instant- *Equal contribution. (a) Forward -facing (LLFF) (b) 360 °, object -centric (NeRF -360) (c) Free trajectory ……… … … Instant -NGP Ground truth DVGO Plenoxels F2-NeRFNDC warp Inv. sphere Warp Figure 1. Top: (a) Forward-facing camera trajectory. (b) 360◦ object-centric camera trajectory. (c) Free camera trajectory. In (c), the camera trajectory is long and contains multiple foreground objects, which is extremely challenging. Bottom: Rendered images of the state-of-the-art fast NeRF training methods and F2-NeRF on a scene with a free trajectory. NGP [25], enable fast training a NeRF within a few minutes. However, the memory consumption of such grid-based rep- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4150 resentations grows in cubic order with the size of the scene. Though various techniques, such as voxel pruning [38, 57], tensor decomposition [6] or hash indexing [25], are proposed to reduce the memory consumption, these methods still can only process bounded scenes when grids are built in the original Euclidean space. To represent unbounded scenes, a commonly-adopted strategy is to use a space-warping method that maps an un- bounded space to a bounded space [3, 24, 61]. There are typically two kinds of warping functions. (1) For forward- facing scenes (Fig. 1 (a)), the Normalized Device Coordinate (NDC) warping is used to map an infinitely-far view frus- tum to a bounded box by squashing the space along the z-axis [24]; (2) For 360◦object-centric unbounded scenes (Fig. 1 (b)), the inverse-sphere warping can be used to map an infinitely large space to a bounded sphere by the sphere inversion transformation [3, 61]. Nevertheless, these two warping methods assume special camera trajectory patterns and cannot handle arbitrary ones. In particular, when a trajectory is long and contains multiple objects of interest, called free trajectories , as shown in Fig. 1 (c), the quality of rendered images degrades severely. The performance degradation on free trajectories is caused by the imbalanced allocation of spatial representation capacity. Specifically, when the trajectory is narrow and long, many regions in the scenes are empty and invisible to any input views. However, the grids of existing methods are regularly tiled in the whole scene, no matter whether the space is empty or not. Thus, much representation capacity is wasted on empty space. Although such wasting can be allevi- ated by using the progressive empty-voxel-pruning [38, 57], tensor decomposition [6] or hash indexing [25], it still causes blurred images due to limited GPU memory. Furthermore, in the visible spaces, multiple foreground objects in Fig. 1 (c) are observed with dense and near input views while back- ground spaces are only covered by sparse and far input views. In this case, for the optimal use of the spatial representation of the grid, dense grids should be allocated for the fore- ground objects to preserve shape details and coarse grids should be put in background space. However, current grid- based methods allocate grids evenly in the space, causing the inefficient use of the representation capacity. To address the above problems, we propose F2-NeRF (Fast-Free-NeRF), the first fast NeRF training method that accommodates free camera trajectories for large, unbounded scenes. Built upon the framework of Instant-NGP [25], F2- NeRF can efficiently be trained on unbounded scenes with diverse camera trajectories and maintains the fast conver- gence speed of the hash-grid representation. In F2-NeRF , we give the criterion on a proper warping function under an arbitrary camera configuration. Based on this criterion, we develop a general space-warping scheme called the perspective warping that is applicable to arbitrarycamera trajectories. The key idea of perspective warping is to first represent the location of a 3D point pby the concate- nation of the 2D coordinates of the projections of pin the input images and then map these 2D coordinates into a com- pact 3D subspace space using Principle Component Analysis (PCA) [51]. We empirically show that the proposed perspec- tive warping is a generalization of the existing NDC warp- ing [24] and the inverse sphere warping [3, 61] to arbitrary trajectories in a sense that the perspective warping is able to handle arbitrary trajectories while could automatically degenerate to these two warping functions in forward-facing scenes or 360◦object-centric scenes. In order to implement the perspective warping in a grid-based NeRF framework, we further propose a space subdivision algorithm to adap- tively use coarse grids for background regions and fine grids for foreground regions. We conduct extensive experiments on the unbounded forward-facing dataset, the unbounded 360◦object-centric dataset, and a new unbounded free trajectory dataset. The experiments show that F2-NeRF uses the same perspective warping to render high-quality images on the three datasets with different trajectory patterns. On the new Free dataset with free camera trajectories, our method outperforms base- line grid-based NeRF methods,while only using ∼12 min- utes on training on a 2080Ti GPU.
Voigtlaender_Connecting_Vision_and_Language_With_Video_Localized_Narratives_CVPR_2023	Abstract We propose Video Localized Narratives, a new form of multimodal video annotations connecting vision and lan- guage. In the original Localized Narratives [ 36], annota- tors speak and move their mouse simultaneously on an im- age, thus grounding each word with a mouse trace segment. However, this is challenging on a video. Our new protocol empowers annotators to tell the story of a video with Local- ized Narratives, capturing even complex events involving multiple actors interacting with each other and with sev- eral passive objects. We annotated 20k videos of the OVIS, UVO, and Oops datasets, totalling 1.7M words. Based on this data, we also construct new benchmarks for the video narrative grounding and video question answering tasks, and provide reference results from strong baseline models. Our annotations are available at https://google. github.io/video-localized-narratives/ .	1. Introduction Vision and language is a very active research area, which experienced much exciting progress recently [ 1,28,38,39, 48]. At the heart of many developments lie datasets con- necting still images to captions, while grounding some of the words in the caption to regions in the image, made with a variety of protocols [ 6,20,23,25,29,34,36,44,46,56]. The recent Localized Narratives (ImLNs) [ 36] offer a particu- larly attractive solution: the annotators describe an image with their voice while simultaneously moving their mouse over the regions they are describing. Speaking is natural and efﬁcient, resulting in long captions that describe the whole scene. Moreover, the synchronization between the voice and mouse pointer yields dense visual grounding for every word. Yet, still images only show one instant in time. Anno- tating videos would be even more interesting, as they show entire stories , featuring a ﬂow of events involving multiple actors and objects interacting with each other. Directly extending ImLNs to video by letting the anno- tator move their mouse and talk while the video is playing would lead to a “race against time”, likely resulting in fol- lowing only one salient object. In this paper, we proposea better annotation protocol which allows the annotator to tell the story of the video in a calm environment (Fig. 1-3). The annotators ﬁrst watch the video carefully, identify the main actors (“man”, “ostrich”), and select a few represen- tative key-frames for each actor. Then, for each actor sep- arately, the annotators tell a story: describing the events it is involved in using their voice, while moving the mouse on the key-frames over the objects and actions they are talking about. The annotators mention the actor name, its attributes, and especially the actions it performs, both on other ac- tors ( e.g., “play with the ostrich”) and on passive objects (e.g., “grabs the cup of food”). For completeness, the an- notators also brieﬂy describe the background in a separate step (bottom row). Working on keyframes avoids the race against time, and producing a separate narration for each ac- tor enables disentangling situations and thus cleanly captur- ing even complex events involving multiple actors interact- ing with each other and with several passive objects. As in ImLN, this protocol localizes each word with a mouse trace segment. We take several additional measures to obtain ac- curate localizations, beyond what was achieved in [ 36]. We annotated the OVIS [ 37], UVO [ 47], and Oops [ 12] datasets with Video Localized Narratives (VidLNs). These datasets cover a general domain, as opposed to only cook- ing [ 10,62] or ﬁrst-person videos [ 16]. Moreover, these videos contain complex scenes with interactions between multiple actors and passive objects, leading to interesting stories that are captured by our rich annotations. In to- tal our annotations span 20k videos, 72k actors, and 1.7 million words. On average, each narrative features tran- scriptions for 3.5actors, totaling 75.1words, including 23.0 nouns,9.5verbs, and 8.5adjectives (Fig. 4). Our analysis demonstrates the high quality of the data. The text descrip- tions almost always mention objects/actions that are actu- ally present in the video, and the mouse traces do lie inside their corresponding object in most cases. The richness of our VidLNs makes them a strong ba- sis for several tasks. We demonstrate this by creating new benchmarks for the tasks of Video Narrative Grounding (VNG) and Video Question Answering (VideoQA). The new VNG task requires a method to localize the nouns in This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2461 1. Watch 2./3. Select actors and key-frames 5. Transcription 4. Speak and mo ve mouse "A man wearing a black t-shirt is holding a cup of food in his right hand . He moves around a piece of food in his left hand to play with the ostrich ." A man wearing a black t-shirt is holding a cup of food in his right hand. He moves around a piece of food in his left hand to play with the ostrich. Figure 1. The ﬁve steps of the annotation process for VidLNs. We are able to cleanly describe even complex interactions and events by disentangling the storylines of different actors. <Man> A man wearing a black t-shirt is holding a cup of food in his right hand. He moves around a piece of food in his left hand to play with the ostrich. <Ostrich> An ostrich is looking at the piece of food held by the man and suddenly grabs the cup of food and starts eating. <Background> In the background, there are hills, white barriers, a ﬂag, the sky, and soil on the ground.Man Ostrich Background Figure 2. An example VidLN annotation. Each row shows the story of the video from the perspective of a different actor. Each mouse trace segment drawn on selected key-frames localizes the word highlighted in the same color. More examples are in the supplement. an input narrative with a segmentation mask on the video frames. This is a complex task that goes beyond open- vocabulary semantic segmentation [ 14], as often the text has multiple identical nouns that need to be disambiguated us- ing the context provided by other words. We construct two VNG benchmarks on OVIS and UVO, with a total of 8k videos involving 45k objects. We also build baseline mod- els that explicitly attempt this disambiguation and report ex- periments showing it’s beneﬁcial on this dataset, while also highlighting these new benchmarks are far from solved. For VideoQA, we construct a benchmark on the Oops dataset, consisting of text-output questions ( e.g., “What is the person in the blue suit doing?”), where the answer is free-form text ( e.g., “paragliding”), and location-output questions ( e.g., “Where is the woman that is wearing a black and white dress?”), where the answer is a spatio-temporal location in the video. The benchmark features a total of 62k questions on 9.5k videos. Answering many of these ques- tions requires a deep understanding of the whole video. Wealso implement baseline methods to establish initial results on this task.
Wang_Towards_Transferable_Targeted_Adversarial_Examples_CVPR_2023	Abstract Transferability of adversarial examples is critical for black-box deep learning model attacks. While most existing studies focus on enhancing the transferability of untargeted adversarial attacks, few of them studied how to generate transferable targeted adversarial examples that can mislead models into predicting a specific class. Moreover, existing transferable targeted adversarial attacks usually fail to suf- ficiently characterize the target class distribution, thus suf- fering from limited transferability. In this paper, we pro- pose the Transferable Targeted Adversarial Attack (TTAA), which can capture the distribution information of the tar- get class from both label-wise and feature-wise perspec- tives, to generate highly transferable targeted adversarial examples. To this end, we design a generative adversar- ial training framework consisting of a generator to produce targeted adversarial examples, and feature-label dual dis- criminators to distinguish the generated adversarial exam- ples from the target class images. Specifically, we design the label discriminator to guide the adversarial examples to learn label-related distribution information about the target class. Meanwhile, we design a feature discrimina- tor, which extracts the feature-wise information with strong cross-model consistency, to enable the adversarial exam- ples to learn the transferable distribution information. Fur- thermore, we introduce the random perturbation dropping to further enhance the transferability by augmenting the di- versity of adversarial examples used in the training process. Experiments demonstrate that our method achieves excel- lent performance on the transferability of targeted adver- sarial examples. The targeted fooling rate reaches 95.13% when transferred from VGG-19 to DenseNet-121, which significantly outperforms the state-of-the-art methods. ∗Zhibo Wang is the corresponding author.	1. Introduction As an important branch of artificial intelligence (AI), deep neural networks (DNNs) contribute to many real- life applications, e.g., image classification [1, 2], speech recognition [3, 4], face detection [5, 6], automatic driv- ing technology [7], etc. Such broad impacts have mo- tivated a wide range of investigations into the adversar- ial attacks on DNNs, exploring the vulnerability and un- certainty of DNNs. For instance, Szegedy et al . showed that DNNs could be fooled by adversarial examples crafted by adding human-indistinguishable perturbations to origi- nal inputs [8]. As successful adversarial examples must be imperceptible to humans but cause DNNs to make a false prediction, how to design adversarial attacks to generate high-quality adversarial examples in a general manner re- mains challenging. Adversarial attack methods can be divided into untar- geted attacks and targeted attacks. Untargeted attacks try to misguide the model to predict arbitrary incorrect la- bels, while targeted adversarial attacks expect the gener- ated adversarial examples can trigger the misprediction for a specific label. The transferability of adversarial examples is crucial for both untargeted and targeted attacks, espe- cially in black-box attack scenarios where the target model is inaccessible. However, most existing studies focus on enhancing the transferability of untargeted adversarial at- tacks, through data augmentation [9, 10], model aggrega- tion [11,12], feature information utilization [13–15], or gen- erative methods [16–18]. Although some of them could be extended to the targeted adversarial attacks by simply mod- ifying the loss function, they could not extract sufficient transferable information about the target class due to the overfitting of the source model and the lack of distribution information of the target class, thus demonstrating limited transferability. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20534 Some recent works [15, 19–21] studied how to boost the transferability of targeted adversarial attacks by learning the target class information from either the label or the feature perspective, leveraging the label probability distribution and feature maps respectively. The label-wise information, of- ten output by the last layer of the classification model, can effectively reflect the direct correlation between image dis- tribution and class labels. However, learning with just the label-wise information is proven to retain high-level seman- tic information of the original class [22], thus leading to low cross-model transferability. The feature-wise information, which can be obtained from the intermediate layer of the classification model, has been proven [23] to have transfer- ability due to the mid-level layer of different DNNs follow- ing similar activation patterns. However, the feature-wise information is not sensitive to the target label and fails to trigger the targeted misclassification. In summary, only the information extracted from the la- bel or the feature cannot generate high-transferability tar- geted adversarial examples. Meanwhile, most existing tar- geted attacks assume that the training dataset of the target model ( i.e., target domain) is accessible and could be lever- aged by the attackers to train a shadow model in the target domain as the simulation of the target model, which how- ever is not practical in realistic scenarios. In this paper, we aim to generate highly transferable tar- geted adversarial examples in a more realistic but challeng- ing scenario where the attacker cannot access the data of the target domain. To this end, we are facing two main challenges. The first challenge is how to generate targeted adversarial examples with both cross-model and cross- domain transferability? Cross-domain images usually vary significantly in characteristic distribution (e.g., have differ- ent attributes even labels), making it difficult for models to transfer the knowledge to unknown domains. Besides, when involved models adopt different architectures, do- main knowledge learned by the source model becomes less transferable, resulting in more difficulties in cross-model and cross-domain transferable adversarial attacks. The sec- ond challenge is how to improve the transferability of tar- geted adversarial attacks? As the targeted adversarial at- tack needs to distort the ground-truth label and trigger the model to predict the target label simultaneously, causing the transferability of the targeted attack hard to achieve due to the dual objective of obfuscating original class information and recognizing target class information. To solve such challenges, we propose Transferable Tar- geted Adversarial Attack (TTAA) to generate highly trans- ferable targeted adversarial examples. The main idea of our method is to capture the distribution information of the target class from both label-wise and feature-wise perspec- tives. We design a generative adversarial network, which generates targeted adversarial examples by a generator andcaptures the distribution information of the target class by label-feature dual discrimination, consisting of the label dis- criminator and the feature discriminator. More specifically, the label discriminator learns the label-related distribution information from the label probability distribution and the feature discriminator extracts transferable distribution in- formation of the target class via feature maps output by the intermediate layer of the label discriminator. Meanwhile, we propose random perturbation dropping to enhance our training samples, which applies random transformations to augment the diversity of the adversarial examples used dur- ing the training process to improve the robustness of the distribution information of the target class on the adversar- ial examples. In generative adversarial training, such dis- tribution information extracted by discriminators guides the generator to acquire the label-related and transferable dis- tribution information of the target class. Therefore the tar- geted adversarial examples achieve both high cross-model and cross-domain transferability. Our main contributions are summarized as follows. • We propose a general framework for targeted adver- sarial attack that works in both non-cross-domain and cross-domain scenarios. This framework could be eas- ily integrated with other targeted adversarial attacks to improve their cross-model and cross-domain targeted transferability. • Our proposed Transferable Targeted Adversarial At- tack could extract the target class distribution from feature-wise and label-wise levels to promote pertur- bations to acquire label-related and transferable distri- bution information of the target class, thus generating highly transferable targeted adversarial examples. • Extensive experiments on diverse classification models demonstrate the superior targeted transferability of ad- versarial examples generated by the proposed TTAA as compared to state-of-the-art transferable attacking methods, no matter whether the attack scenario is cross-domain or non-cross-domain.
Wang_Semi-Supervised_Parametric_Real-World_Image_Harmonization_CVPR_2023	Abstract Learning-based image harmonization techniques are usu- ally trained to undo synthetic random global transforma- tions applied to a masked foreground in a single ground truth photo. This simulated data does not model many of the important appearance mismatches (illumination, object boundaries, etc.) between foreground and background in real composites, leading to models that do not generalize well and cannot model complex local changes. We propose a new semi-supervised training strategy that addresses this problem and lets us learn complex local appearance harmo- nization from unpaired real composites, where foreground and background come from different images. Our model is fully parametric. It uses RGB curves to correct the global colors and tone and a shading map to model local vari- ations. Our method outperforms previous work on estab- lished benchmarks and real composites, as shown in a userstudy, and processes high-resolution images interactively. Code, and project page available at: https://kewang0622.github.io/sprih/ .	1. Introduction Image harmonization [12, 22, 23, 26, 28, 32] aims to iron out visual inconsistencies created when compositing a fore- ground subject onto a background image that was captured under different conditions [18, 32], by altering the fore- ground’s colors, tone, etc., to make the composite more re- alistic. Despite significant progress, the practicality of to- day’s most sophisticated learning-based image harmoniza- tion techniques [3, 4, 9, 10, 13, 14, 16, 32] is limited by a se- vere domain gap between the synthetic data they are trained on and real-world composites. As shown in Figure 2, the standard approach to generat- ing synthetic training composites applies global transforms This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5927 Global adjustmentsReal image Paste foreground F BSyntheticcompositeSyntheticcomposite Same lighting environmentConsistentshadingPerfectboundaryImperfectboundaryInconsistent shadingBackgroundRealcomposite RealcompositeDifferent lighting environment Lighting source of Foregroundand Background F BSynthetic composite imageReal-world composite imageFigure 2. Domain Gap between synthetic and real-world com- posites. The existing synthetic composites [4] (left), generated by applying global transforms (e.g., color, brightness), are unable to simulate many of the appearance mismatches that occur in real composites (right). This leads to a domain gap: models trained on synthetic data do not generalize well to real composites. In real composites (right), the foreground and background are captured under different conditions. They have different illuminations, the shadows do not match, and the object’s boundary is inconsistent. Such mismatches do not happen in the synthetic case (left). (color, brightness, contrast, etc.) to a masked foreground subject in a ground truth photo. This is how the iHarmony Dataset [2, 4] was constructed. A harmonization network is then trained to recover the ground truth image from the syn- thetic input. While this approach makes supervised training possible, it is unsatisfying in simulating the real composite in that synthetic data does not simulate mismatch in illu- mination, shadows, shading, contacts, perspective, bound- aries, and low-level image statistics like noise, lens blur, etc. However, in real-world composites, the foreground subject and the background are captured under different conditions, which can have more diverse and arbitrary differences in any aspects mentioned above. We argue that using realistic composites for training is essential for image harmonization to generalize better to real-world use cases. Because collecting a large dataset of artist-created before/after real composite pairs would be costly and cumbersome, our strategy is to use a semi- supervised approach instead. We propose a novel dual- stream training scheme that alternates between two data streams. Similar to previous work, the first is a supervised training stream, but crucially, it uses artist-retouched image pairs. Different from previous datasets, these artistic adjust- ments include global color editing but also dodge and burn shading corrections and other local edits. The second stream is fully unsupervised. It uses a GAN [8] training procedure, in which the critic comparesour harmonized results with a large dataset of realistic im- age composites. Adversarial training requires no paired ground truth. The foreground and background for the com- posite in this dataset are extracted from different images so that their appearance mismatch is consistent with what the model would see at test time. To reap the most benefits from our semi-supervised train- ing, we also introduce a new model that is fully paramet- ric. To process a high-resolution input composite at test time, our proposed network first creates a down-sampled copy of the image at 512×512resolution, from which it predicts global RGB curves and a smooth, low-resolution shading map . We then apply the RGB curves pointwise to the high-resolution input and multiply them by the upsam- pled shading map. The shading map enables more realistic local tonal variations, unlike previous harmonization meth- ods limited to global tone and color changes, either by con- struction [14, 16, 31] or because of their training data [4]. Our parametric approach offers several benefits. First, by restricting the model’s output space, it regularizes the adver- sarial training. Unrestricted GAN generators often create spurious image artifacts or other unrealistic patterns [36]. Second, it exposes intuitive controls for an artist to adjust and customize the harmonization result post-hoc. This is unlike the black-box nature of most current learning-based approaches [3, 4, 9, 10], which output an image directly. And, third our parametric model runs at an interactive rate, even on very high-resolution images (e.g., 4k), whereas sev- eral state-of-the-art methods [4, 9, 10] are limited to low- resolution (e.g., 256×256) inputs. To summarize, we make the following contributions: • A novel dual-stream semi-supervised training strategy that, for the first time, enables training from real com- posites, which contains much richer local appearance mismatches between foreground and background. • A parametric harmonization method that can capture these more complex, local effects (using our shading map) and produces more diverse and photorealistic harmonization results. • State-of-the-art results on both synthetic and real com- posite test sets in terms of quantitative results and visual comparisons, together with a new evaluation benchmark.
Xie_Revealing_the_Dark_Secrets_of_Masked_Image_Modeling_CVPR_2023	Abstract Masked image modeling (MIM) as pre-training is shown to be effective for numerous vision downstream tasks, but how and where MIM works remain unclear. In this paper, we compare MIM with the long-dominant supervised pre- trained models from two perspectives, the visualizations and the experiments, to uncover their key representational dif- ferences. From the visualizations, we find that MIM brings locality inductive bias to all layers of the trained models, but supervised models tend to focus locally at lower layers but more globally at higher layers. That may be the rea- son why MIM helps Vision Transformers that have a very large receptive field to optimize. Using MIM, the model can maintain a large diversity on attention heads in all lay- ers. But for supervised models, the diversity on attention heads almost disappears from the last three layers and less diversity harms the fine-tuning performance. From the exper- iments, we find that MIM models can perform significantly better on geometric and motion tasks with weak semantics or fine-grained classification tasks, than their supervised counterparts. Without bells and whistles, a standard MIM pre-trained SwinV2-L could achieve state-of-the-art perfor- mance on pose estimation (78.9 AP on COCO test-dev and 78.0 AP on CrowdPose), depth estimation (0.287 RMSE on NYUv2 and 1.966 RMSE on KITTI), and video object track- ing (70.7 SUC on LaSOT). For the semantic understanding datasets where the categories are sufficiently covered by the supervised pre-training, MIM models can still achieve highly competitive transfer performance. With a deeper understand- ing of MIM, we hope that our work can inspire new and solid research in this direction. Code will be available at https: //github.com/zdaxie/MIM-DarkSecrets .	1. Introduction Pre-training of effective and general representations ap- plicable to a wide range of tasks in a domain is the key to the success of deep learning. In computer vision, supervised *Equal Contribution. The work is done when Zhenda Xie, Zigang Geng, and Jingcheng Hu are interns at Microsoft Research Asia.†Contact person.classification on ImageNet [14] has long been the dominant pre-training task which is manifested to be effective on a wide range of vision tasks, especially on the semantic un- derstanding tasks, such as image classification [17, 18, 36, 38, 51], object detection [23, 29, 61, 63], semantic segmenta- tion [53, 69], video action recognition [7, 52, 65, 67] and so on. Over the past several years, “masked signal modeling”, which masks a portion of input signals and tries to predict these masked signals, serves as a universal and effective self- supervised pre-training task for various domains, including language, vision, and speech. After (masked) language mod- eling repainted the NLP field [15, 49], recently, such task has also been shown to be a competitive challenger to the su- pervised pre-training in computer vision [3, 8, 18, 27, 74, 80]. That is, masked image modeling (MIM) pre-trained models achieve very high fine-tuning accuracy on a wide range of vision tasks of different nature and complexity. However, there still remain several questions: 1.What are the key mechanisms that contribute to the excellent performance of MIM? 2.How transferable are MIM and supervised models across different types of tasks, such as semantic un- derstanding, geometric and motion tasks? To investigate these questions, we compare MIM with su- pervised models from two perspectives, the visualization perspective and the experimental perspective, trying to un- cover key representational differences between these two pre-training tasks and deeper understand the behaviors of MIM pre-training. We start with studying the attention maps of the pre- trained models. Firstly, we visualize the averaged attention distance in MIM models, and we find that masked image modeling brings locality inductive bias to the trained model, that the models tend to aggregate near pixels in part of the attention heads, and the locality strength is highly correlated with the masking ratio and masked patch size in the pre-training stage. But the supervised models tend to focus locally at lower layers but more globally at higher layers. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14475 We next probe how differently the attention heads in MIM trained Transformer behave. We find that different atten- tion heads tend to aggregate different tokens on all layers in MIM models , according to the large KL-divergence on attention maps of different heads. But for supervised models, the diversity on attention heads diminishes as the layer goes deeper and almost disappears in the last three layers. We drop the last several layers for supervised pre-trained models during fine-tuning and find that it benefits the fine-tuning per- formance on downstream tasks, however this phenomenon is not observed for MIM models. That is, less diversity on attention heads would somewhat harm the performance on downstream tasks . Then we examine the representation structures in the deep networks of MIM and supervised models via the simi- larity metric of Centered Kernel Alignment (CKA) [37]. We surprisingly find that in MIM models, the feature represen- tations of different layers are of high similarity, that their CKA values are all very large (e.g., [0.9, 1.0]). But for supervised models, as in [59], different layers learn different representation structures, that their CKA similarities vary greatly (e.g., [0.5,1.0]). To further verify this, we load the pre-trained weights of randomly shuffled layers during fine- tuning and find that supervised pre-trained models suffer more than the MIM models. From the experimental perspective, a fundamental pre- training task should be able to benefit a wide range of tasks, or at least it is important to know for which types of tasks MIM models work better than the supervised counterparts. To this end, we conduct a large-scale study by comparing the fine-tuning performance of MIM and supervised pre-trained models, on three types of tasks, semantic understanding tasks, geometric and motion tasks, and the combined tasks which simultaneously perform both. For semantic understanding tasks, we select several rep- resentative and diverse image classification benchmarks, in- cluding Concept Generalization (CoG) benchmark [62], the widely-used 12-dataset benchmark [38], as well as a fine- grained classification dataset iNaturalist-18 [68]. For the classification datasets whose categories are sufficiently cov- ered by ImageNet categories (e.g. CIFAR-10/100), super- vised models can achieve better performance than MIM models. However, for other datasets, such as fine-grained classification datasets (e.g., Food, Birdsnap, iNaturalist), or datasets with different output categories (e.g., CoG), most of the representation power in supervised models is diffi- cult to transfer, thus MIM models remarkably outperform supervised counterparts. For geometric and motion tasks that require weaker se- mantics and high-resolution object localization capabilities, such as pose estimation on COCO [48] and CrowdPose [44], depth estimation on NYUv2 [64] and KITTI [22], and video object tracking on GOT10k [32], TrackingNet [55], and La-SOT [20], MIM models outperform supervised counterparts by large margins. Note that, without bells and whistles, Swin-L with MIM pre-training could achieve state-of-the-art performance on these benchmarks, e.g., 80.5AP on COCO val,78.9AP on COCO test-dev, and 78.0AP on Crowd- Pose of pose estimation, 0.287RMSE on NYUv2 and 1.966 RMSE on KITTI of depth estimation, and 70.7SUC on LaSOT of video object tracking. We select object detection on COCO as the combined task which simultaneously performs both semantic understanding and geometric learning. For object detection on COCO, MIM models would outperform supervised counterparts. Via investigating the training losses of object classification and localization, we find that MIM models help localization task converge faster, and supervised models benefit more for object classification, that categories of COCO are fully covered by ImageNet. In general, MIM models tend to exhibit improved perfor- mance on geometric/motion tasks with weak semantics or fine-grained classification tasks compared to their supervised counterparts. For tasks/datasets where supervised models excel in transfer, MIM models can still achieve competitive transfer performance. Masked image modeling appears to be a promising candidate for a general-purpose pre-trained model. We hope our paper contributes to this understanding within the community and stimulates further research in this direction.
Wu_Virtual_Sparse_Convolution_for_Multimodal_3D_Object_Detection_CVPR_2023	Abstract Recently, virtual/pseudo-point-based 3D object detec- tion that seamlessly fuses RGB images and LiDAR data by depth completion has gained great attention. However, virtual points generated from an image are very dense, in- troducing a huge amount of redundant computation during detection. Meanwhile, noises brought by inaccurate depth completion significantly degrade detection precision. This paper proposes a fast yet effective backbone, termed Vir- ConvNet , based on a new operator VirConv (Virtual Sparse Convolution), for virtual-point-based 3D object detection. VirConv consists of two key designs: (1) StVD (Stochas- tic Voxel Discard) and (2) NRConv (Noise-Resistant Sub- manifold Convolution). StVD alleviates the computation problem by discarding large amounts of nearby redundant voxels. NRConv tackles the noise problem by encoding voxel features in both 2D image and 3D LiDAR space. By integrating VirConv, we first develop an efficient pipeline VirConv-L based on an early fusion design. Then, we build a high-precision pipeline VirConv-T based on a trans- formed refinement scheme. Finally, we develop a semi- supervised pipeline VirConv-S based on a pseudo-label framework. On the KITTI car 3D detection test leader- board, our VirConv-L achieves 85% AP with a fast run- ning speed of 56ms . Our VirConv-T and VirConv-S attains a high-precision of 86.3% and87.2% AP , and currently rank 2nd and 1st1, respectively. The code is available at https://github.com/hailanyi/VirConv .	1. Introduction 3D object detection plays a critical role in autonomous driving [32, 45]. The LiDAR sensor measures the depth of scene [4] in the form of a point cloud and enables re- liable localization of objects in various lighting environ- ments. While LiDAR-based 3D object detection has made rapid progress in recent years [19, 23, 25, 27, 28, 42, 43, 49], its performance drops significantly on distant objects, which inevitably have sparse sampling density in the scans. Unlike *Corresponding author 1On the date of CVPR deadline, i.e., Nov.11, 2022 Inference speed (ms) 40 60 80 100 78 80 82 84 86 LiDAR-only Multimodal Ours SE-SSD Voxel-RCNN VirConv-L VirConv-T UberATG-MMF Fast PointRCNN 3D-CVF Focals Conv SFD VPFNet Graph-VoI PV-RCNN CasA 40 60 80 100 88 89 90 91 92 93 SE-SSD Voxel-RCNN VirConv-L VirConv-T UberATG-MMF Fast PointRCNN 3D-CVF Focals Conv SFD VPFNet Graph-VoI PV-RCNN CasA 3D AP (%) BEV AP (%) Inference speed (ms)Figure 1. Our VirConv-T achieves top average precision (AP) on both 3D and BEV moderate car detection in the KITTI benchmark (more details are in Table 1). Our VirConv-L runs fast at 56ms with competitive AP. LiDAR scans, color image sensors provide high-resolution sampling and rich context data of the scene. The RGB im- age and LiDAR data can complement each other and usu- ally boost 3D detection performance [1, 6, 20, 21, 24]. Early methods [29–31] extended the features of LiDAR points with image features, such as semantic mask and 2D CNN features. They did not increase the number of points; thus, the distant points still remain sparse. In contrast, the methods based on virtual/pseudo points (for simplic- ity, both denoted as virtual points in the following) enrich the sparse points by creating additional points around the LiDAR points. For example, MVP [45] creates the vir- tual points by completing the depth of 2D instance points from the nearest 3D points. SFD [36] creates the virtual points based on depth completion networks [16]. The vir- tual points complete the geometry of distant objects, show- ing the great potential for high-performance 3D detection. However, virtual points generated from an image are generally very dense. Taking the KITTI [9] dataset as an example, an 1242 ×375 image generates 466k virtual points (∼27×more than the LiDAR scan points). This brings a huge computational burden and causes a severe efficiency issue (see Fig. 2 (f)). Previous work addresses the density problem by using a larger voxel size [19, 44] or by ran- domly down-sampling [17] the points. However, applying such methods to virtual points will inevitably sacrifice use- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21653 Distance 0 1 2 3 4 AP improvement LiDAR points per object Virtual points per object 97% vir. points 0.18% AP gain 3% vir. points 2.2% AP gain AP comparison Late fusion Virtual points LiDAR-only Early fusion Speed comparison (e) (f) (b) (d) 0 10 20 30 40 40 50 60 70 75 80 85 90 0 40 80 120 160 0 5k 10k 15k 20k 25k 30k (a) (c) 3D AP (%) Time (ms) Number of points 3D AP (%)Figure 2. The noise problem and density problem of virtual points. (a) Virtual points in 3D space. (b) Virtual points in 2D space. (c) Noises (red) in 3D space. (d) Noises (red) distributed on 2D in- stance boundaries. (e) Virtual points number versus AP improve- ment along different distances by using V oxel-RCNN [7] with late fusion (details see Sec. 3.1). (f) Car 3D AP and inference time us- ing V oxel-RCNN [7] with LiDAR-only, virtual points-only, early fusion, and late fusion (details see Sec. 3.1), respectively. ful shape cues from faraway points and result in decreased detection accuracy. Another issue is that the depth completion can be inac- curate, and it brings a large amount of noise in the virtual points (see Fig. 2 (c)). Since it is very difficult to distinguish the noises from the background in 3D space, the localiza- tion precision of 3D detection is greatly degraded. In addi- tion, the noisy points are non-Gaussian distributed, and can not be filtered by conventional denoising algorithms [8,12]. Although recent semantic segmentation network [15] show promising results, they generally require extra annotations. To address these issues, this paper proposes a VirCon- vNet pipeline based on a new Virtual Sparse Convolution (VirConv) operator. Our design builds on two main obser- vations . (1) First, geometries of nearby objects are often relatively complete in LiDAR scans. Hence, most virtual points of nearby objects only bring marginal performance gain (see Fig. 2 (e)(f)), but increase the computational cost significantly. (2) Second, noisy points introduced by inac- curate depth completions are mostly distributed on the in- stance boundaries (see Fig. 2 (d)). They can be recognized in 2D images after being projected onto the image plane. Based on these two observations, we design a StVD (Stochastic V oxel Discard) scheme to retain those most im- portant virtual points by a bin-based sampling, namely, dis- carding a huge number of nearby voxels while retaining far- away voxels. This can greatly speed up the network com- putation. We also design a NRConv (Noise-Resistant Sub- manifold Convolution) layer to encode geometry features of voxels in both 3D space and 2D image space. The ex-tended receptive field in 2D space allows our NRConv to distinguish the noise pattern on the instance boundaries in 2D image space. Consequently, the negative impact of noise can be suppressed. We develop three multimodal detectors to demon- strate the superiority of our VirConv: (1) a lightweight VirConv-L constructed from V oxel-RCNN [7]; (2) a high- precision VirConv-T based on multi-stage [34] and multi- transformation [35] design; (3) a semi-supervised VirConv- Sbased on a pseudo-label [33] framework. The effective- ness of our design is verified by extensive experiments on the widely used KITTI dataset [9] and nuScenes dataset [3]. Our contributions are summarized as follows: • We propose a VirConv operator, which effectively en- codes voxel features of virtual points by StVD and NRConv . The StVD discards a huge number of redun- dant voxels and substantially speeds up the 3D detec- tion prominently. The NRConv extends the receptive field of 3D sparse convolution to the 2D image space and significantly reduces the impact of noisy points. • Built upon VirConv, we present three new multimodal detectors: a VirConv-L , aVirConv-T , and a semi- supervised VirConv-S for efficient, high-precision, and semi-supervised 3D detection, respectively. • Extensive experiments demonstrated the effectiveness of our design (see Fig. 1). On the KITTI leaderboard, our VirConv-T and VirConv-S currently rank 2nd and 1st, respectively. Our VirConv-L runs at 56ms with competitive precision.
Wang_Selective_Structured_State-Spaces_for_Long-Form_Video_Understanding_CVPR_2023	Abstract Effective modeling of complex spatiotemporal dependencies in long-form videos remains an open problem. The recently proposed Structured State-Space Sequence (S4) model with its linear complexity offers a promising direction in this space. However, we demonstrate that treating all image- tokens equally as done by S4 model can adversely affect its efficiency and accuracy. To address this limitation, we present a novel Selective S4 (i.e., S5) model that employs a lightweight mask generator to adaptively select infor- mative image tokens resulting in more efficient and accu- rate modeling of long-term spatiotemporal dependencies in videos. Unlike previous mask-based token reduction meth- ods used in transformers, our S5 model avoids the dense self-attention calculation by making use of the guidance of the momentum-updated S4 model. This enables our model to efficiently discard less informative tokens and adapt to various long-form video understanding tasks more effec- tively. However, as is the case for most token reduction methods, the informative image tokens could be dropped in- correctly. To improve the robustness and the temporal hori- zon of our model, we propose a novel long-short masked contrastive learning (LSMCL) approach that enables our model to predict longer temporal context using shorter in- put videos. We present extensive comparative results using three challenging long-form video understanding datasets (LVU, COIN and Breakfast), demonstrating that our ap- proach consistently outperforms the previous state-of-the- art S4 model by up to 9.6%accuracy while reducing its memory footprint by 23%.	1. Introduction Video understanding is an active research area where a variety of different models have been explored including e.g., two-stream networks [19, 20, 52], recurrent neural net- works [3, 63, 72] and 3-D convolutional networks [59–61]. However, most of these methods have primarily focused on short-form videos that are typically with a few seconds in length, and are not designed to model the complex long- eggsaucepanstovepickwhiskpouractions:objects: timecook eggsmake cereal•••Figure 1. Illustration of long-form videos – Evenly sampled frames from two long-form videos, that have long duration (more than 1 minute) and distinct categories in the Breakfast [36] dataset (grayscale frames are shown for better visualization). The video on top shows the activity of making scrambled eggs, while the one on the bottom shows the activity of making cereal. These two videos heavily overlap in terms of objects ( e.g., eggs, saucepan and stove), and actions ( e.g., picking, whisking and pouring). To effectively distinguish these two videos, it is important to model long-term spatiotemporal dependencies, which is also the key in long-form video understanding. term spatiotemporal dependencies often found in long-form videos (see Figure 1 for an illustrative example). The re- cent vision transformer (ViT) [14] has shown promising ca- pability in modeling long-range dependencies, and several variants [1,4,15,41,45,49,65] have successfully adopted the transformer architecture for video modeling. However, for a video with T frames and S spatial tokens, the complexity of standard video transformer architecture is O(S2T2), which poses prohibitively high computation and memory costs when modeling long-form videos. Various attempts [54,68] have been proposed to improve this efficiency, but the ViT pyramid architecture prevents them from developing long- term dependencies on low-level features. In addition to ViT, a recent ViS4mer [29] method has tried to apply the Structured State-Spaces Sequence (S4) model [23] as an effective way to model the long-term video dependencies. However, by introducing simple mask- ing techniques we empirically reveal that the S4 model can have different temporal reasoning preferences for different downstream tasks. This makes applying the same image to- ken selection method as done by ViS4mer [29] for all long- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6387 form video understanding tasks suboptimal. To address this challenge, we propose a cost-efficient adaptive token selection module, termed S 5(i.e., selective S4) model, which adaptively selects informative image to- kens for the S4 model, thereby learning discriminative long- form video representations. Previous token reduction meth- ods for efficient image transformers [37, 42, 50, 66, 70, 71] heavily rely on a dense self-attention calculation, which makes them less effective in practice despite their theoret- ical guarantees about efficiency gains. In contrast, our S5 model avoids the dense self-attention calculation by lever- aging S4 features in a gumble-softmax sampling [30] based mask generator to adaptively select more informative im- age tokens. Our mask generator leverages S4 feature for its global sequence-context information and is further guided by the momentum distillation from the S4 model. To further improve the robustness and the temporal pre- dictability of our S5 model, we introduce a novel long-short mask contrastive learning (LSMCL) to pre-train our model. In LSMCL, randomly selected image tokens from long and short clips include the scenario that the less informative im- age tokens are chosen, and the representation of them are learned to match each other. As a result, the LSMCL not only significantly boosts the efficiency compared to the pre- vious video contrastive learning methods [17, 51, 64], but also increases the robustness of our S5 model when deal- ing with the mis-predicted image tokens. We empirically demonstrate that the S5 model with LSMCL pre-training can employ shorter-length clips to achieve on-par perfor- mance with using longer-range clips without incorporating LSMCL pre-training. We summarize our key contributions as the following: •We propose a Selective S4 (S5) model that leverages the global sequence-context information from S4 features to adaptively choose informative image tokens in a task- specific way. •We introduce a novel long-short masked contrastive learn- ing approach (LSMCL) that enables our model to be tol- erant to the mis-predicted tokens and exploit longer dura- tion spatiotemporal context by using shorter duration input videos, leading to improved robustness in the S5 model. •We demonstrate that two proposed novel techniques (S5 model and LSMCL) are seamlessly suitable and effective for long-form video understanding, achieving the state-of- the-art performance on three challenging benchmarks. No- tably, our method achieves up to 9.6% improvement on LVU dataset compared to the previous state-of-the-art S4 method, while reducing the memory footprint by 23% .
Xie_Exploring_and_Exploiting_Uncertainty_for_Incomplete_Multi-View_Classification_CVPR_2023	Abstract Classifying incomplete multi-view data is inevitable since arbitrary view missing widely exists in real-world applica- tions. Although great progress has been achieved, existing incomplete multi-view methods are still difficult to obtain a trustworthy prediction due to the relatively high uncertainty nature of missing views. First, the missing view is of high uncertainty, and thus it is not reasonable to provide a single deterministic imputation. Second, the quality of the imputed data itself is of high uncertainty. To explore and exploit the uncertainty, we propose an Uncertainty-induced Incomplete Multi-View Data Classification (UIMC) model to classify the incomplete multi-view data under a stable and reliable framework. We construct a distribution and sample multiple times to characterize the uncertainty of missing views, and adaptively utilize them according to the sampling quality. Accordingly, the proposed method realizes more perceivable imputation and controllable fusion. Specifically, we model each missing data with a distribution conditioning on the available views and thus introducing uncertainty. Then an evidence-based fusion strategy is employed to guarantee the trustworthy integration of the imputed views. Extensive ex- periments are conducted on multiple benchmark data sets and our method establishes a state-of-the-art performance in terms of both performance and trustworthiness.	1. Introduction Learning from multiple complementary views has the potential to yield more generalizable models. Benefiting from the power of deep learning, multi-view learning has further exhibited remarkable benefits against the single-view paradigm in clustering [1 –3], classification [4, 5] and rep- resentation learning [6, 7]. However, real-world data are usually incomplete. For instance, in the medical field, pa- tients with same condition may choose different medical examinations producing incomplete/unaligned multi-view ‡Equal contribution. ∗Corresponding author.data; similarly, sensors in cars at times may be out of order and thus only part of information can be collected. Flexible and trustworthy utilization of incomplete multi-view data is still very challenging due to the high uncertainty of missing issues. For the task of incomplete multi-view classification (IMVC), there have been plenty of studies which could be roughly categorized into two main lines. The methods [8, 9] only use available views without imputation to conduct clas- sification. While the other line [10 –13] reconstructs miss- ing data based on deep learning methods such as autoen- coder [14,15] or generative adversarial network (GAN) [16], and then utilizes the imputed complete data for classification. Although significant progress has been achieved by exist- ing IMVC methods, there are still limitations: (1) Methods that simply neglect missing views are usually ineffective especially under high missing rate due to the limitation in exploring the correlation among views; (2) Methods that impute missing data based on deep learning methods are short in interpretability and the deterministic imputation way fails to characterize the uncertainty of missing resulting in unstable classification; (3) Few IMVC methods can handle multi-view data with complex missing patterns especially for the data with more than two views, which makes these methods inflexible. In view of above limitations, we propose a simple yet effective, stable and flexible incomplete multi-view classi- fication model. First, the proposed model characterizes the uncertainty of each missing view by imputing a distribution instead of a deterministic value. The necessity of charac- terizing the uncertainty for missing data on single view has been well recognized in recent works [17, 18]. Second, we conduct sampling multiple times from the above distribution and each one is combined with the observed views to form multiple completed multi-view samples. The quality of the sampled data is of high uncertainty, and thus we adaptively integrate them according to their quality on the single view and multi-view fusion. For single view, the uncertainty of the low-quality sampled data tends to be large and should not affect the learning of other views. Therefore, we con- struct an evidence-based classifier for each view to obtain the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19873 opinion including subjective probabilities and uncertainty masses. While for multi-view fusion, the opinions from mul- tiple views are integrated based on DS rule, which ensures the trustworthy utilization of arbitrary-quality views. The main contributions of this work are summarized as follows: •We propose an exploration-and-exploitation strategy for classifying incomplete multi-view data by characteriz- ing the uncertainty of missing data, which promotes both effectiveness and trustworthiness in utilizing im- puted data. To the best of our knowledge, the proposed UIMC is the first work asserting uncertainty in incom- plete multi-view classification. •To fully exploit the high-quality imputed data and re- duce the affect of the low-quality imputed views, we propose to weight the imputed data from two aspects, avoiding the negative effect on single view and multi- view fusion. The uncertainty-aware training and fusion significantly ensure the effectiveness and reliability of integrating uncertain imputed data. •We conduct experiments on classification for data with multiple types of features or modalities, and evaluate the results with diverse metrics, which validates that the proposed UIMC outperforms existing methods in the above tasks and is trustworthy with reliable uncertainty.
Wang_Hunting_Sparsity_Density-Guided_Contrastive_Learning_for_Semi-Supervised_Semantic_Segmentation_CVPR_2023	Abstract Recent semi-supervised semantic segmentation methods combine pseudo labeling and consistency regularization to enhance model generalization from perturbation-invariant training. In this work, we argue that adequate supervi- sion can be extracted directly from the geometry of fea- ture space. Inspired by density-based unsupervised clus- tering, we propose to leverage feature density to locate sparse regions within feature clusters defined by label and pseudo labels. The hypothesis is that lower-density fea- tures tend to be under-trained compared with those densely gathered. Therefore, we propose to apply regularization on the structure of the cluster by tackling the sparsity to in- crease intra-class compactness in feature space. With this goal, we present a Density-Guided Contrastive Learning (DGCL) strategy to push anchor features in sparse regions toward cluster centers approximated by high-density posi- tive keys. The heart of our method is to estimate feature density which is defined as neighbor compactness. We de- sign a multi-scale density estimation module to obtain the density from multiple nearest-neighbor graphs for robust density modeling. Moreover, a unified training framework is proposed to combine label-guided self-training and density- guided geometry regularization to form complementary su- pervision on unlabeled data. Experimental results on PAS- CAL VOC and Cityscapes under various semi-supervised settings demonstrate that our proposed method achieves state-of-the-art performances. The project is available athttps://github.com/Gavinwxy/DGCL .	1. Introduction Semantic segmentation, as an essential computer vision task, has seen significant advances along with the rise of deep learning [4, 30, 46]. Nevertheless, training segmenta- tion models requires massive pixel-level annotations which can be time-consuming and laborious to obtain. Therefore, *Corresponding author. (a) Feature clusters (b) Local density estimationFigure 1. Illustration of feature density within clusters. (a) Pixel- level features of 5 classes extracted from PASCAL VOC 2012 [12] with prediction confidence over 0.95. (b) Feature density esti- mated by the averaged distance of 16 nearest neighbor features. semi-supervised learning is introduced in semantic segmen- tation and is drawing growing interest. It aims to design a label-efficient training scheme with limited annotated data to allow better model generalization by leveraging addi- tional unlabeled images. The key in semi-supervised semantic segmentation lies in mining extra supervision from unlabeled samples. Re- cent studies focus on learning consistency-based regular- ization [13, 29, 32, 51] and designing self-training pipelines [15, 19, 43, 44]. Inspired by the advances in representation learning [17,24], another line of works [28,40,47–49] intro- duce contrastive learning on pixel-level features to enhance inter-class separability. Though previous works have shown effectiveness, their label-guided learning scheme solely re- lies on classifier knowledge, while the structure information of feature space is under-explored. In this work, we argue that effective supervision can be extracted from the geome- try of feature clusters to complement label supervision. Feature density, measured by local compactness, has shown its potential to reveal feature patterns in unsuper- vised clustering algorithms such as DBSCAN [11]. The density-peak assumption [33] states that cluster centers are more likely located in dense regions. Inversely, features in This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3114 sparse areas tend to be less representative within the clus- ter, so they require extra attention. The sparsity exists even within features that are confidently predicted by the classi- fier. As shown in Fig. 1, pixel-level features are extracted on labeled and unlabeled images from 5 classes in PASCAL VOC 2012 dataset, all with high prediction confidence. Ev- ident density variation still exists within each cluster, indi- cating varying learning difficulty among features, which the classifier fails to capture. In this work, we propose a learning strategy named Density-Guided Contrastive Learning (DGCL) to mine ef- fective supervision from cluster structure of unlabeled data. Specifically, we initialize categorical clusters based on la- beled features and enrich them with unlabeled features which are confidently predicted. Then, sparsity hunting is conducted in each in-class feature cluster to locate low- density features as anchors. Meanwhile, features in dense regions are selected to approximate the cluster centers and serve as the positive keys. Then, feature contrast is applied to push the anchors toward their positive keys, explicitly shrinking sparse regions to enforce more compact clusters. The core of our method is feature density estimation. We measure local density by the average distance between the target feature and its nearest neighbors. For robust esti- mation, categorical memory banks are proposed to break the limitation on mini-batch, so in-class density can be es- timated in a feature-to-bank style where class distribution can be approximated globally. When building the nearest neighbor graph, densities estimated by fewer neighbors tend to focus on the local region, which prevents capturing true cluster centers. On the other hand, graphs with too many neighbors cause over-smoothed estimation, which harms accurate sparsity mining. Therefore, we propose multi- scale nearest neighbor graphs to determine the final density by combining estimations from graphs of different sizes. We evaluate the proposed method on PASCAL VOC 2012 [12] and Cityscapes [8] under various semi-supervised settings, where our approach achieves state-of-the-art per- formances. Our contributions are summarized as follows: • We propose a density-guided contrastive learning strat- egy to tackle semi-supervised semantic segmentation by mining effective supervision from the geometry in feature space. • We propose a multi-scale density estimation module combined with dynamic memory banks to capture fea- ture density robustly. • We propose a unified learning framework consisting of label-guided self-training and density-guided fea- ture learning, in which two schemes complement each other. Experiments show that our method achieves state-of-the-art performances.
Xie_On_Data_Scaling_in_Masked_Image_Modeling_CVPR_2023	Abstract Scaling properties have been one of the central issues in self-supervised pre-training, especially the data scalabil- ity, which has successfully motivated the large-scale self- supervised pre-trained language models and endowed them with significant modeling capabilities. However, scaling properties seem to be unintentionally neglected in the re- cent trending studies on masked image modeling (MIM), and some arguments even suggest that MIM cannot ben- efit from large-scale data. In this work, we try to break down these preconceptions and systematically study the scal- ing behaviors of MIM through extensive experiments, with data ranging from 10% of ImageNet-1K to full ImageNet- 22K, model parameters ranging from 49-million to one- billion, and training length ranging from 125K to 500K iterations. And our main findings can be summarized in two folds: 1) masked image modeling remains demand- ing large-scale data in order to scale up computes and model parameters; 2) masked image modeling cannot bene- fit from more data under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-trained language models or supervised pre-trained vi- sion models. In addition, we reveal several intriguing prop- erties in MIM, such as high sample efficiency in large MIM models and strong correlation between pre-training vali- dation loss and transfer performance. We hope that our findings could deepen the understanding of masked im- age modeling and facilitate future developments on large- scale vision models. Code and models will be available at https://github.com/microsoft/SimMIM .	1. Introduction Masked Image Modeling (MIM) [3, 18, 44], which has recently emerged in the field of self-supervised visual pre- training, has attracted widespread interest and extensive ap- plications throughout the community for unleashing the su- perior modeling capacity of attention-based Transformer The work is done when Zhenda Xie, Yutong Lin, and Yixuan Wei are interns at Microsoft Research Asia.†Project co-leaders.architectures [13, 26] and demonstrating excellent sample efficiency and impressive transfer performance on a vari- ety of vision tasks. However, most recent practices are focused on the design of MIM methods, while the study of scaling properties of MIM is unintentionally neglected, especially the data scaling property, which successfully mo- tivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capa- bilities. Although previous works [8, 10, 17, 37, 45] have explored several conclusions about the scaling properties of vision models, most of their findings were obtained under a supervised pre-training scheme or under a contrastive learn- ing framework, so the extent to which these findings could be transferred to MIM still needs to be investigated. Meanwhile, with the emergence of Transformers [41] and masked language modeling (MLM) [12, 31], the systematic studies of scaling laws have already been explored in natural language processing field [21,23,35], which provided ample guidance for large models in recent years. The core finding drawn from scaling laws [21] for neural language models is that the performance has a power-law relationship with each of the three scale factors – model parameters N, size of dataset D, and amount of compute C respectively – when not bottlenecked by the other two . This conclusion implies that better performance can be obtained by scaling up these three factors to the extent that the scaling laws are in effect, which led to the subsequent developments of large scale language models [16,28,29,32,33] that exhibit excellent modeling ca- pabilities [4] on most language tasks. Therefore, it is natural to ask whether MIM possesses the same scaling signatures for vision models as the MLM method for language models, so that the scaling up of vision models can catch up with language models. Though masked image modeling and masked language modeling both belong to masked signal prediction, their prop- erty differences are also non-negligible due to the different nature of vision and language. That is, the images are highly redundant raw signals and words/sentences are semantically rich tokens, which may result in different abilities of data utilization, and it is thus debatable whether the observations from language models could be reproduced in vision models. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10365 Figure 1. The curves of validation loss of pre-training models w.r.t. the relative compute, dataset size and model size. MIM performance improves smoothly as we increase the relative compute and model size, but not improve when the dataset size is sufficient to prevent model from overfitting. We set the relative compute of SwinV2-S for 125K iterations as the value of 1. Best viewed in color. Moreover, recent studies [14, 38] have shown that using a small amount of training data in masked image modeling can achieve comparable performance to using large datasets, which also motivates our explorations as it disagrees with previous findings and intuitions. In this paper, we systematically investigate the scaling properties, especially the data scaling capability of masked image modeling in terms of different dataset sizes ranging from 10% of ImageNet ( ∼0.1 million) to full ImageNet-22K (∼14 million), model sizes ranging from 49-million Swin- V2 Small to 1-billion Swin-V2 giant and training lengths ranging from 125K to 500K iterations. We use Swin Trans- former V2 [25] as the vision encoder for its proven train- ability of large models and applicability to a wide range of vision tasks, and adopt SimMIM [44] for masked image modeling pre-training because it has no restrictions on en- coder architectures. We also conduct experiments with other MIM frameworks like MAE [18] and other vision encoder like the widely used ViT [13] to verify the generalizability of our findings. With these experimental setups, our main findings could be summarized into two folds: i)Masked image modeling remains demanding large- scale data in order to scale up computes and model parame- ters. We empirically find that MIM performance has a power- law relationship with relative compute and model size when not bottlenecked by the dataset size (Figure 1). Besides, we observe that smaller datasets lead to severe overfitting phe- nomenon for training large models (Figure 2), and the size of the dataset to prevent model from overfitting increases clearly as the model increases (Figure 5-Left). Therefore, from the perspective of scaling up model, MIM still demands large-scale data. Furthermore, if we train large-scale models of different lengths, we find that relatively small datasets are adequate at shorter training lengths, but still suffer from overfitting at longer training lengths(Figure 5-Right), which further demonstrate the data scalability of MIM for scaling up compute. 2)Masked image modeling cannot benefit from more dataunder a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-traind language models or supervised pre-trained vision models. In MIM, we find that increasing the number of unique samples for a non-overfitting model does not provide additional bene- fits to performance (Figure 1). This behavior differs from previous observations in supervised vision transformers and self-supervised language models, where the model perfor- mance increases as the number of unique samples increases. In addition, we also demonstrate some intriguing proper- ties about masked image modeling, such as larger models possessing higher sample efficiency, i.e., fewer optimization steps are required for larger models to achieve same perfor- mance (Section 4.2); and the consistency between transfer and test performance, i.e., test performance could be used to indicate the results on downstream tasks (Section 4.3). These observations also correspond to those in previous practices. These findings on the one hand confirm the effect of MIM on scaling up model size and compute, and raise new con- cerns and challenges on the data scalability of MIM on the other. We hope that our findings could deepen the under- standing of masked image modeling and facilitate future developments on large-scale vision models.
Xia_VecFontSDF_Learning_To_Reconstruct_and_Synthesize_High-Quality_Vector_Fonts_via_CVPR_2023	Abstract Font design is of vital importance in the digital con- tent design and modern printing industry. Developing al- gorithms capable of automatically synthesizing vector fonts can significantly facilitate the font design process. How- ever, existing methods mainly concentrate on raster image generation, and only a few approaches can directly syn- thesize vector fonts. This paper proposes an end-to-end trainable method, VecFontSDF , to reconstruct and synthe- size high-quality vector fonts using signed distance func- tions (SDFs). Specifically, based on the proposed SDF- based implicit shape representation, VecFontSDF learns to model each glyph as shape primitives enclosed by sev- eral parabolic curves, which can be precisely converted to quadratic B ´ezier curves that are widely used in vec- tor font products. In this manner, most image generation methods can be easily extended to synthesize vector fonts. Qualitative and quantitative experiments conducted on a publicly-available dataset demonstrate that our method ob- tains high-quality results on several tasks, including vector font reconstruction, interpolation, and few-shot vector font synthesis, markedly outperforming the state of the art.	1. Introduction Traditional vector font designing process relies heavily on the expertise and effort from professional designers, set- ting a high barrier for common users. With the rapid de- velopment of deep generative models in the last few years, a large amount of effective and powerful methods [1, 7, 32] have been proposed to synthesize visually-pleasing glyph images. In the meantime, how to automatically reconstruct and generate high-quality vector fonts is still considered as a challenging task in the communities of Computer Vi- sion and Computer Graphics. Recently, several methods *Denotes equal contribution. †Corresponding author. E-mail: [email protected] This work was supported by National Language Committee of China (Grant No.: ZDI135-130), Center For Chinese Font Design and Research, and Key Laboratory of Science, Technology and Standard in Press Industry (Key Laboratory of Intelligent Press Media Technology). (a) Vector Font Reconstruction (b) Vector Font Interpolation Figure 1. Examples of results obtained by our method in the tasks of vector font reconstruction (a) and vector font interpolation (b). based on sequential generative models [3,11,20,30,33] have been reported that treat a vector glyph as a draw-commands sequence and use Recurrent Neural Networks (RNNs) or Transformer [31] to encode and decode the sequence. How- ever, this explicit representation of vector graphics is ex- tremely difficult for the learning and comprehension of deep neural networks, mainly due to the long-range dependence and the ambiguity in how to draw the outlines of glyphs. More recently, DeepVecFont [33] was proposed to use dual- modality learning to alleviate the problem, showing state- of-the-art performance on this task. Its key idea is to use a CNN encoder and an RNN encoder to extract features from both image and sequence modalities. Despite using richer information of dual-modality data, it still needs repetitively random samplings of synthesis results to find the optimal one and then uses Diffvg [18] to refine the vector glyphs un- der the guidance of generated images in the inference stage. Another possible solution to model the vector graphic is to use implicit functions which have been widely used to represent 3D shapes in recent years. For instance, DeepSDF [25] adopts a neural network to predict the values of signed distance functions for surfaces, but it fails to con- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1848 vert those SDF values to the explicit representation. BSP- Net [4] uses hyperplanes to split the space in 2D shapes and 3D meshes, but it generates unsmooth and disordered re- sults when handling glyphs consisting of numerous curves. Inspired by the above-mentioned existing methods, we propose an end-to-end trainable model, VecFontSDF, to reconstruct and synthesize high-quality vector fonts using signed distance functions (SDFs). The main idea is to treat a glyph as several primitives enclosed by parabolic curves which can be translated to quadratic B ´ezier curves that are widely used in common vector formats like SVG and TTF. Specifically, we use the feature extracted from the input glyph image by convolutional neural networks and decode it to the parameters of parabolic curves. Then, we calculate the value of signed distance function for the nearest curve of every sampling point and leverage these true SDF values as well as the target glyph image to train the model. The work most related to ours is [19], which also aims to provide an implicit shape representation for glyphs, but possesses a serious limitation mentioned below. For conve- nience, we name the representation proposed in [19] IGSR, standing for “Implicit Glyph Shape Representation”. The quadratic curves used in IGSR are not strictly parabolic curves, which cannot be translated to quadratic B ´ezier curves. As a consequence, their model is capable of synthe- sizing high-resolution glyph images but not vector glyphs. Furthermore, it only uses raster images for supervision which inevitably leads to inaccurate reconstruction results. On the contrary, our proposed VecFontSDF learns to recon- struct and synthesize high-quality vector fonts that consist of quadratic B ´ezier curves by training on the corresponding vector data in an end-to-end manner. Major contributions of our paper are threefold: • We design a new implicit shape representation to pre- cisely reconstruct high-quality vector glyphs, which can be directly converted into commonly-used vector font formats (e.g., SVG and TTF). • We use the true SDF values as a strong supervision instead of only raster images to produce much more precise reconstruction and synthesis results compared to previous SDF-based methods. • The proposed VecFontSDF can be flexibly integrated with other generative methods such as latent space in- terpolation and style transfer. Extensive experiments have been conducted on these tasks to verify the supe- riority of our method over other existing approaches, indicating its effectiveness and broad applications.
Wu_Semi-Supervised_Stereo-Based_3D_Object_Detection_via_Cross-View_Consensus_CVPR_2023	Abstract Stereo-based 3D object detection, which aims at detect- ing 3D objects with stereo cameras, shows great potential in low-cost deployment compared to LiDAR-based methods and excellent performance compared to monocular-based algorithms. However, the impressive performance of stereo- based 3D object detection is at the huge cost of high-quality manual annotations, which are hardly attainable for any given scene. Semi-supervised learning, in which limited an- notated data and numerous unannotated data are required to achieve a satisfactory model, is a promising method to address the problem of data deficiency. In this work, we pro- pose to achieve semi-supervised learning for stereo-based 3D object detection through pseudo annotation generation from a temporal-aggregated teacher model, which tempo- rally accumulates knowledge from a student model. To fa- cilitate a more stable and accurate depth estimation, we introduce Temporal-Aggregation-Guided (TAG) disparity consistency, a cross-view disparity consistency constraint between the teacher model and the student model for robust and improved depth estimation. To mitigate noise in pseudo annotation generation, we propose a cross-view agreement strategy, in which pseudo annotations should attain high degree of agreements between 3D and 2D views, as well as between binocular views. We perform extensive exper- iments on the KITTI 3D dataset to demonstrate our pro- posed method’s capability in leveraging a huge amount of unannotated stereo images to attain significantly improved detection results.	1. Introduction 3D object detection, as one of the most significant per- ception tasks in the computer vision community, has wit- nessed great progress in recent years, especially after the advent of neural-network-based deep learning. Most state- of-the-art works which focus on 3D object detection mainly *Corresponding author. Figure 1. Illustration of the proposed Temporal-Aggregation- Guided (TAG) disparity consistency constraint, in which the teacher model, temporally collecting knowledge from the student model, leads the disparity estimation of the student model across the views. rely on LiDAR data [17,32,33,43,51] to extract accurate 3D information, such as 3D structure and depth of the points. However, LiDAR data are costly to harvest and annotate, and have limited sensing ranges in some cases. Instead, vision-based methods, which detect 3D objects based on images only, have drawn more attention in recent years. While it is convenient to collect image data, there are signif- icant challenges in applying image-based methods to depth sensing, which is an ill-posed problem when localizing ob- jects with only images. This problem is further exacerbated in monocular-based 3D object detection [24,25,48]. Stereo images, in which image pairs took at different viewpoints are available, can be used to reconstruct depth information through pixel-to-pixel correspondence or stereo geometry, making them more useful for detecting 3D objects without the introduction of expensive LiDAR data. With more attention focusing on stereo-based 3D object detection, the performance of these methods gets continu- ously improved, and the gap between LiDAR-based meth- ods and stereo-based methods becomes progressively nar- rower. However, the improving performance is built on large-scale manual annotation, which is costly in terms of both time and human resources. When the amount of an- notation is limited, the performance of stereo-based meth- ods deteriorates rapidly. Semi-supervised learning, which This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17471 makes use of a limited set of annotated data and plenty of unannotated data, is a promising method for solving the data deficiency problem. In this work, we propose an effective and efficient semi-supervised stereo-based 3D object detec- tion method to alleviate this limited annotation problem. With limited annotated stereo images, depth estima- tion, a key stage for accurately localizing 3D objects, be- comes increasingly unstable, which in turn leads to poor object localization. However, the inherent constraint be- tween the left and right image in each stereo pair can be adopted to promote the model’s performance in depth es- timation. Inspired by the left-right disparity consistency constraint proposed by Godard et al . [11], we propose a Temporal-Aggregation-Guided (TAG) disparity consis- tency constraint, as shown in Fig. 1. In this method, dis- parity estimation of the base model, as the student model, should pursue that of the teacher model, with continuous knowledge accumulation from the base model, across the views. With the proposed TAG method, the base model can enhance its capability in depth estimation when provided with extra data without annotations. In addition, to gener- ate more accurate pseudo annotations for the unannotated data, we propose a cross-view agreement strategy, which comprises a 3D-2D agreement constraint, in which pseudo annotations with high localization consistency in both 2D view and 3D view are kept, and a left-right agreement con- straint, in which pseudo annotations with high similarities in both the left and right view at the feature space are re- tained for pseudo-supervision. With the proposed cross- view agreement strategy, the remaining pseudo annotations can effectively encompass the objects and provide high- quality supervision on the unannotated data, further improv- ing the detection performance of the base model. We evaluate the proposed method on the KITTI 3D dataset [10]. With our proposed method, the base model achieves relative performance gains of up to 18 percentage points under the evaluation metric of AP 3Dand 20 percent- age points under the evaluation metric of APBEV with the IoU threshold of 0.7 in the car category when only 5% of training data are annotated, thus verifying the effectiveness of our proposed method in making full use of data without annotations. We summarize our main contributions as follows: 1) We propose a semi-supervised method for stereo- based 3D object detection, in which plentiful and easy- to-access unannotated images are fully utilized to en- hance the base model. 2) We propose a Temporal-Aggregation-Guided (TAG) disparity consistency constraint to direct the dispar- ity estimation of the base model through the teacher model, with cumulative knowledge from the base model, across the views.3) We introduce a cross-view agreement strategy to re- fine the generated pseudo annotations through enforc- ing agreements between the 3D view and 2D view, and between the left view and right view. 4) Our proposed semi-supervised method leads to signifi- cant performance gains without requiring extensive an- notations on the KITTI 3D dataset.
Wei_Sparsifiner_Learning_Sparse_Instance-Dependent_Attention_for_Efficient_Vision_Transformers_CVPR_2023	Abstract Vision Transformers (ViT) have shown competitive advan- tages in terms of performance compared to convolutional neural networks (CNNs), though they often come with high computational costs. To this end, previous methods explore different attention patterns by limiting a fixed number of spatially nearby tokens to accelerate the ViT’s multi-head self-attention (MHSA) operations. However, such struc- tured attention patterns limit the token-to-token connections to their spatial relevance, which disregards learned seman- tic connections from a full attention mask. In this work, we propose an approach to learn instance-dependent at- tention patterns, by devising a lightweight connectivity pre- dictor module that estimates the connectivity score of each pair of tokens. Intuitively, two tokens have high connectiv- ity scores if the features are considered relevant either spa- tially or semantically. As each token only attends to a small number of other tokens, the binarized connectivity masks are often very sparse by nature and therefore provide the opportunity to reduce network FLOPs via sparse compu- tations. Equipped with the learned unstructured attention pattern, sparse attention ViT (Sparsifiner) produces a supe- rior Pareto frontier between FLOPs and top-1 accuracy on ImageNet compared to token sparsity. Our method reduces 48%∼69% FLOPs of MHSA while the accuracy drop is within 0.4%. We also show that combining attention and token sparsity reduces ViT FLOPs by over 60%.	1. Introduction Vision Transformers (ViTs) [15] have emerged as a dom- inant model for fundamental vision tasks, such as image classification [15], object detection [3], and semantic seg- mentation [6, 7]. However, scaling ViTs to a large number of tokens is challenging due to the quadratic computational complexity of multi-head self-attention (MHSA) [37]. This is particularly disadvantageous for large-scale vi- sion tasks because processing high-resolution and high- *Equal contribution. (c) Axial Attention Pattern (Fixed) (d) Proposed Instance -Dependent Attention Pattern (Dynamic)(b) Window Attention Pattern (Fixed) Query Patch Attended Patch (a) Local Attention Pattern (Fixed) Input ImageFigure 1. Comparison of Sparsifiner and fixed attention patterns. Twins [10] (a), Swin [24] (b), and Axial [18] (c) address quadratic MHSA complexity using fixed attention patterns, which does not consider the instance-dependent nature of semantic information in images. To address this, we propose Sparsifiner (d): an efficient module for sparse instance-dependent attention pattern prediction. dimensionality inputs is desirable. For example, input modalities such as video frames and 3D point clouds have a large number of tokens even for basic use cases. New algo- rithms are needed to continue to scale ViTs to larger, more complex vision tasks. Prior works have largely taken two approaches to im- prove the computational efficiency of ViTs: token pruning and using fixed sparse attention patterns in MHSA. To- ken pruning methods [29] reduce the number of tokens by a fixed ratio called the keep rate, but accuracy degrades quickly when pruning early layers in the network [17, 32, 33]. For example, introducing token pruning into shallower layers of EViT [17] causes a significant 3.16% top-1 accu- racy drop on ImageNet [14]. This issue is due to the restric- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22680 tion of pruning an entire token, which amounts to pruning an entire row and column of the attention matrix at once. One way to alleviate this is to prune connectivity between individual tokens in the attention matrix. Existing meth- ods that take this attention matrix connectivity-pruning ap- proach use fixed sparse attention patterns [8]. For example, local and strided fixed attention patterns are used [8, 16], in combination with randomly-initialized connectivities [42]. However, such fixed attention patterns limit the capacity of the self-attention connections to a fixed subset of to- kens (Fig. 1). The fixed nature of these attention patterns is less effective compared to the direct communication be- tween tokens in full self-attention. For example, Swin trans- former [23,24] has a limited the receptive field at shallower layers and needs many layers to model long-range depen- dencies. And BigBird [42] needs to combine multiple fixed attention patterns to achieve good performance. Rather, it is desirable to design sparse attention algorithms that mimic full self attention’s instance-dependent nature [37], thereby capturing the variable distribution of semantic information in the input image content. To address these challenges, we propose a method called Sparsifiner that learns to compute sparse connectivity pat- terns over attention that are both instance-dependent and unstructured . The instance-dependent nature of the atten- tion pattern allows each token to use its limited attention budget of nonzero elements more efficiently compared to fixed sparse attention patterns. For example, in attention heads that attend to semantic rather than positional con- tent [37, 38], tokens containing similar semantic informa- tion should be considered to have high connectivity scores despite their spatial distance. Similarly, nearby tokens with irrelevant semantic relations should have lower connectivity scores despite their spatial proximity. Furthermore, Spar- sifiner improves attention pattern flexibility compared to to- ken pruning by pruning individual connectivities instead of entire rows and columns of the attention matrix. This al- lows Sparsifiner to reduce FLOPs in the early layers of the network without incurring significant top-1 accuracy degra- dation (§4). By pruning individual connectivities dependent on image content, Sparsifiner generalizes prior approaches to sparsifying MHSA in ViTs, and in doing so, produces a favourable trade-off between accuracy and FLOPs. Our contributions are the following: • We propose a novel efficient algorithm called Spar- sifiner to predict instance-dependent sparse attention patterns using low-rank connectivity patterns. Our in- vestigation into instance-dependent unstructured spar- sity is to the best of our knowledge novel in the context of ViTs. • We show that such learned unstructured attention spar- sity produces a superior Pareto-optimal tradeoff be-tween FLOPs and top-1 accuracy on ImageNet com- pared to token sparsity. Furthermore, we show that Sparsifiner is complementary to token sparsity meth- ods, and the two approaches can be combined to achieve superior performance-accuracy tradeoffs. • We propose a knowledge distillation-based approach for training Sparsifiner from pretrained ViTs using a small number of training epochs.
Wang_Object-Aware_Distillation_Pyramid_for_Open-Vocabulary_Object_Detection_CVPR_2023	Abstract Open-vocabulary object detection aims to provide ob- ject detectors trained on a fixed set of object categories with the generalizability to detect objects described by arbi- trary text queries. Previous methods adopt knowledge dis- tillation to extract knowledge from Pretrained Vision-and- Language Models (PVLMs) and transfer it to detectors. However, due to the non-adaptive proposal cropping and single-level feature mimicking processes, they suffer from information destruction during knowledge extraction and inefficient knowledge transfer. To remedy these limitations, we propose an Object-Aware Distillation Pyramid (OADP) framework, including an Object-Aware Knowledge Extrac- tion (OAKE) module and a Distillation Pyramid (DP) mech- anism. When extracting object knowledge from PVLMs, the former adaptively transforms object proposals and adopts object-aware mask attention to obtain precise and com- plete knowledge of objects. The latter introduces global and block distillation for more comprehensive knowledge transfer to compensate for the missing relation information in object distillation. Extensive experiments show that our method achieves significant improvement compared to cur- rent methods. Especially on the MS-COCO dataset, our OADP framework reaches 35.6mAPN 50, surpassing the cur- rent state-of-the-art method by 3.3mAPN 50. Code is released athttps://github.com/LutingWang/OADP .	1. Introduction Open-vocabulary object detection (OVD) [49] aims to endow object detectors with the generalizability to detect open categories including both base andnovel categories where only the former are annotated in the training phase. Pretrained Vision-and-Language Models (PVLMs, e.g., CLIP [32] and ALIGN [19]) have witnessed great progress in recent years, and Knowledge Distillation (KD) [17] has led to a wave of unprecedented advances transferring the zero-shot visual recognition ability from PVLMs to detec- *Corresponding author ([email protected]) Center Crop w/o Transform Center Crop w/ Transform (a) Knowledge Extraction StudentGlobal KDBlock KDObject KD (b) Knowledge Transfer Figure 1. An overview of our OADP framework. (a) Directly applying center crop on proposals may throw informative object parts away, resulting in ambiguous image regions. In contrast, our OAKE module extracts complete objects and reduces the influence of surrounding distractors. (b) Our DP mechanism includes global, block, and object KD to achieve effective knowledge transfer. tors [12, 25, 28, 29, 48, 53]. KD typically comprises two es- sential steps, i.e.,knowledge extraction and then knowledge transfer . A common practice in OVD is to crop objects with class-agnostic proposals and use the teacher ( e.g., CLIP vi- sual encoder) to extract knowledge of the proposals. The knowledge is then transferred to the detector ( e.g., Mask R- CNN [15]) via feature mimicking. Despite significant development, we argue that con- ventional approaches still have two main limitations: 1) Dilemma between comprehensiveness and purity during knowledge extraction . As proposals have diverse aspect ra- tios, the fixed center crop strategy to square them may cut out object parts (fig. 1a). Enlarging those proposals via re- sizing function may alleviate this problem, but additional surrounding distractors may confuse the teacher to extract accurate proposal knowledge. 2) Missing global scene understanding during knowledge transfer . Conventional approaches merely concentrate on object-level knowledge transfer by directly mimicking the teacher’s features of indi- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11186 vidual proposals. As a result, the student cannot fully grasp the contextual characteristics describing the interweaving of different objects. In light of the above discussions, we pro- pose an Object-Aware Distillation Pyramid (OADP) frame- work to excavate the teacher’s knowledge accurately and effectively transfer the knowledge to the student. To preserve the complete information of proposals while extracting their CLIP image embeddings, we propose an Object-Aware Knowledge Extraction (OAKE) module. Concretely, given a proposal, we square it with an adaptive resizing function to avoid destroying the object structure and involve object information as much as possible. How- ever, the resizing process inevitably introduces environmen- tal context, which may contain some distractors that confuse the teacher. Therefore, we propose to utilize an object to- ken[OBJ] whose interaction manner during the forward process is almost the same as the class token [CLS] except that it only attends to patch tokens covered by the origi- nal proposal. In this way, the extracted embeddings contain precise and complete knowledge of the proposal object. To facilitate complete and effective knowledge trans- fer, we propose a Distillation Pyramid (DP) mechanism (fig. 1b). As previous works only adopt object distillation to align the feature space of detectors and PVLMs, the re- lation between different objects is neglected. Therefore, we propose global and block distillation to compensate for the missing relation information in object distillation. For global distillation, we optimize the L1distance between the detector backbone and the CLIP visual encoder so that the detector learns to encode rich semantics implied in the im- age scene. However, the CLIP visual encoder is prone to ignore background information, which may also be valu- able for detection. Therefore, we take a finer step to divide the input image into several blocks and optimize the L1dis- tance between the block embeddings of the detector and the CLIP image encoder. Overall, the above three distillation modules constitute a hierarchical distillation pyramid, al- lowing for the transfer of more diversified knowledge from CLIP to the detectors. We demonstrate the superiority of our OADP framework on MS-COCO [27] and LVIS [14] datasets. On MS-COCO, it improves the state-of-the-art results of mAPN 50from 32.3 to35.6. On the LVIS dataset, our OADP framework reaches 21.9APron the object detection task and 21.7APron the instance segmentation task, leading the former methods by more than 1.1APrand1.9APrrespectively.
Xiang_SQUID_Deep_Feature_In-Painting_for_Unsupervised_Anomaly_Detection_CVPR_2023	Abstract Radiography imaging protocols focus on particular body regions, therefore producing images of great similarity and yielding recurrent anatomical structures across pa- tients. To exploit this structured information, we propose the use of Space-aware Memory Queues for In-painting and Detecting anomalies from radiography images (abbreviated asSQUID ). We show that SQUID can taxonomize the in- grained anatomical structures into recurrent patterns; and in the inference, it can identify anomalies (unseen/modiﬁed patterns) in the image. SQUID surpasses 13 state-of-the- art methods in unsupervised anomaly detection by at least 5 points on two chest X-ray benchmark datasets measured by the Area Under the Curve (AUC). Additionally, we have cre- ated a new dataset ( DigitAnatomy ), which synthesizes the spatial correlation and consistent shape in chest anatomy. We hope DigitAnatomy can prompt the development, eval- uation, and interpretability of anomaly detection methods.	1. Introduction Vision tasks in photographic imaging and radiogra- phy imaging are different. For example, when identify- ing objects in photographic images, we assume translation invariance—a cat is a cat no matter if it appears on the left or right of the image. In radiography imaging, on the other hand, the relative location and orientation of a structure are important characteristics that allow the identiﬁcation of nor- mal anatomy and pathological conditions [ 20,83]. Since ra- diography imaging protocols assess patients in a fairly con- sistent orientation, the generated images have great similar- ity across various patients, equipment manufacturers, and facility locations (see examples in Figure 1d). The consis- tent and recurrent anatomy facilitates the analysis of numer- ous critical problems and should be considered a signiﬁcant advantage for radiography imaging [ 85]. Several investiga- *Corresponding author: Zongwei Zhou ( [email protected] ) Figure 1. Anomaly detection in radiography images can be both easier and harder than in photographic images. It is easier be- cause radiography images are spatially structured due to consistent imaging protocols. It is harder because anomalies in radiography images are subtle and require medical expertise to annotate. tions have demonstrated the value of this prior knowledge in enhancing Deep Nets’ performance by adding location fea- tures, modifying objective functions, and constraining co- ordinates relative to landmarks in images [ 3,47,49,69,86]. Our work seeks to answer this critical question: Can we exploit consistent anatomical patterns and their spatial in- formation to strengthen Deep Nets’ detection of anomalies from radiography images without manual annotation? Unsupervised anomaly detection only uses healthy im- ages for model training and requires no other annotations such as disease diagnosis or localization [ 5]. As many as 80% of clinical errors occur when the radiologist misses the abnormality in the ﬁrst place [ 7]. The impact of anomaly detection is to reduce that 80% by clearly pointing out to ra- diologists that there exists a suspicious lesion and then hav- ing them look at the scan in depth. Unlike previous anomaly detection methods, we formulate the task as an in-painting task to exploit the anatomical consistency in appearance, position, and layout across radiography images. Specif- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23890 ically, we propose Space-aware Memory Queues for In- painting and Detecting anomalies from radiography images (abbreviated as SQUID ). During training, our model can dynamically maintain a visual pattern dictionary by taxono- mizing recurrent anatomical patterns based on their spatial locations. Due to the consistency in anatomy, the same body region across healthy images is expected to express similar visual patterns, which makes the total number of unique pat- terns manageable. During inference, since anomaly patterns do not exist in the dictionary, the generated radiography im- age is expected to be unrealistic if an anomaly is present. As a result, the model can identify the anomaly by discrim- inating the quality of the in-painting task. The success of anomaly detection has two basic assumptions [ 89]:ﬁrst, anomalies only occur rarely in the data; second , anomalies differ from the normal patterns signiﬁcantly. We have conducted experiments on two large-scale, pub- licly available radiography imaging datasets. Our SQUID is signiﬁcantly superior to predominant methods in unsuper- vised anomaly detection by over 5 points on the ZhangLab dataset [ 32]; remarkably, we have demonstrated a 10-point improvement over 13 recent unsupervised anomaly detec- tion methods on the Stanford CheXpert dataset [ 29]. In addition, we have created a new dataset ( DigitAnatomy ) to elucidate spatial correlation andconsistent shape of the chest anatomy in radiography (see Figure 1c).Digi- tAnatomy is dedicated to easing the development, eval- uation, and interpretability of anomaly detection methods. The qualitative visualization clearly shows the superiority of our SQUID over the current state-of-the-art methods. In summary, our contributions include: (I)the best per- forming unsupervised anomaly detection method for chest radiography imaging; (II)a synthetic dataset to promote anomaly detection research; (III) SQUID overcomes lim- itations in dominant unsupervised anomaly detection meth- ods [ 1,17,35,61,82] by inventing Space-aware Memory Queue (§ 3.2), and Feature-level In-painting (§ 3.3).
Wang_Glocal_Energy-Based_Learning_for_Few-Shot_Open-Set_Recognition_CVPR_2023	Abstract Few-shot open-set recognition (FSOR) is a challenging task of great practical value. It aims to categorize a sam- ple to one of the pre-defined, closed-set classes illustrated by few examples while being able to reject the sample from unknown classes. In this work, we approach the FSOR task by proposing a novel energy-based hybrid model. The model is composed of two branches, where a classification branch learns a metric to classify a sample to one of closed- set classes and the energy branch explicitly estimates the open-set probability. To achieve holistic detection of open- set samples, our model leverages both class-wise and pixel- wise features to learn a glocal energy-based score, in which a global energy score is learned using the class-wise fea- tures, while a local energy score is learned using the pixel- wise features. The model is enforced to assign large energy scores to samples that are deviated from the few-shot exam- ples in either the class-wise features or the pixel-wise fea- tures, and to assign small energy scores otherwise. Exper- iments on three standard FSOR datasets show the superior performance of our model.1	1. Introduction In recent years, deep learning has flourished in various fields with the ever-increasing scale of the training data un- der the closed-world learning settings, i.e., the training and test sets share exactly the same set of classes. However, such settings often do not hold in many real applications. This is because 1) it is difficult or costly to obtain a large amount of labeled data, and 2) models deployed in open- world environments need to constantly deal with samples from unknown classes. For example, in the application of deep learning for diagnosing rare diseases, the number of samples is limited. In this case, the model is prone to over- fitting, resulting in a significant degradation in performance. Further, there can be unknown variants of those diseases due *H. Wang, G. Pang and P. Wang contributed equally in this work. †Corresponding author. 1Code is available at https://github.com/00why00/Glocal support: linnet query: robin support: cliff query: Arctic wolf GT - Unknown PEELER[18] - linnet SnaTCHer[15] - linnet ATT-G[13] - Unknown Ours - Unknown GT - Unknown PEELER[18] - cliff SnaTCHer[15] - cliff ATT-G[13] - cliff Ours - Unknown Figure 1. Two typical errors with existing methods. Our method can detect unknown open-set samples that are similar to the closed-set samples in either the global class level or the local fea- ture level. to our limited understanding of the diseases. Thus, the mod- els are required to perform the classification accurately for the classes illustrated by limited samples, while at the same time detecting the samples from unknown classes. The lat- ter ability is important, especially for healthcare or safety- critical applications, e.g., to alert the unknown cases for hu- man investigation in the disease diagnosis example, or to request human intervention for handling unknown objects in autonomous driving. Few shot learning [16, 23, 28, 32, 35, 36] (FSL) and open set recognition [3,7,9,22,29,30] (OSR) are two techniques This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7507 dedicated to solve these two problems, respectively. FSL methods are trained to achieve a good generalization abil- ity on the new task with only a few training samples. But FSL approaches are developed under a closed-set setting. It lacks the ability to distinguish the classes unseen during training. The goal of OSR is, on the other hand, to recognize open-set samples while maintaining the classification abil- ity of closed-set samples. However, its classification ability is often built upon the availability of a large number of train- ing samples. Thus, OSR approaches fail to work effectively when only a few training samples are available. Few-shot open-set recognition (FSOR), which combines the FSL and OSR problems, is a largely under-explored area. FSOR requires the model to utilize only a few train- ing samples to effectively achieve the ability of both closed- set classification and open-set recognition. Existing FSOR methods [6,13,15] are based on the prototype network [32], which performs classification by measuring the distance be- tween the prototype of each class and the query embedding feature of a sample. They improve the original closed-set classifier in recognizing open-set samples by learning an additional open-set class using pseudo open-set samples. However, only the class-wise information of the sample is considered, and the pixel-wise spatial information of the sample is ignored. As shown in Figure 1, these methods fail to distinguish open-set samples from closed-set class sam- ples that share similar global semantic appearances. Fur- ther, the optimization objectives of FSL and OSR are differ- ent from each other. Thus, training using only an open-set classifier can limit the performance of these models. To solve the above problems, we propose a novel FSOR method, called Glocal Energy-based Learning (GEL). Dif- ferent from previous methods, GEL consists of two classi- fication components: one for closed-set classification and one for open-set recognition. Specifically, in addition to use the class-wise features to classify closed-set samples in the closed-set classifier, GEL leverages both class-wise and pixel-wise features to learn a new energy-based open- set classifier, in which a global energy score is learned using the class-wise features while a local energy score is learned using the pixel-wise features. GEL is enforced to assign large energy scores to samples that are deviated from the few-shot examples in either the class-wise features or the pixel-wise features, and to assign small energy scores oth- erwise. In doing so, GEL can detect unknown class samples that are deviated from the known classes in either high-level abstractions or fine-grained appearances, as shown in Fig- ure 1. In summary, this work makes the following three main contributions: • We propose a novel FSOR framework that learns glo- cal open scores for detecting unknown samples from the class-wise (global) and pixel-wise (local) scales.• We further propose a novel energy-based FSOR model, dubbed GEL, that learns glocal energy-based open scores based on the class-wise and pixel-wise similarities of query samples to the support set. • Through extensive experiments on three widely-used datasets, we show that GEL outperforms state-of-the- art competing methods and achieves state-of-the-art re- sults on these benchmarks.
Wang_Imagen_Editor_and_EditBench_Advancing_and_Evaluating_Text-Guided_Image_Inpainting_CVPR_2023	Abstract Text-guided image editing can have a transformative im- pact in supporting creative applications. A key challenge is to generate edits that are faithful to input text prompts, while consistent with input images. We present Imagen Ed- itor, a cascaded diffusion model built, by fine-tuning Im- agen [36] on text-guided image inpainting. Imagen Ed- itor’s edits are faithful to the text prompts, which is ac- complished by using object detectors to propose inpaint- ing masks during training. In addition, Imagen Editor cap- tures fine details in the input image by conditioning the cas- caded pipeline on the original high resolution image. To im- prove qualitative and quantitative evaluation, we introduce EditBench , a systematic benchmark for text-guided image inpainting. EditBench evaluates inpainting edits on natu- ral and generated images exploring objects, attributes, and scenes. Through extensive human evaluation on EditBench, we find that object-masking during training leads to across- the-board improvements in text-image alignment – such that Imagen Editor is preferred over DALL-E 2 [31] and Stable Diffusion [33] – and, as a cohort, these models are better at object-rendering than text-rendering, and handle mate- rial/color/size attributes better than count/shape attributes. ∗Equal contribution.†Equal advisory contribution.	1. Introduction Text-to-image generation has seen a surge of recent inter- est [31, 33, 36, 50, 51]. While these generative models are surprisingly effective, users with specific artistic and de- sign needs do not typically obtain the desired outcome in a single interaction with the model. Text-guided image edit- ingcan enhance the image generation experience by sup- porting interactive refinement [13, 17, 34, 46]. We focus on text-guided image inpainting, where a user provides an im- age, a masked area, and a text prompt and the model fills the masked area, consistent with both the prompt and the image context (Fig. 1). This complements mask-free edit- ing [13, 17, 46] with the precision of localized edits [5, 27]. This paper contributes to the modeling and evaluation of text-guided image inpainting. Our modeling contribution is Imagen Editor ,2a text-guided image editor that combines large scale language representations with fine-grained con- trol to produce high fidelity outputs. Imagen Editor is a cascaded diffusion model that extends Imagen [36] through finetuning for text-guided image inpainting. Imagen Edi- tor adds image and mask context to each diffusion stage via three convolutional downsampling image encoders (Fig. 2). A key challenge in text-guided image inpainting is ensur- ing that generated outputs are faithful to the text prompts. The standard training procedure uses randomly masked re- 2https://imagen.research.google/editor/ This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18359 Figure 2. Imagenator is an image editing model built by fine- tuning Imagen. All of the diffusion models, i.e., the base model and super-resolution (SR) models, condition on high-resolution 1024×1024 image and mask inputs. To this end, new convolu- tional image encoders are introduced. gions of input images [27, 35]. We hypothesize that this leads to weak image-text alignment since randomly cho- sen regions can often be plausibly inpainted using only the image context, without much attention to the prompt. We instead propose a novel object masking technique that en- courages the model to rely more on the text prompt during training (Fig. 3). This helps make Imagen Editor more con- trollable and substantially improves text-image alignment. Observing that there are no carefully-designed standard datasets for evaluating text-guided image inpainting, we propose EditBench , a curated evaluation dataset that cap- tures a wide variety of language, types of images, and lev- els of difficulty. Each EditBench example consists of (i) a masked input image, (ii) an input text prompt, and (iii) a high-quality output image that can be used as reference for automatic metrics. To provide insight into the relative strengths and weaknesses of different models, edit prompts are categorized along three axes: attributes (material, color, shape, size, count), objects (common, rare, text rendering), and scenes (indoor, outdoor, realistic, paintings). Finally, we perform extensive human evaluations on Ed- itBench, probing Imagen Editor alongside Stable Diffu- sion (SD) [33], and DALL-E 2 (DL2) [31]. Human an- notators are asked to judge a) text-image alignment – how well the prompt is realized (both overall and assessing the presence of each object/attribute individually) and b) image quality – visual quality regardless of the text prompt. In terms of text-image alignment, Imagen Editor trained with object-masking is preferred in 68% of comparisons with its counterpart configuration trained with random masking (a commonly adopted method [27, 35, 41]). Improvements are across-the-board in all object and attribute categories. Imagen Editor is also preferred by human annotators rel- ative to SD and DL2 (78% and 77% respectively). As a cohort, models are better at object-rendering than text- rendering, and handle material/color/size attributes better than count/shape attributes. Comparing automatic evalua- tion metrics with human judgments, we conclude that while Figure 3. Random masks (left) frequently capture background or intersect object boundaries, defining regions that can be plausibly inpainted just from image context alone. Object masks (right) are harder to inpaint from image context alone, encouraging models to rely more on text inputs during training. (Note: This example image was generated by Imagen and is not in the training data.) human evaluation remains indispensable, CLIPScore [14] is the most useful metric for hyperparameter tuning and model selection. In summary, our main contributions are: (i) Imagen Ed- itor, a new state-of-the-art diffusion model for high fidelity text-guided image editing (Sec. 3); (ii) EditBench, a man- ually curated evaluation benchmark for text-guided image inpainting that assesses fine-grained details such as object- attribute combinations (Sec. 4); and (iii) a comprehensive human evaluation on EditBench, highlighting the relative strengths and weaknesses of current models, and the use- fulness of various automated evaluation metrics for text- guided image editing (Sec. 5).
Wang_Dynamic_Graph_Learning_With_Content-Guided_Spatial-Frequency_Relation_Reasoning_for_Deepfake_CVPR_2023	Abstract With the springing up of face synthesis techniques, it is prominent in need to develop powerful face forgery de-tection methods due to security concerns. Some existing methods attempt to employ auxiliary frequency-aware in- formation combined with CNN backbones to discover theforged clues. Due to the inadequate information inter- action with image content, the extracted frequency fea- tures are thus spatially irrelavant, struggling to general- ize well on increasingly realistic counterfeit types. To ad- dress this issue, we propose a Spatial-Frequency Dynamic Graph method to exploit the relation-aware features in spa- tial and frequency domains via dynamic graph learning. Tothis end, we introduce three well-designed components: 1)Content-guided Adaptive Frequency Extraction module to mine the content-adaptive forged frequency clues. 2) Multi- ple Domains Attention Map Learning module to enrich thespatial-frequency contextual features with multiscale atten-tion maps. 3) Dynamic Graph Spatial-Frequency FeatureFusion Network to explore the high-order relation of spa-tial and frequency features. Extensive experiments on sev- eral benchmark show that our proposed method sustainedly exceeds the state-of-the-arts by a considerable margin.	1. Introduction Recent years have witnessed the continuous advances in deepfake creation [ 11,27,36]. Utilizing booming open- source tools such as Deepfakes [ 41], novices can readily manipulate the expression and identity of faces to generate visually untraceable videos. Face forgery technology hasstimulated many applications [ 12,14,44,46] with wide ac- ceptance. These techniques can whereas be abused by ma- *Chen Chen is the corresponding author. This work is supported by the National Key R&D Program of China under Grant 2021YFF0602101,Alibaba Group through Alibaba Research Intern Program. SFDG-Detector Spat. Freq.Spat. Freq. 漏b漐Correlation Heatmap 漏c漐Graph Reasoning漏a漐SFDG Framework 漏 b 漐 Cl t i H t Figure 1. The motivation of our proposed approach. Our SFDG method (a) delves into the high-order relationships (b) of spatial and frequency domain via cross-domain graph reasoning (c). licious intentions to make pornographic movies, fake news and political rumors. In the circumstances, it is desperatelyin need to develop powerful forgery detection methods. Early face forgery detection methods [ 7,30,52] treat this challenge as vanilla dichotomy tasks in the prevailing view.They use off-the-shelf backbones to extract the global fea-ture of faces and a binary classiﬁer follow-up to identify the real and counterfeit faces. However, as the counter-feits become increasingly realistic, it is intractable for thesemethods to spot subtle and local forgery traces. One recent study [ 50] reformulates deepfake detection as a ﬁne-grained classiﬁcation task and designs a multi-attentional frame-work to extract local discriminative features from multiple attention maps. It is susceptible to common disturbances and the generalized features remain therefore poorly un- derstood. Some other works resort to speciﬁc forgery pat-terns to encourage better classiﬁcation such as DCT [ 24,28], SRM [ 25] and steganalysis features [ 45]. Although promis- ing advances have been achieved by these previous works, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7278 they always extract frequency features with hand-crafted ﬁlter banks which are content-irrelevant, thus incapable toadapt the changes of complex scenarios. Moreover, theyfuse multi-domain information via adding directly or atten-tional projection. However, these approaches devote little efforts to discover the high-order relation of spatial and fre- quency features and integrating them in a reasonable way. In this paper, we provide seminal insights to exploit adaptive frequency features and delve into the interactions of spatial-frequency domains. To this end, we propose anadaptive extraction-multiscale enhancement-graph fusion paradigm for deepfake detection via dynamic graph learn-ing, which prompts to excavate content-aware frequency clues and the high-order relation of multiple domains. Firstly, for adaptive frequency extraction, we tailor a Content-guided Adaptive Frequency Extraction (CAF ´E) module with coarse-grained DCT and ﬁne-grained DCT tocapture the local frequency cues guided by content-awaremasks. Different from PEL [ 9] and F 3Net [ 28], our cus- tomized frequency learning protocol provides potential formore combinations of frequency features, which is indis- pensable to spot complicated counterfeit patterns. To further enhance the representation of content-guided frequency features, we introduce a Multiple Domains At- tention Map Learning (MDAML) module to generate mul- tiscale spatial and frequency attention maps with high-level semantic features. Speciﬁcally, we ﬁrst propose a Multi-Scale Attention Ensemble (MSAE) module, which pro- duces multi-scale semantic attention maps with large re- ceptive ﬁelds and endows rich contextual information tospatial and frequency domains. Moreover, an AttentionMap Reﬁnement Block (AMRB) is included in the MSAEmodule to reﬁne the obtained semantic attention maps con-ducive to the following feature learning. In comparison withMADD [ 50] that merely emphasizes the spatial domain, we further introduce the semantic-relevant frequency attentionmap with rich semantic information retained for the subse- quent spatial-frequency relation-discovery paradigm. Finally, to fully discover the spatial and frequency re- lationships, we propose a Dynamic Graph-based Spatial-Frequency Feature Fusion network (DG-SF 3Net) to formu- late the interaction of two domains via a graph-based rela-tion discovery protocol. Speciﬁcally, DG-SF 3Net is com- posed of two ingredients: Dynamic GCN [ 16] and Graph Information Interaction layers. The former constructs kNN-graph dynamically and performs graph convolution to rea- son high-order relationships in spatial and frequency do- mains, while the latter is designed to enhance the mutualrelation with several graph-weighted MLP-Mixer [ 40] lay- ers via channel-wise and node-wise interaction. The achievements, including contributions are threefold: • From a new perspective, we propose a novel Spatial- Frequency Dynamic Graph (SFDG) framework, whichis qualiﬁed to exploit relation-aware spatial-frequency features to promote generalized forgery detection. • We ﬁrst harness a CAF ´E module for content-aware fre- quency feature extraction, and then tailor an MDAMLscheme to dig deeper into multiscale spatial-frequencyattention maps with rich contextual understanding offorgeries. Finally, a seminal DG-SF 3Net module is proposed to discover the multi-domain relationshipswith a graph-based relation-reasoning approach. • Our method achieves state-of-the-art performance on six benchmark datasets. The cross-dataset experiment and the perturbation analysis show the robustness and generalization ability of the proposed SFDG method.
Wang_NeMo_Learning_3D_Neural_Motion_Fields_From_Multiple_Video_Instances_CVPR_2023	Abstract The task of reconstructing 3D human motion has wide- ranging applications. The gold standard Motion capture (MoCap) systems are accurate but inaccessible to the general public due to their cost, hardware, and space constraints. In contrast, monocular human mesh recovery (HMR) methods are much more accessible than MoCap as they take single- view videos as inputs. Replacing the multi-view MoCap systems with a monocular HMR method would break the cur- rent barriers to collecting accurate 3D motion thus making exciting applications like motion analysis and motion-driven animation accessible to the general public. However, the performance of existing HMR methods degrades when the video contains challenging and dynamic motion that is not in existing MoCap datasets used for training. This reducesits appeal as dynamic motion is frequently the target in 3D motion recovery in the aforementioned applications. Our study aims to bridge the gap between monocular HMR and multi-view MoCap systems by leveraging information shared across multiple video instances of the same action. We in- troduce the Neural Motion (NeMo) field. It is optimized to represent the underlying 3D motions across a set of videos of the same action. Empirically, we show that NeMo can re- cover 3D motion in sports using videos from the Penn Action dataset, where NeMo outperforms existing HMR methods in terms of 2D keypoint detection. To further validate NeMo using 3D metrics, we collected a small MoCap dataset mim- icking actions in Penn Action, and show that NeMo achieves better 3D reconstruction compared to various baselines. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22129	1. Introduction Reconstructing 3D human motion has wide-ranging ap- plications from the production of animation movies like Avatar [ 4], human motion synthesis [ 17,44,45] and biome- chanical analysis of motion [ 3,10,16,27,39,40]. Tra- ditional marker-based MoCap systems work by recording 2D infrared images of light reflected by markers placed on the human subject. Placing, calibrating, and labeling the markers are tedious, and the markers can potentially restrict the range-of-motion of the subject. Alternatively, marker- lessMoCap systems work with RGB videos and use com- puter vision techniques to extract the 3D human pose, which eliminates the need of placing physical markers on human subjects. Given a set of synchronized video captures from multiple views, one can run 2D keypoint detection methods like OpenPose [ 5] and perform triangulation to recover the 3D pose [ 35]. Markerless MoCap approaches, however, are still restricted by the assumption of having synchronized multi-view video capture of the same instance of the motion. Meanwhile, there has been a rapid development of 3D monocular human pose estimation (HPE) and human mesh recovery (HMR) methods. These methods aim to recover the 3D human motion from a single-view video capture. The accessibility of monocular HMR makes it an attractive alter- native to existing MoCap systems, especially in situations where setting up additional hardware is challenging like when the human subject is on a bike [13] or underwater [9]. The accuracy of monocular methods is generally lower be- cause single-view input only provides partial information about the underlying 3D motion. The model needs to over- come complications like depth ambiguity and self-occlusion. Like other machine learning systems, HMR models over- come these difficulties by learning from paired training data. Because of the difficulty and cost of collecting MoCap data, paired datasets of videos and 3D motions are scarce. Ad- ditionally, publically available MoCap datasets are often restricted to simple everyday motions like Human3.6M [ 19] and AMASS [ 12]. As a result, existing HMR methods gen- eralize less well in domains with less available MoCap data, such as motions in sports, a dominant application domain of 3D human motion recovery [ 8,15,33,34]. As an example, see Figure 1 for where existing HMR methods struggled to capture the dynamic range of athletic motions. We bridge the gap between multi-view MoCap and monocular HMR by assuming there is shared and comple- mentary information in multiple instances of video captures of the same action, similar to what is in the (same instance) multi-view setup. These multiple video instances can be different repetitions of the same action from the same per- son, or even from different people executing the same action in different settings (See the left side of Figure 2 for illus- tration). For application domains like sports, a key feature of its motions is that they are well-defined and structured.For example, an athlete is also often instructed to practice by performing many repetitions of the same action and the variation across the repetitions is oftentimes slight. Even when we look at the executions of the same action from different athletes, these different motion “instances” often contain shared information. In this work, we aim to better re- construct the underlying 3D motion from videos of multiple instances of the same action in sports. We parametrize each motion as a neural network. It takes as input a scalar phase value indicating the phase/progress of the action and an instance code vector for variation and outputs human joint angles, root orientation, and translation. Since the different sequences are not synchronized, and the actions might progress at slightly different rates (e.g., a faster versus slower pitch), we use an additional learned phase net- work for synchronization. The neural network is shared across all the instances while other components including the instance codes and phase networks are instance-specific. All the components are learned jointly. We optimize using the 2D reprojection loss with respect to the 2D joint key- points of the input videos and prior 3D loss w.r.t. initial predictions from HMR methods to enforce 3D prior. We call the resulting neural network a Neural Motion ( NeMo ) field. NeMo can also be seen as a new test-time optimization scheme for better domain adaptation of 3D HMR similar in spirit to SMPLify [ 2,36]. A key difference is that we leverage shared 3D information at the group level of many instances to learn a canonical motion and its variations. To summarize, our contributions are: •We propose the neural motion (NeMo) field and an optimization framework that improves 3D HMR results by jointly reasoning about different video instances of the same action. •We optimize NeMo fields on sports actions selected from the Penn Action dataset [ 49]. Since the Penn Action dataset only has 2D keypoint annotations, we collected a small MoCap dataset with 3D groundtruth where the actor was instructed to mimic these motions, which we will refer to as our NeMo-MoCap dataset. We show improved 3D motion reconstruction compared to various baseline HMR methods using both 3D metrics, and also improved results on the Penn Action dataset using 2D metrics. •The NeMo field also recovers global root transla- tion. Compared to the recently proposed global HMR method, recovered global motion from NeMo is sub- stantially more accurate on our NeMo-MoCap dataset.
Wimbauer_Behind_the_Scenes_Density_Fields_for_Single_View_Reconstruction_CVPR_2023	Abstract Inferring a meaningful geometric scene representation from a single image is a fundamental problem in computer vision. Approaches based on traditional depth map predic- tion can only reason about areas that are visible in the im- age. Currently, neural radiance fields (NeRFs) can capture true 3D including color, but are too complex to be gener- ated from a single image. As an alternative, we propose to predict an implicit density field from a single image. It maps every location in the frustum of the image to volumetric den- sity. By directly sampling color from the available views instead of storing color in the density field, our scene rep- resentation becomes significantly less complex compared to NeRFs, and a neural network can predict it in a single for- ward pass. The network is trained through self-supervision from only video data. Our formulation allows volume ren- dering to perform both depth prediction and novel view syn- thesis. Through experiments, we show that our method is able to predict meaningful geometry for regions that are oc- cluded in the input image. Additionally, we demonstrate the potential of our approach on three datasets for depth pre- diction and novel-view synthesis.	1. Introduction The ability to infer information about the geometric structure of a scene from a single image is of high impor- tance for a wide range of applications from robotics to aug-mented reality. While traditional computer vision mainly focused on reconstruction from multiple images, in the deep learning age the challenge of inferring a 3D scene from merely a single image has received renewed attention. Traditionally, this problem has been formulated as the task of predicting per-pixel depth values ( i.e. depth maps). One of the most influential lines of work showed that it is possible to train neural networks for accurate single-image depth prediction in a self-supervised way only from video sequences. [14–16, 29, 44, 51, 58, 59, 61] Despite these ad- vances, depth prediction methods are notmodeling the true 3D of the scene: they model only a single depth value per pixel. As a result, it is not directly possible to obtain depth values from views other than the input view without con- sidering interpolation and occlusion. Further, the predicted geometric representation of the scenes does not allow rea- soning about areas that lie behind another object in the im- age ( e.g. a house behind a tree), inhibiting the applicability of monocular depth estimation to 3D understanding. Due to the recent advance of 3D neural fields, the related task of novel view synthesis has also seen a lot of progress. Instead of directly reasoning about the scene geometry, the goal here is to infer a representation that allows rendering views of the scene from novel viewpoints. While geomet- ric properties can often be inferred from the representation, they are usually only a side product and lack visual quality. Even though neural radiance field [32] based methods achieve impressive results, they require many training im- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9076 ages per scene and do not generalize to new scenes. To en- able generalization, efforts have been made to condition the neural network on global or local scene features. However, this has only been shown to work well on simple scenes, for example, scenes containing an object from a single cat- egory [43, 57]. Nevertheless, obtaining a neural radiance field from a single image has not been achieved before. In this work, we tackle the problem of inferring a ge- ometric representation from a single image by generaliz- ing the depth prediction formulation to a continuous den- sity field. Concretely, our architecture contains an encoder- decoder network that predicts a dense feature map from the input image. This feature map locally conditions a density field inside the camera frustum, which can be evaluated at any spatial point through a multi-layer perceptron (MLP). The MLP is fed with the coordinates of the point and the feature sampled from the predicted feature map by repro- jecting points into the camera view. To train our method, we rely on simple image reconstruction losses. Our method achieves robust generalization and accu- rate geometry prediction even in very challenging outdoor scenes through three key novelties: 1. Color sampling. When performing volume render- ing, we sample color values directly from the input frames through reprojection instead of using the MLP to predict color values. We find that only predicting density drasti- cally reduces the complexity of the function the network has to learn. Further, it forces the model to adhere to the multi-view consistency assumption during training, leading to more accurate geometry predictions. 2. Shifting capacity to the feature extractor. In many previous works, an encoder extracts image features to con- dition local appearance, while a high-capacity MLP is ex- pected to generalize to multiple scenes. However, on com- plex and diverse datasets, the training signal is too noisy for the MLP to learn meaningful priors. To enable robust train- ing, we significantly reduce the capacity of the MLP and use a more powerful encoder-decoder that can capture the entire scene in the extracted features. The MLP then only evaluates those features locally. 3.Behind the Scenes loss formulation. The continuous nature of density fields and color sampling allow us to re- construct a novel view from the colors of any frame, not just the input frame. By applying a reconstruction loss between two frames that both observe areas occluded in the input frame, we train our model to predict meaningful geometry everywhere in the camera frustum, not just the visible areas. We demonstrate the potential of our new approach in a number of experiments on different datasets regarding the aspects of capturing true 3D, depth estimation, and novel view synthesis. On KITTI [12] and KITTI-360 [26], we show both qualitatively and quantitatively that our model can indeed capture true 3D, and that our modelachieves state-of-the-art depth estimation accuracy. On RealEstate10K [45] and KITTI, we achieve competitive novel view synthesis results, even though our method is purely geometry-based. Further, we perform thorough ab- lation studies to highlight the impact of our design choices.
Xu_Learning_To_Generate_Image_Embeddings_With_User-Level_Differential_Privacy_CVPR_2023	Abstract Small on-device models have been successfully trained with user-level differential privacy (DP) for next word pre- diction and image classification tasks in the past. However, existing methods can fail when directly applied to learn em- bedding models using supervised training data with a large class space. To achieve user-level DP for large image- to-embedding feature extractors, we propose DP-FedEmb, a variant of federated learning algorithms with per-user sensitivity control and noise addition, to train from user- partitioned data centralized in the datacenter. DP-FedEmb combines virtual clients, partial aggregation, private local fine-tuning, and public pretraining to achieve strong pri- vacy utility trade-offs. We apply DP-FedEmb to train im- age embedding models for faces, landmarks and natural species, and demonstrate its superior utility under same privacy budget on benchmark datasets DigiFace, EMNIST, GLD and iNaturalist. We further illustrate it is possible to achieve strong user-level DP guarantees of ϵ <2while con- trolling the utility drop within 5%, when millions of users can participate in training .	1. Introduction Representation learning, by training deep neural net- works as feature extractors to generate compact embedding vectors from images, is a fundamental component in com- puter vision. Metric learning, a kind of representation learn- ing using supervised data, has been widely applied to im- age recognition, clustering, and retrieval [61, 75, 77]. Ma- chine learning models have the capacity to memorize train- ing data [10, 11], leading to privacy risks when the models are deployed. Privacy risk can also be audited by member- ship inference attacks [9, 63], i.e. detecting whether cer- tain data was used to train a model and potentially expos- ing users’ usage behaviors. Defending against such risks is a critical responsibility when training on privacy-sensitive data. *The first two authors contributed equally. Correspondence to Zheng [email protected] Privacy (DP) [23] is an extensively used quantifiable measurement of privacy risk, now generally ac- cepted as a standard notion of privacy in both industry and government [5, 18, 50, 70]. Applied to machine learning, DP requires a training procedure with explicit randomness, and guarantees that the distribution over output models is quantifiably similar given a certain scope of change to the training dataset. A DP guarantee with respect to the change of a single arbitrary training example is known as example- level DP , which provides plausible deniability (in the binary hypothesis testing sense of [38]) that any single example (e.g., image) occurred in the training dataset. If we instead consider how the distribution of output models changes if the data (including even the number of examples) from any single user change arbitrarily, we have user-level DP [22]. This ensures model training is quantifiably insensitive to all of the data from any one user, and hence it is impossible to tell if a user has participated in training with high con- fidence. This guarantee can be exponentially stronger than example-level DP if one user may contribute many exam- ples to training. Recently, DP-SGD [1] (essentially, SGD with the ad- ditional steps of clipping each individual gradient to have a maximum norm, and adding correspondingly calibrated noise) has been used to achieve example-level DP for rel- atively large models in language modeling and image clas- sification tasks [4, 16, 42, 45, 81], often utilizing techniques like large batch training and pretraining on public data. DP- SGD can be modified to guarantee user-level DP, which is often combined with federated learning algorithms and called DP-FedAvg [49]. User-level DP has only been stud- ied for small on-device models that have less than 10 mil- lion parameters [36, 49, 57]. We consider user-level DP for relatively large models in representation learning with supervised data. In our set- ting, similar to federated learning (FL), the data are user- partitioned; but in contrast to decentralized FL, we are pri- marily motivated
Xu_Learning_Dynamic_Style_Kernels_for_Artistic_Style_Transfer_CVPR_2023	Abstract Arbitrary style transfer has been demonstrated to be ef- ficient in artistic image generation. Previous methods ei- ther globally modulate the content feature ignoring local details, or overly focus on the local structure details lead- ing to style leakage. In contrast to the literature, we propose a new scheme “style kernel” that learns spatially adaptive kernels for per-pixel stylization, where the convolutional kernels are dynamically generated from the global style- content aligned feature and then the learned kernels are applied to modulate the content feature at each spatial posi- tion. This new scheme allows flexible both global and local interactions between the content and style features such that the wanted styles can be easily transferred to the content im- age while at the same time the content structure can be eas- ily preserved. To further enhance the flexibility of our style transfer method, we propose a Style Alignment Encoding (SAE) module complemented with a Content-based Gating Modulation (CGM) module for learning the dynamic style kernels in focusing regions. Extensive experiments strongly demonstrate that our proposed method outperforms state- of-the-art methods and exhibits superior performance in terms of visual quality and efficiency.	1. Introduction Artistic style transfer [48] refers to a hot computer vi- sion technology that allows us to recompose the content of an image in the style of an artistic work. Figure 1 shows several vivid examples. We might have ever imagined what a photo might look like if it were painted by a famous artist like Pablo Picasso or Van Gogh. Now style transfer is the computer vision technique that turns this into a reality. It has great potential values in various real-world applications and therefore attracts a lot of researchers to constantly put efforts to make progress towards both quality and efficiency. Most of existing style transfer works [6, 15, 18, 26, 33] either globally modulate the content feature ignoring local details or overly focus on the local structure details leading to style leakage. In particular, [15,18] seeks to match global statistics between content and style images, resulting in in- consistent stylizations that either remain large parts of the content unchanged or contain local content with distorted style patterns. SANet [33] and AdaAttN [31] improve the content similarity by learning attention maps that match the semantics between the style and stylized images, but these methods tend to distort object instances when improper style patterns are involved. Recently, IEC [5] adopts con- trastive learning to pull close both the content and style rep- resentations between input and output images. MAST [16] This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10083 seeks to balance the style and content via manifold align- ment. Although both IEC and MAST have achieved some success in most cases, they are still far from satisfactory to well balance two important requirements, i.e., style consis- tency and structure similarity. In this paper, we develop a novel style transfer frame- work in a manner of encoder-decoder structure, as illus- trated in Figure 2, which contains two important modules: (1) a Style Alignment Encoding (SAE) module enhanced by Content-based Gating Modulation (CGM), which is used to generate content-style aligned features, and (2) a Style Ker- nel Generation (SKG) module used to transform the output features of SAE to convolutional kernels. In such a frame- work, we propose a new style transfer scheme, “style ker- nel”, which treats the features of the style images as dy- namic convolutional kernels and then transfer the style in- formation to the content image by convolving the content image by the learned dynamic style kernels. This allows fine-grained local interactions between the style and content features, the flexibility of which makes both style transfer- ring and content structure preserving much easier. To enforce the global correlation between the content and style features, we simulate the self-attention mecha- nism of Transformer [41] in the SAE module. This treats the content and style features as the queries and keys of the self-attention operator respectively and computes a large at- tention map the elements of which measure for each local content context the similarity to each local style context. With the attention map, we can aggregate for each pixel of the content image all the style features in a weighted man- ner to obtain content-style aligned features. We also design a Content-based Gating Modulation (CGM) operator that further thresholds the attention map and zeros the places where similarities are too small, seeking an adaptive num- ber of correlated neighbors as a focusing region for each query point. Then, we use another set of convolutions that constitute the SKG module to further transform the content- style aligned features, and view (reshape) the output of SKG as dynamic style kernels. In order to improve the efficiency of the dynamic convolutions, SKG predicts separable local filters (two 1D filters and a bias) instead of a spatial 2D filter that has more parameters. The learned style kernels already integrate the information of the given content and style im- ages. We further apply the learned dynamic style kernels to the content image feature for transferring the target style to the content image. We shall emphasize that unlike “style code” in AdaIN [15], DRB-GAN [48], and AdaConv [2] modeled as the dynamic parameters ( e.g. mean, variance, convolution kernel) which are shared over all spatial positions globally without distinguishing the semantic regions, the learned dy- namic kernels via our “style kernel” are point-wisely mod- eled upon the globally style-content aligned feature, andtherefore is able to make full use of the point-wise semantic structure correlation to modulate content feature for better artistic style transfer. Our learned dynamic style kernels are good at transferring the globally aggregated and locally semantic-aligned style features to local regions. This en- sures our style transfer model that works well on consis- tent stylization with well preserved structure, overcoming the shortages of existing style transfer methods which either globally modulate the content feature ignoring local details (e.g. AdaIN, WCT [28]) or overly focus on the local struc- ture details leading to style leakage ( e.g. AdaAttN [31] and MAST [16]). The main contributions of this work are 3-fold as fol- lows: • We are the first to propose the “style kernel” scheme for artistic style transfer, which converts the globally style-content alignment features into position point- wisely dynamic convolutional kernels to convolve the features of content images together for style transfer. • We design a novel architecture, i.e. SAE and CGM, to generate the dynamic style kernels by employing the correlation between the content and style image adap- tively. • Extensive experiments demonstrate that our method produces visually plausible stylization results with fine-grained style details and coherence content with the input content images.
Wang_Learning_Transformation-Predictive_Representations_for_Detection_and_Description_of_Local_Features_CVPR_2023	Abstract The task of key-points detection and description is to es- timate the stable location and discriminative representa- tion of local features, which is a fundamental task in vi- sual applications. However, either the rough hard positive or negative labels generated from one-to-one correspon- dences among images may bring indistinguishable samples, like false positives or negatives, which acts as inconsis- tent supervision. Such resultant false samples mixed with hard samples prevent neural networks from learning de- scriptions for more accurate matching. To tackle this chal- lenge, we propose to learn the transformation-predictive representations with self-supervised contrastive learning. We maximize the similarity between corresponding views of the same 3D point (landmark) by using none of the neg- ative sample pairs and avoiding collapsing solutions. Fur- thermore, we adopt self-supervised generation learning and curriculum learning to soften the hard positive labels into soft continuous targets. The aggressively updated soft la- bels contribute to overcoming the training bottleneck (de- rived from the label noise of false positives) and facili- tating the model training under a stronger transformation paradigm. Our self-supervised training pipeline greatly de- creases the computation load and memory usage, and out- performs the sota on the standard image matching bench- marks by noticeable margins, demonstrating excellent gen- eralization capability on multiple downstream tasks.	1. Introduction Local visual descriptors are fundamental to various com- puter vision applications such as camera calibration [37], 3D reconstruction[19], visual simultaneous localization and mapping (VSLAM) [33], and image retrieval [38]. The descriptors indicate the representation vector of the patch around the key-points and can be used to generate dense *Corresponding authorcorrespondences between images. The discriptors are highly dependent on the effective rep- resentation, which has been always trained with the Siamese architecture and contrastive learning loss [12, 29, 39]. The core idea of contrastive learning is “learn to compare”: given an anchor key-point, distinguish a similar (or pos- itive) sample from a set of dissimilar (or negative ) sam- ples, in a projected embedding space. The induced repre- sentations present two key properties: 1) alignment of fea- tures from positive pairs, 2) and uniformity of representation on the hypersphere [51]. Negative samples are thus intro- duced to keep the uniformity property and avoid model col- lapse, i.e., preventing the convergence to one constant solu- tion [13]. Therefore, various methods have been proposed to mine hard negatives[6, 21, 55]. However, these methods raise the computational load and memory resources usage heavily [17]. More Importantly, within the hard negatives, some samples are labeled as negatives, but actually have the identical semantics of the anchor ( i.e.,false negatives ). These false negatives act as inconsistent supervision and prevent the learning-based models from achieving higher accuracy[4]. More concretely, the false negatives represent the instance located on the repetitive texture in the structural dataset, as shown in Figure 1. It is challenging to recognize such false negatives from true negatives[39]. The recent active self-supervised learning methods[5, 8, 9, 15] motivate us to rethink the effectiveness of negatives in descriptors learning. We propose to learn the transfor- mation predictive representations (TPR) for visual descrip- tors only with the positives and avoids collapsing solutions. Furthermore, using none of negatives greatly improves the training efficiency by reducing the scale of the similarity matrix from O(n2)toO(n), and reduces the computation load and memory usage. To further improve the generalization performance of descriptors, hard positives (i.e., corresponding pairs with large-scale transformation) are encouraged as training data to expose novel patterns. However, recent experiments have shown that directly contrastive learning on stronger This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11464 Query EasyNegativesHardNegativesFalseNegativesSemanticallydissimilarSemanticallysimilar,butnotidenticalSemanticallyidentical,butlabeldissimilar t-SNEvisualizationFigure 1. Illustrative schematic of different negative samples. Easy negatives (green) are easy to distinguish from the query (blue) and not sufficient for good performance. Hard negatives (orange) are important for better performance and are still semantically dissimilar from the query. False negatives (red) are practically impossible to distinguish and semantically identical with the query, which is harmful to the performance. The right figure visualizes the embedding of different samples in latent space. transformed images can not learn representations effec- tively [52]. In addition, current contrastive learning meth- ods label all positives with different transformation strength as coarse “1”, which prevent learning refined representa- tion. We propose to learn transformation predictive repre- sentation with soft positive labels from (0,1] instead of “1” to supervise the learning of local descriptors. Furthermore, we propose a self-supervised curriculum learning module to generate controllable stronger positives with gradually re- fined soft supervision as the network iterative training. Finally, our TPR with soft labels is trained on natural images in a fully self-supervised paradigm. Different from previous methods trained on datasets with SfM or cam- era pose information [20, 29, 39, 49], our training datasets are generally easy to collect and scale up since there is no extra annotation requirement to capture dense correspon- dences. Experiments show that our self-supervised method outperforms the state-of-the-art on standard image match- ing benchmarks by noticeable margins and shows excel- lent generalization capability on multiple downstream tasks (e.g., visual odometry, and localization). Our contributions to this work are as follows: i) we pro- pose to learn transformation-predictive representations for joint local feature learning, using none of the negative sam- ple pairs and avoiding collapsing solutions. ii) We adopt self-supervised generation learning and curriculum learn- ing to soften the hard positives into continuous soft labels, which can alleviate the false positives and train the model with stronger transformation. iii) The overall pipeline is trained with the self-supervised paradigm, and the training data are computed from random affine transformation and augmentation on natural images.
Vaksman_Patch-Craft_Self-Supervised_Training_for_Correlated_Image_Denoising_CVPR_2023	Abstract Supervised neural networks are known to achieve excel- lent results in various image restoration tasks. However, such training requires datasets composed of pairs of cor- rupted images and their corresponding ground truth tar- gets. Unfortunately, such data is not available in many ap- plications. For the task of image denoising in which the noise statistics is unknown, several self-supervised training methods have been proposed for overcoming this difficulty. Some of these require knowledge of the noise model, while others assume that the contaminating noise is uncorrelated, both assumptions are too limiting for many practical needs. This work proposes a novel self-supervised training tech- nique suitable for the removal of unknown correlated noise. The proposed approach neither requires knowledge of the noise model nor access to ground truth targets. The in- put to our algorithm consists of easily captured bursts of noisy shots. Our algorithm constructs artificial patch-craft images from these bursts by patch matching and stitching, and the obtained crafted images are used as targets for the training. Our method does not require registration of the images within the burst. We evaluate the proposed frame- work through extensive experiments with synthetic and real image noise.	1. Introduction Supervised neural networks have proven themselves powerful, achieving impressive results in solving image restoration problems (e.g., [14–16, 18, 33, 41–43]). In the commonly deployed supervised training for such tasks, one needs a dataset consisting of pairs of corrupted and ground truth images. The degraded images are fed to the network input, while the ground truth counterparts are used as guid- ing targets. When the degradation model is known and easy to implement, one can construct such a dataset by applying This research was partially supported by the Israel Science Foundation (ISF) under Grant 335/18 and the Council For Higher Education - Planning & Budgeting Committee.the degradation to clean images. However, a problem arises when the degradation model is unknown. In such cases, while it is relatively easy to acquire distorted images, ob- taining their ground truth counterparts can be challenging. For this reason, there is a need for self-supervised methods that use corrupted images only in the training phase. More on these methods is detailed in Section 2. In this work we focus on the problem of image denoising with an unknown noise model. More specifically, we as- sume that the noise is additive, zero mean, but not necessar- ily Gaussian, and one that could be cross-channel and short- range spatially correlated1. We additionally assume that the noise is (mostly) independent of the image and nearly ho- mogeneous, i.e., having low to moderate spatially variant statistics. Examples of such noise could be Gaussian corre- lated noise or real image noise in digital cameras. Several recent papers propose methods for self-supervised training under similar such challenging conditions. However, they all assume an uncorrelated noise or a noise with a known model, thus limiting their coverage of the need posed. This work proposes a novel self-supervised training framework for addressing the problem of image denois- ing of an unknown correlated noise. The proposed algo- rithm gets as input bursts of shots, where each frame in the burst captures nearly the same scene, up to moderate movements of the camera and objects. Such sequences of images are easily captured in many digital cameras. Our algorithm uses one image from the burst as the input, utiliz- ing the rest of the frames of the same burst for constructing (noisy) targets for training. For creating these target images, we harness the concept of patch-craft frames introduced in PaCNet [35]. Similar to PaCNet, we split the input shot into fully overlapping patches. For each patch, we find its nearest neighbor within the rest of the burst images. Note that, unlike PaCNet, we strictly omit the input shot from the neighbor search. We proceed by building mpatch-craft 1By short-term we refer to the case in which the auto-correlation func- tion decays fast, implying that only nearby noise pixels may be highly cor- related. The correlation range we consider is governed by the patch size in our algorithm - see Section 3 for more details. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5795 images by stitching the found neighbor patches, where m is the patch size, and use these frames as denoising targets. The above can be easily extended by using more than one nearest neighbor per patch, this way enriching dramatically the number of patch-craft frames and their diversity. The proposed technique for creating artificial target im- ages is sensitive to the possibility of getting statistical de- pendency between the input and the target noise. To com- bat this flaw, we propose a method for statistical analysis of the target noise. This analysis suggests simple actions that reduce dependency between the target noise and the denoiser’s input, leading to a significant boost in perfor- mance. We evaluate the proposed framework through ex- tensive experiments with synthetic and real image noise, showing that the proposed framework outperforms leading self-supervised methods. To summarize, the contributions of this work are the following: • We propose a novel self-supervised framework for training an image denoiser, where the noise may be cross-channel and short-range spatially correlated. Our approach relies simply on the availability of bursts of noisy images; the ground truth is unavailable, and the noise model is unknown. • We suggest a method for statistical analysis of the tar- get noise that leads to a boost in performance. • We demonstrate superior denoising performance com- pared to leading alternative self-supervised denoising methods.
Xie_Category_Query_Learning_for_Human-Object_Interaction_Classification_CVPR_2023	Abstract Unlike most previous HOI methods that focus on learn- ing better human-object features, we propose a novel and complementary approach called category query learning . Such queries are explicitly associated to interaction cat- egories, converted to image specific category representa- tion via a transformer decoder, and learnt via an auxil- iary image-level classification task. This idea is motivated by an earlier multi-label image classification method, but is for the first time applied for the challenging human- object interaction classification task. Our method is sim- ple, general and effective. It is validated on three rep- resentative HOI baselines and achieves new state-of-the- art results on two benchmarks. Code will be available at https://github.com/charles-xie/CQL .	1. Introduction Human-Object Interaction (HOI) detection has attracted a lot of interests in recent years [3, 8–10, 21, 33]. The task consists of two sub-tasks. The first is human and object de- tection. It is usually performed by common object detection methods. The second is interaction classification of each human-object (HO) pair. This sub-task is very challenging due to the complex appearance variations in the interaction categories. See Fig. 1 for examples. It is the focus of most previous HOI methods, as well as this work. Most previous HOI methods focus on learning bet- ter human-object features , including modeling relation and context via GNN [7, 29, 34, 37] or attention mecha- nism [8, 34, 46], decoupling localization and classification [22, 41, 48], leveraging vision-language knowledge [6, 22] and introducing multi-scale feature to transformer [16]. However, for interaction classification they all adopt the simple linear classifier that performs the dot product of the *Corresponding author. †Work done during internship at MEGVII technology. ‡Work done while worked at MEGVII. (a) personfly kite (b) personholdforkpersonholdelephantpersonflyairplaneFigure 1. Interaction classification is inherently challenging. In (a), “fly” is semantically polysemic, resulting in different objects, poses and relative positions. In (b), when “hold” is associated with different objects, the appearance, scene background, and human poses are largely different. human-object feature and a static weight vector, which rep- resents an interaction category. In this work, we propose a new approach that enhances the above paradigm and complements most previous HOI methods. It is motivated by the recent work Query2label [25], a transformer-based classification network. It pro- poses a new concept we call category-specific query . Unlike the queries in other transformer methods, each query is as- sociated to a specific and fixed image category during train- ing and inference. This one-to-one binding makes the query learn to model each category more effectively. The queries are converted to image specific category representations via a transformer decoder. This method achieves excellent per- formance on multi-label image classification task. We extend this approach for human-object interaction classification. Essentially, our approach replaces traditional category representation as a static weight vector in previ- ous HOI methods with category queries learnt as described above. The same linear classifier is adopted. Such cate- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15275 gory queries are more effective, and adaptive for different images, giving rise to better modeling of the complex vari- ations in each interaction category. This is the crucial dif- ference between this work and a simple adaption of [25] to HOI. Notably, this work is the first to address the category weight representation problem in the HOI community. Note that our proposed category specific query is differ- ent and not related to those queries in other transformer- based HOI methods [4, 15, 33, 49]. Specifically, category queries extract image-level features as the category repre- sentation. The queries in other methods are human-object instance-level features and category-agnostic. Our method is simple, lightweight and general. The overview is in Fig. 2. It is complementary to any off-the- shelf HOI method that provides human-object features. The modification of both inference and training is small. The in- curred additional cost is marginal. In experiments, our approach is validated on three representative and strong HOI baseline methods, two transformer-based methods [22, 33] and a traditional two- stage method [42]. They are all significantly improved by our approach. New state-of-the-art results are obtained on two benchmarks. Specifically, we obtain 36.03 mAP on HICO-DET. Comprehensive ablation studies and in-depth discussion are also provided to verify the effectiveness of implementation details in our approach. It turns out that our method is more effective on challenging images that con- tain more human-object instances, a property that is rarely discussed by previous HOI methods.
Wu_Learning_Semantic-Aware_Knowledge_Guidance_for_Low-Light_Image_Enhancement_CVPR_2023	Abstract Low-light image enhancement (LLIE) investigates how to improve illumination and produce normal-light images. The majority of existing methods improve low-light images via a global and uniform manner, without taking into ac- count the semantic information of different regions. With- out semantic priors, a network may easily deviate from a region’s original color. To address this issue, we propose a novel semantic-aware knowledge-guided framework (SKF) that can assist a low-light enhancement model in learning rich and diverse priors encapsulated in a semantic segmen- tation model. We concentrate on incorporating semantic knowledge from three key aspects: a semantic-aware em- bedding module that wisely integrates semantic priors in feature representation space, a semantic-guided color his- togram loss that preserves color consistency of various in- stances, and a semantic-guided adversarial loss that pro- duces more natural textures by semantic priors. Our SKF is appealing in acting as a general framework in LLIE task. Extensive experiments show that models equipped with the SKF significantly outperform the baselines on mul- tiple datasets and our SKF generalizes to different models and scenes well. The code is available at Semantic-Aware- Low-Light-Image-Enhancement	1. Introduction In real world, low-light imaging is fairly common due to unavoidable environmental or technical constraints such asinsufficient illumination and limited exposure time. Low- light images not only have poor visibility for human per- ception, but also are unsuitable for subsequent multime- dia computing and downstream vision tasks designed for high-quality images [4, 9, 36]. Thus, low-light image en- hancement (LLIE) is proposed to reveal buried details in low-light images and avoid degraded performance in sub- sequent vision tasks. Mainstream traditional methods for LLIE include Histogram Equalization-based methods [2] and Retinex model-based methods [18]. Recently, many deep learning-based LLIE methods have proposed, such as end-to-end frameworks [5,7,34,45,46,48] and Retinex-based frameworks [29, 41, 43, 44, 49, 53, 54]. Benefiting from their ability in modeling the mapping be- tween the low-light and high-quality image, deep LLIE methods commonly achieve better results than traditional approaches. However, existing methods typically improve low-light images globally and uniformly, without taking into account the semantic information of different regions, which is crucial for enhancement. As shown in Fig. 1(a), a network that lacks the utilization of semantic priors can easily deviate from a region’s original hue [22]. Further- more, studies have demonstrated the significance of incor- porating semantic priors into low-light enhancement. Fan et al. [8] utilize semantic map as prior and incorporated it into the feature representation space, thereby enhancing image quality. Rather than relying on optimizing intermediate fea- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1662 tures, Zheng et al. [58] adopt a novel loss to guarantee the semantic consistency of the enhanced images. These meth- ods successfully combine the semantic priors with LLIE task, demonstrating the superiority of semantic constraints and guidance. However, their methods fail to fully ex- ploit the knowledge that semantic segmentation networks can provide, limiting the performance gain by semantic pri- ors. Furthermore, the interaction between segmentation and enhancement is designed for specific methods, limiting the possibility of incorporating semantic guidance into LLIE task. Hence, we wonder two questions: 1. How can we obtain various and available semantic knowledge? 2. How does semantic knowledge contribute to image quality im- provement in LLIE task? We attempt to answer the first question. First, a semantic segmentation network pre-trained on large-scale datasets is introduced as a semantic knowledge bank (SKB). The SKB can provide richer and more diverse semantic priors to im- prove the capability of enhancement networks. Second, ac- cording to previous works [8, 19, 58], the available priors provided by the SKB primarily consist of intermediate fea- tures and semantic maps. Once training a LLIE model, the SKB yields above semantic priors and guides the enhance- ment process. The priors can not only refine image fea- tures by employing techniques like affinity matrices, spatial feature transformations [40], and attention mechanisms, but also guide the design of objective functions by explicitly incorporating regional information into LLIE task [26]. Then we try to answer the second question. We design a series of novel methods to integrate semantic knowledge into LLIE task based on the above answers, formulating in a novel semantic-aware knowledge-guided framework (SKF). First, we use the High-Resolution Network [38] (HRNet) pre-trained on the PASCAL-Context dataset [35] as the previously mentioned SKB. In order to make use of intermediate features, we develop a semantic-aware embed- ding (SE) module. It computes the similarity between the reference and target features and employs cross-modal in- teractions between heterogeneous representations. As a re- sult, we quantify the semantic awareness of image features as a form of attention and embed semantic consistency in enhancement network. Second, some methods [20, 55] propose to optimize im- age enhancement using color histogram in order to preserve the color consistency of the image rather than simply en- hancing the brightness globally. The color histogram, on the other hand, is still a global statistical feature that cannot guarantee local consistency. Hence, we propose a semantic- guided color histogram (SCH) loss to refine color consis- tency. Here, we intend to make use of local geometric information derived from the scene semantics and global color information derived from the content. In addition to guarantee original color of the enhanced image, it can alsoadd spatial information to the color histogram, performing a more nuanced color recovery. Third, existing loss functions are not well aligned with human perception and fail to capture an image’s intrinsic signal structure, resulting in unpleasing visual results. To improve visual quality, EnlightenGAN [16] employs global and local image-content consistency and randomly chooses the local patch. However, the discriminator do not know where the regions are likely to be ‘fake’. Thus, we propose a semantic-guided adversarial (SA) loss. Specifically, the ability of the discriminator is improved by using segmen- tation map to determine the fake areas, which can improve the image quality further. The main contributions of our work are as follows: • We propose a semantic-aware knowledge-guided framework (SKF) to boost the performance of exist- ing methods by jointly maintaining color consistency and improving image quality. • We propose three key techniques to take full advantage of semantic priors provided by semantic knowledge bank (SKB): semantic-aware embedding (SE) mod- ule, semantic-guided color histogram (SCH) loss, and semantic-guided adversarial (SA) loss. • We conduct experiments on LOL/LOL-v2 datasets and unpaired datasets. The experimental results demon- strate large performance improvements by our SKF, verifying its effectiveness in resolving the LLIE task.
Wang_Compression-Aware_Video_Super-Resolution_CVPR_2023	Abstract Videos stored on mobile devices or delivered on the In- ternet are usually in compressed format and are of various unknown compression parameters, but most video super- resolution (VSR) methods often assume ideal inputs result- ing in large performance gap between experimental set- tings and real-world applications. In spite of a few pio- neering works being proposed recently to super-resolve the compressed videos, they are not specially designed to deal with videos of various levels of compression. In this pa- per, we propose a novel and practical compression-aware video super-resolution model, which could adapt its video enhancement process to the estimated compression level. A compression encoder is designed to model compression levels of input frames, and a base VSR model is then condi- tioned on the implicitly computed representation by insert- ing compression-aware modules. In addition, we propose to further strengthen the VSR model by taking full advan- tage of meta data that is embedded naturally in compressed video streams in the procedure of information fusion. Ex- tensive experiments are conducted to demonstrate the ef- fectiveness and efficiency of the proposed method on com- pressed VSR benchmarks. The codes will be available at https://github.com/aprBlue/CAVSR	1. Introduction Video super-resolution aims at restoring a sequence of high-resolution (HR) frames by utilizing the comple- mentary temporal information within low-resolution (LR) frames. There have been many efforts [1–3, 7, 13, 15–18, 20, 24, 35, 38, 47] made on this task, especially after the rise of deep learning. Most of these methods, however, *T. ISOBE and Y . WANG contributed equally to this work. †Corresponding author. CRF15CRF35 MobilePhoneLRVideoTranscoderCloud ServerVSRDecoderBitstreamFramesPreviousMethodsMetaDataCompression-awareVSROurMethodCompressedVSRUploading BasicVSROursBicubicCRF15CRF35Train TestSTDF+BasicVSR CRF15CRF35（a）（b）（c）Figure 1. Motivation of this work. (a) Application scenario of the VSR task this work focuses on, (b) performance of existing VSR methods on compressed VSR task, and (c) performance of previous VSR models trained on videos of different compression. assume an ideal input such as directly taking either bicu- bicly downsampled frames or Gaussian smoothed and dec- imated ones as the degraded inputs. In real world, videos stored on mobile devices or delivered on the Internet are all in a compressed format with different compression lev- els [10, 11, 14, 25, 30, 33, 41, 45]. Unless the compression is very lightweight, directly applying an existing VSR model would give unsatisfactory results with magnified compres- sion artifacts, as shown in Fig. 1(a). One straightforward so- lution is to first apply a multi-frame decompression method [5, 9, 41, 45] to remove blocking artifacts and then feed the enhanced frames to an uncompressed VSR model. How- ever, as shown in Fig. 1(b), the performance is still not good with artifacts remained. In addition, the decompression net- work usually cannot handle video frames of different com- pression levels adaptively, which will cause over-smoothing This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2012 and accumulate errors in super-resolution stage. Recently a few pioneering works have been proposed to investigate the video super-resolution task on compressed videos. In [46], Yang et al. took into consideration com- plex degradations in real video scenarios and built a real- world VSR dataset using iPhone 11 Pro Max. A special training strategy is proposed to handle misalignment and il- lumination/color difference. In RealBasicVSR [3], Chan et al. synthesized real-world-like low-quality videos based on a combination of several degradations and proposed a two- stage method with an image pre-cleaning module followed by an existing VSR model. COMISR [22] and FTVSR [27] are proposed to address streamed videos compressed in dif- ferent levels rather than degradations like noise and blur. Although COMISR and FTVSR have improved perfor- mance on compressed videos, they are not specially de- signed to deal with videos of various levels of compres- sion. Being aware of compression with input videos would allow a model to exert its power on those videos adap- tively. Otherwise, video frames with less compression would be oversmoothed while the ones with heavy com- pression would still remain magnified artifacts, as shown in Fig. 1(c). These methods feed themselves with only com- pressed video frames as input, however, meta data such as frame type, motion vectors and residual maps that are nat- urally encoded with a compressed video are ignored. Mak- ing full use of such meta data and the decoded video frames could help further improve super-resolution performance on compressed videos. Based on the above observations, we propose a compression-aware video super-resolution model, a com- pression encoder module is designed to implicitly model compression level with the help of meta data of a com- pressed video. It would also take into account both frames and their frame types in computing compression represen- tation. A base bidirectional recurrent-based VSR model is then conditioned on that representation by inserting compression-aware modules such that it could adaptively deal with videos of different compression levels. To fur- ther strengthen the power of the base VSR model, we take advantage of meta data in a further step. Motion vectors and residual maps are employed to achieve fast and accu- rate alignment between different time steps and frame types are leveraged again to update hidden state in bidirectional recurrent network. Extensive experiments demonstrate that the specially designed VSR model for compressed videos performs favorably against state-of-the-art methods. Our contributions are summarized as follows: • A compression encoder to perceive compression levels of frames is proposed. It is supervised with a ranking- based loss and the computed compression representa- tion is used to modulate a base VSR model. • Meta data that comes naturally with compressedvideos are fully explored in fusion process of spatial and temporal information to strengthen the power of a bidirectional RNN-based VSR model. • Extensive experiments demonstrate the effectiveness and efficiency of the proposed method on compressed VSR benchmarks.
Wang_ARO-Net_Learning_Implicit_Fields_From_Anchored_Radial_Observations_CVPR_2023	Abstract We introduce anchored radial observations (ARO), a novel shape encoding for learning implicit field representation of 3D shapes that is category-agnostic andgeneralizable amid significant shape variations. The main idea behind our work is to reason about shapes through partial observations from a set of viewpoints, called anchors. We develop a general and unified shape representation by employing a fixed set of an- chors, via Fibonacci sampling, and designing a coordinate- based deep neural network to predict the occupancy value of a query point in space. Differently from prior neural implicit models that use global shape feature, our shape encoder operates on contextual, query-specific features. To predict point occupancy, locally observed shape information from the perspective of the anchors surrounding the input query point are encoded and aggregated through an attention mod- ule, before implicit decoding is performed. We demonstrate the quality and generality of our network, coined ARO-Net, on surface reconstruction from sparse point clouds, with tests on novel and unseen object categories, “one-shape” training, and comparisons to state-of-the-art neural and classical methods for reconstruction and tessellation.	1. Introduction Despite the substantial progress made in deep learning in recent years, transferability and generalizability issues still persist due to domain shifts and the need to handle diverse, out-of-distribution test cases. For 3D shape representation learning, a reoccurring challenge has been the inability of trained neural models to generalize to unseen object cate- gories and diverse shape structures, as these models often overfit to or “memorize" the training data. In this paper, we introduce a novel shape encoding for learning an implicit field [34] representation of 3D shapes that is category-agnostic andgeneralizable amid significant shape variations. In Figure 1 (top), we show 3D reconstruc- tions from sparse point clouds obtained by our approach that is trained on chairs but tested on airplanes, riffles, and *Equal contribution †Corresponding author. E-mail: [email protected] Airplanes Rifles Animals Input ARO-Net GT mesh Input ARO-Net GT mesh Trainingdata 3D Reconstruction from sparse point clouds by ARO -Net trained on 4K chairs ARO-Net trained on only the Fertility model with rotation and scaling ......Figure 1. Neural 3D reconstruction using ARO-Net from sparse point clouds (1,024 or 2,048 points). Top: reconstruction results on airplanes, riffles, and animals when the network was trained only on chairs. Bottom: reconstruction of a variety of shapes when the network was train on numerous versions of one model - the Fertility. See comparison to other methods in Section 4. animals. Our model can even reconstruct a variety of shapes when the training data consists of only one shape, augmented with rotation and scaling; see Figure 1 (bottom). The main idea behind our work is to reason about shapes from partial observations at a set of viewpoints, called an- chors , and apply this reasoning to learn implicit fields. We develop a general and unified shape representation by des- ignating a fixed set of anchors and designing a coordinate- based neural network to predict the occupancy at a query point. In contrast to classical neural implicit models such as IM-Net [8], OccNet [23], and DeepSDF [24], which learn global shape features for occupancy/distance prediction, our novel encoding scheme operates on local shape features obtained by viewing the query point from the anchors. Specifically, for a given query point x, we collect locally observable shape information surrounding x, as well as direc- tional and distance information toward x, from the perspec- tive of the set of anchors, and encode such information using This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3572 /gid00009/gid00064/gid00010/gid00001/gid00034/gid00077/gid00066/gid00071/gid00078/gid00081/gid00082/gid00001/gid00064/gid00077/gid00067/gid00001/gid00072/gid00077/gid00079/gid00084/gid00083/gid00001/gid00082/gid00071/gid00064/gid00079/gid00068 (b) ARO fixated on query (c) ARO-Net PointNet Attention Module Implicit Decoder Figure 2. 2D illustration of ARO and ARO-Net architecture: (a) Input point cloud (in grey) and a set of mfixed anchors (coloured dots). (b) Radial observation Oifrom each anchor aitoward the query point xconsists of closed points inside the cone apexed at ai, with axis ri=x−ai. (c) Each radial observation Oiis passed to a PointNet encoder to obtain an embedding feature fi, which is concatenated with ri and its norm to form the query-specific ARO encoding of xwith respect to ai. Finally, all the ARO features are decoded into the occupancy value occ(x)though an attention module and several MLPs. For 3D reconstruction, the ARO features are computed for each query point, while the PointNet, attention module, and implicit decoder are fixed during inference – their weights were determined during training. (e) GT (d) ARO-Net (c) ConvONet (b) OccNet (a) Input Figure 3. A somewhat extreme toy example comparing ARO-Net to prior occupancy prediction networks on 3D reconstruction from a sparse point cloud of a cube (a), with training on a single sphere. The results from OccNet [23] and ConvONet) [25] show more signs of overfitting to the training sphere than ARO-Net (d). PointNet [27]; see Figure 2(b). The PointNet features are then aggregated through an attention module, whose output is fed to an implicit decoder to produce the occupancy value forx. We call our query-specific shape encoding Anchored Radial Observations , orARO . The prediction network is coined ARO-Net, as illustrated in Figure 2. The advantages of ARO for learning implicit fields are three-fold. First, shape inference from partial and local observations is not bound by barriers set by object categories or structural variations. Indeed, local shape features are more prevalent, and hence more generalizable, across categories than global ones [8, 23, 24]. Second, ARO is query-specific and the directional and distance information it includes is intimately tied to the occupancy prediction at the query point, as explained in Section 3.1. Last but not least, by aggregating observations from all the anchors, the resulting encoding is not purely local, like the voxel-level encoding or latent codes designed to better capture geometric details across large-scale scenes [3,15,25]. ARO effectively captures more global and contextual query-specific shape features. In Figure 3, we demonstrate using a toy example the difference between ARO-Net and representative neural im- plicit models based on global (OccNet [23]) and local gridshape encodings (convolutional occupancy network or Con- vONet [25]), for the 3D reconstruction task. All three net- works were trained on a single sphere shape, with a single anchor for ARO inside the sphere. When tested on recon- structing a cube, the results show that both OccNet and ConvONet exhibit more memorization of the training sphere either globally or locally, while the ARO-Net reconstruction is more faithful to the input without special fine-tuning. In our current implementation of ARO-Net, we adopt Fi- bonacci sampling [19] to obtain the fixed set of anchors. We demonstrate the quality and generalizability of our method on surface reconstruction from sparse point clouds, with testing on novel and unseen object categories, as well as “one-shape” training (see bottom of Figure 1). We report extensive quantitative and qualitative comparison results to state-of-the-art methods including both neural models for re- construction [8, 11, 23, 25, 26] and tessellation [7], as well as classical schemes such as screen Poisson reconstruction [18]. Finally, we conduct ablation studies to evaluate our choices for the number of anchors, the selection strategies, and the decoder architecture: MLP vs. attention modules.
Wen_Crowd3D_Towards_Hundreds_of_People_Reconstruction_From_a_Single_Image_CVPR_2023	Abstract Image-based multi-person reconstruction in wide-field large scenes is critical for crowd analysis and security alert. However, existing methods cannot deal with large scenes containing hundreds of people, which encounter the chal- lenges of large number of people, large variations in hu- man scale, and complex spatial distribution. In this paper, we propose Crowd3D, the first framework to reconstruct the 3D poses, shapes and locations of hundreds of people with global consistency from a single large-scene image. The core of our approach is to convert the problem of complex crowd localization into pixel localization with the help of our newly defined concept, Human-scene Virtual Interac- tion Point (HVIP). To reconstruct the crowd with global consistency, we propose a progressive reconstruction net- work based on HVIP by pre-estimating a scene-level cam- era and a ground plane. To deal with a large number of per- sons and various human sizes, we also design an adaptive human-centric cropping scheme. Besides, we contribute a benchmark dataset, LargeCrowd, for crowd reconstruc- tion in a large scene. Experimental results demonstrate the effectiveness of the proposed method. The code and the †Equal contribution. * Corresponding author.dataset are available at http://cic.tju.edu.cn/ faculty/likun/projects/Crowd3D .	1. Introduction 3D pose, shape and location reconstruction for hundreds of people in a large scene will help with modeling crowd be- havior for simulation and security monitoring. However, no existing methods can achieve this with global consistency. In this paper, we aim to reconstruct the 3D poses, shapes and locations of hundreds of people in the global camera space from a single large-scene image, as shown in Fig. 1. Although monocular human pose and shape estimation [15, 37, 43] has been extensively explored over the past years, estimating global space locations together with hu- man poses and shapes for multiple people from a single im- age is still a difficult problem due to the depth ambiguity. Existing methods [13,35] reconstruct 3D poses, shapes and relative positions of the reconstructed human meshes by as- suming a constant focal length. But the methods are limited to small scenes with a common FoV (Field of View). These methods cannot regress the people from a whole large-scene image [41] due to the relatively small and varying human scales in comparison to the image size. Even with an image cropping strategy, these methods cannot obtain consistent This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8937 reconstructions in the global camera space due to indepen- dent inference from the cropped images. Besides, existing methods hardly consider the coherence of the reconstructed people with the outdoor scene, especially with the ground, since the ground is a common and significant element of outdoor scenes. Taking a usual urban scene as an example, these methods may include wrong positions and rotations so that the reconstructed people do not appear to be standing or walking on the ground. In general, there are three challenges in reconstructing hundreds of people with global consistency from a single large-scene image: 1) there are a large number of people with relatively small and highly varying 2D scales; 2) due to the depth ambiguity from a single view, it is difficult to directly estimate absolute 3D positions and 3D poses of people in the large scene; 3) there is no large-scene image datasets with hundreds of people for supervising crowd re- construction in large scenes. In this paper, to address these challenges, we propose Crowd3D , the first framework for crowd reconstruction from a single large-scene image. To deal with the large number of people and various human scales, we propose an adaptive human-centric cropping scheme for a consistent scale proportion of people among different cropped images by leveraging the observation of pyramid-like changes in the scales of people in large-scene images. To ensure the globally consistent spatial locations and coherence with the scene, we propose a progressive ground-guided reconstruc- tion network Crowd3DNet to reconstruct globally consis- tent human body meshes from the cropped images by pre- estimating a global scene-level camera and a ground plane. To alleviate the ambiguity brought in by directly estimat- ing absolute 3D locations from a single image, we present a novel concept called Human-scene Virtual Interaction Point (HVIP) for effectively converting the 3D crowd spatial lo- calization problem into a progressive 2D pixel localization problem with intermediate supervisions. Benefiting from HVIP, our model can reconstruct the people with various poses including non-standing. We also construct LargeCrowd , a benchmark dataset with over 100K labeled humans (2D bounding boxes, 2D keypoints, 3D ground plane and HVIPs) in 733 gigapixel images (19200×6480) of 9 different scenes. To our best knowledge, this is the first large-scene crowd dataset, which enables the training and evaluation on large-scene images with hundreds of people. Experimental results demonstrate that our method achieves globally consistent crowd recon- struction in a large scene. Fig. 1 gives an example. To summarize, our main contributions include: 1) We propose Crowd3D, a multi-person 3D pose, shape and location estimation framework for large-scale scenes with hundreds of people. We design an adap- tive human-centric cropping scheme and a joint localand global strategy to achieve the globally consistent reconstruction. 2) We propose a progressive reconstruction network with the newly defined HVIP, to alleviate the depth ambigu- ity and obtain global reconstructions in harmony with the scene. 3) We contribute LargeCrowd , a benchmark dataset with over 100K labeled crowded people in 733 gigapixel large-scene images (19200×6480), which are valuable for the training and evaluation of crowd reconstruction and spatial reasoning in large scenes.
Xie_Active_Finetuning_Exploiting_Annotation_Budget_in_the_Pretraining-Finetuning_Paradigm_CVPR_2023	Abstract Given the large-scale data and the high annotation cost, pretraining-ﬁnetuning becomes a popular paradigm in mul- tiple computer vision tasks. Previous research has coveredboth the unsupervised pretraining and supervised ﬁnetuningin this paradigm, while little attention is paid to exploiting the annotation budget for ﬁnetuning. To ﬁll in this gap, weformally deﬁne this new active ﬁnetuning task focusing on the selection of samples for annotation in the pretraining-ﬁnetuning paradigm. We propose a novel method called Ac-tiveFT for active ﬁnetuning task to select a subset of datadistributing similarly with the entire unlabeled pool and maintaining enough diversity by optimizing a parametricmodel in the continuous space. We prove that the Earth Mover’s distance between the distributions of the selectedsubset and the entire data pool is also reduced in this pro-cess. Extensive experiments show the leading performanceand high efﬁciency of ActiveFT superior to baselines onboth image classiﬁcation and semantic segmentation. Our code is released at https://github.com/yichen928/ActiveFT .	1. Introduction Recent success of deep learning heavily relies on abun- dant training data. However, the annotation of large-scaledatasets often requires intensive human labor. This dilemmainspires a popular pretraining-ﬁnetuning paradigm where models are pretrained on a large amount of data in an unsu- pervised manner and ﬁnetuned on a small labeled subset. Existing literature pays signiﬁcant attention to both the unsupervised pretraining [ 7,13,14,18] and supervised ﬁne- tuning [ 26]. In spite of their notable contributions, these researches build upon an unrealistic assumption that we al- ready know which samples should be labeled . As shown in Fig. 1, given a large unlabeled data pool, it is necessary to pick up the most useful samples to exploit the limited an- notation budget. In most cases, this selected subset only Unlabeled Pool EncoderStage 1: Unsupervised Pretraining Data Selection Selected Samples Labeled Pool Oracle giraffe horse Our Focus: Active Finetuning TaskEncoderTask ModuleStage 2: Supervised Finetuning Finetune Figure 1. Pretraining-Finetuning Paradigm: We focus on the selection strategy of a small subset from a large unlabeled data pool for annotation, named as active ﬁnetuning task, which isunder-explored for a long time. counts a small portion ( e.g.<10%) of this large unlabeled pool. Despite the long-standing under-exploration, the se-lection strategy is still crucial since it may signiﬁcantly af- fect the ﬁnal results. Active learning algorithms [ 34,38,39] seem to be a potential solution, which aims to select the most suit- able samples for annotation when models are trained from scratch. However, their failures in this pretraining-ﬁnetuning paradigm are revealed in both [ 4] and our ex- periments (Sec. 4.1). A possible explanation comes from the batch-selection strategy of most current active learn- ing methods. Starting from a random initial set, this strat-egy repeats the model training and data selection processes multiple times until the annotation budget runs out. De- spite their success in from-scratch training, it does not ﬁt this pretraining-ﬁnetuning paradigm well due to the typi- cally low annotation budget, where too few samples in each batch lead to harmful bias inside the selection process. To ﬁll in this gap in the pretraining-ﬁnetuning paradigm, we formulate a new task called active ﬁnetuning , concen- trating on the sample selection for supervised ﬁnetuning. In this paper, a novel method, ActiveFT , is proposed to deal with this task. Starting from purely unlabeled data, Ac- tiveFT fetches a proper data subset for supervised ﬁnetuning in a negligible time. Without any redundant heuristics, we This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23715 directly bring close the distributions between the selected subset and the entire unlabeled pool while ensuring the di-versity of the selected subset. This goal is achieved by con-tinuous optimization in the high-dimensional feature space,which is mapped with the pretrained model. We design a parametric model p θSto estimate the distri- bution of the selected subset. Its parameter θSis exactly the high-dimensional features of those selected samples.We optimize this model via gradient descent by minimizingour designed loss function. Unlike traditional active learn-ing algorithms, our method can select all the samples fromscratch in a single-pass without iterative batch-selections. We also mathematically show that the optimization in the continuous space can exactly reduce the earth mover’s dis-tance (EMD) [ 35,36] between the entire pool and selected subset in the discrete data sample space. Extensive experiments are conducted to evaluate our method in the pretraining-ﬁnetuning paradigm. After pre- training the model on ImageNet-1k [ 37], we select subsets of data from CIFAR-10, CIFAR-100 [ 23], and ImageNet- 1k [ 37] for image classiﬁcation, as well as ADE20k [ 50] for semantic segmentation. Results show the signiﬁcant perfor- mance gain of our ActiveFT in comparison with baselines. Our contributions are summarized as follows: • To our best knowledge, we are the ﬁrst to iden- tify the gap of data selection for annotation and supervised ﬁnetuning in the pretraining-ﬁnetuning paradigm, which can cause inefﬁcient use of annota-tion budgets as also veriﬁed in our empirical study. Meanwhile, we formulate a new task called active ﬁne- tuning to ﬁll in this gap. • We propose a novel method, ActiveFT, to deal with the active ﬁnetuning task through parametric model opti- mization which theoretically reduces the earth mover’s distance (EMD) between the distributions of the se- lected subset and entire unlabeled pool. To our bestknowledge, we are the ﬁrst to directly optimize sam- ples to be selected in the continuous space for data se-lection tasks. • We apply ActiveFT to popular public datasets, achiev- ing leading performance on both classiﬁcation and seg- mentation tasks. In particular, our ablation study re- sults justify the design of our method to ﬁll in the dataselection gap in the pretraining-ﬁnetuning paradigm.The source code will be made public available.
Wang_Cooperation_or_Competition_Avoiding_Player_Domination_for_Multi-Target_Robustness_via_CVPR_2023	Abstract Despite incredible advances, deep learning has been shown to be susceptible to adversarial attacks. Numerous approaches have been proposed to train robust networks both empirically and certiﬁably. However, most of them de- fend against only a single type of attack, while recent work takes steps forward in defending against multiple attacks. In this paper, to understand multi-target robustness, we view this problem as a bargaining game in which different players (adversaries) negotiate to reach an agreement on a joint direction of parameter updating. We identify a phenomenon named player domination in the bargaining game, namely that the existing max-based approaches, such as MAX and MSD, do not converge. Based on our theoretical analysis, we design a novel framework that adjusts the budgets of differ- ent adversaries to avoid any player dominance. Experiments on standard benchmarks show that employing the proposed framework to the existing approaches signiﬁcantly advances multi-target robustness.	1. Introduction Machine learning (ML) models [ 15,47,48] have been shown to be susceptible to adversarial examples [ 39], where human-imperceptible perturbations added to a clean example might arbitrarily change the output of machine learning mod- els. Adversarial examples are generated by maximizing the loss within a small perturbation region around a clean exam- ple,e.g.,`1,`1and`2balls. On the other hand, numerous heuristic defenses have been proposed to be robust against *Corresponding author.Figure 1. Robust accuracy against PGD attacks and AutoAttack (“AA” in this ﬁgure) on CIFAR-10. “All” means that the model suc- cessfully defends against the `1,`2, and `1(PGD or AutoAttack) attacks simultaneously. Compared with the previously best-known methods, our proposed framework achieves improved performance. “w.`1” and “w. `2” refer to the model training with our proposed AdaptiveBudget algorithm with `1or`2norms, respectively. adversarial examples, e.g., distillation [ 31], logit-pairing [ 19] and adversarial training [ 25]. However, most of the existing defenses are only robust against one type of attacks [ 11,25,33,49], while they fail to defend against other adversaries. For example, existing work [ 18,26] showed that robustness in the `pthreat model does not necessarily generalize to other `qthreat models when p6=q. However, for the sake of the safety of ML systems, it has been argued that one should target robustness against multiple adversaries simultaneously [ 7]. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20564 Recently, various methods [ 26,35,41] have been proposed to address this problem. Multi-target adversarial training, which targets defending against multiple adversarial per- turbations, has attracted signiﬁcant attention: a variational autoencoder-based model [ 35] learns a classiﬁer robust to multiple perturbations; after that, MAX and A VG strategies, which aggregate different adversaries for adversarial training against multiple threat models, have been shown to enjoy improved performance [ 41]. To further advance the robust- ness against multiple adversaries, MSD [ 26] is proposed and outperformed MAX and A VG by taking the worst case over all steepest descent directions. These methods follow a general scheme similar to the (single-target) adversarial training. They ﬁrst sample adversarial examples by different adversaries and then update the model with the aggregation of the gradients from these adversarial examples. This general scheme for multi-target adversarial training can be seen as an implementation of a cooperative bargaining game [ 40]. In this game, different parties have to decide how to maximize the surplus they jointly get. In the multi-target adversarial training, we view each party as an adversary, and they negotiate to reach an agreed gradient direction that maximizes the overall robustness. Inspired by the bargaining game modelling for multi- target adversarial training, we ﬁrst analyze the convergence property of existing methods, i.e., MAX [ 41], MSD [ 26], and A VG [ 41], and identify a phenomenon namely player domination. Speciﬁcally, it refers to the case where one player dominates the bargaining game at any time t, and the gradient at any time tis the same as this player’s gradi- ent. Furthermore, we notice that under the SVM and linear model setups, player domination always occurs when using MAX and MSD, which leads to non-convergence. Based on such theoretical results, we propose a novel mechanism that adaptively adjusts the budgets of adversaries to avoid the player domination. We show that with our proposed mecha- nism, the overall robust accuracy of MAX, A VG and MSD improves on three representative datasets. We also illustrate the performance improvement on CIFAR- 10in Figure 1. In this paper, we present the ﬁrst theoretical analysis of the convergence of multi-target robustness on three algo- rithms under two models. Building on our theoretical results, we introduce a new method called AdaptiveBudget , de- signed to prevent the player domination phenomenon that can cause MSD and MAX to fail to converge. Our exten- sive experimental results demonstrate the superiority of our approach over previous methods.
Wang_AttriCLIP_A_Non-Incremental_Learner_for_Incremental_Knowledge_Learning_CVPR_2023	Abstract Continual learning aims to enable a model to incre- mentally learn knowledge from sequentially arrived data. Previous works adopt the conventional classification architecture, which consists of a feature extractor and a classifier. The feature extractor is shared across sequen- tially arrived tasks or classes, but one specific group of weights of the classifier corresponding to one new class should be incrementally expanded. Consequently, the parameters of a continual learner gradually increase. Moreover, as the classifier contains all historical arrived classes, a certain size of the memory is usually required to store rehearsal data to mitigate classifier bias and catastrophic forgetting. In this paper, we propose a non- incremental learner, named AttriCLIP , to incrementally extract knowledge of new classes or tasks. Specifically, AttriCLIP is built upon the pre-trained visual-language model CLIP . Its image encoder and text encoder are fixed to extract features from both images and text. Text consists of a category name and a fixed number of learnable parameters which are selected from our designed attribute word bank and serve as attributes. As we compute the visual and textual similarity for classification, AttriCLIP is a non-incremental learner. The attribute prompts, which encode the common knowledge useful for classification, can effectively mitigate the catastrophic forgetting and avoid constructing a replay memory. We evaluate our AttriCLIP and compare it with CLIP-based and previous state-of-the- art continual learning methods in realistic settings with domain-shift and long-sequence learning. The results show that our method performs favorably against previous state- of-the-arts. The implementation code will be available at https://gitee.com/mindspore/models/tree/master/research/ cv/AttriCLIP.	1. Introduction In recent years, deep neural networks have achieved re- markable progress in classification when all the classes (or *Co-First Author. †Corresponding Author.tasks) are jointly trained. However, in real scenarios, the tasks or classes usually sequentially arrive. Continual learn- ing [13, 21, 28] aims to train a model which incrementally expands its knowledge so as to deal with all the histori- cal tasks or classes, behaving as if those tasks or classes are jointly trained. The conventional continual learning methods learn sequentially arrived tasks or classes with a shared model, as shown in Fig. 1(a). Such processing that fine-tunes the same model in sequence inevitably results in subsequent values of the parameters overwriting previous ones [20], which leads to catastrophic forgetting. Besides, the classification ability on historical data can be easily de- stroyed by current-stage learning. In the conventional con- tinual learning methods, a classifier on top of the feature extractor is employed to perform recognition. As one group of weights in the classifier is responsible for the prediction of one specific class, the classifier needs to be expanded sequentially to make a continual learner able to recognize novel classes. Moreover, extra replay data is usually re- quired to reduce the classifier bias and the catastrophic for- getting of learned features. It is still challenging if we ex- pect a non-incremental learner ,i.e., the trainable parame- ters of the model do not incrementally increase and no re- play data is needed to avoid the classifier bias and the catas- trophic forgetting. To address the above issues, this paper proposes a con- tinual learning method named AttriCLIP , which adopts the frozen encoders of CLIP [19]. It is a typical visual-language model that conducts image classification by contrasting the features of images and their descriptive texts. In the face of increasing instances, we design a prompt tuning scheme for continual learning. As shown in Fig. 1(b), there are simi- lar attributes in images with different categories, such as “a brown-white dog lying on the grass” and “a brown-white cat lying on the grass”. They belong to different categories but both have “lying on the grass” and “brown-white” at- tributes, so the distance between these two images of differ- ent categories may be close in the feature space. Therefore, we selectively train different prompts based on the attributes of the images rather than the categories. In this way, there is This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3654 🔥Trainable parametersTask 1Task 2…Task 𝑇Encoder 🔥Classifier 🔥 Fixed number of classesUnlimited number of classes [CLS]on the grassbrown-whitecatdogbirdon the grass brown-whitebeakon the grass （a）（b）Contrastive learningMemorydata ……Task 1Task 2Task Tbrown-whiteon the grassbeak 🔥 🔥Attribute Word Bank…Figure 1. (a) Traditional framework for continual learning. The encoder and the classifier are trained by tasks in sequence, some of which even need extra memory data. In the framework, the model parameters of the current task are fine-tuned from the parameters trained by the previous last task and then are used for the classification of all seen tasks. The total number of categories the model can classify is fixed in the classifier. (b) Our proposed AttriCLIP for continual learning. AttriCLIP is based on CLIP, which classifies images by contrasting them with their descriptive texts. The trainable prompts are selected by the attributes of the current image from a prompt pool. The prompts are different if the attributes of the image are different. The trained prompts are concatenated with the class name of the image, which serve as a more accurate supervised signal for image classification than labels. no problem of knowledge overwriting caused by sequential training of the same model with increasing tasks. Specifically, an attribute word bank is constructed as shown in Fig 1(b), which consists of a set of (key, prompt) pairs. The keys represent the local features (attributes) of images and the prompts represent the descriptive words cor- responding to the keys. Several prompts are selected ac- cording to the similarities between their keys and the in- put image. The selected prompts are trained to capture an accurate textual description of the attributes of the image. If images with different labels have similar attributes, it is also possible to select the same prompts, i.e., the same prompts can be trained with images of different categories. Similarly, different prompts can be trained by the images of the same category. The trained prompts are concate- nated with the class names and put into the text encoder to contrast with the image feature from the image encoder. This process makes our AttriCLIP distinct from all previous classifier-based framework as it serves as a non-incremental learner without the need to store replay data. The goal of our method is to select the existing attributes of the current image as the text description in the inference process, to classify the image. In addition, the structure of AttriCLIP can also avoid the problem of the increasing classifier pa- rameters with the increase of the tasks and the problem of the inefficiency about memory data in the traditional con- tinual learning methods. The experimental setup of existing continual learning methods is idealized which divides one dataset into sev- eral tasks for continual learning. The model can set the output dimensionality of the classifier according to the to- tal number of categories in the dataset. In practical appli- cations, with the continuous accumulation of data, the to- tal number of categories of samples usually cannot be ob-tained when the model is established. When the total num- ber of categories exceeds the preset output dimensionality of the classifier, the model has to add parameters to classi- fier and requires the previous samples for fine-tuning, which greatly increases the training burden. Therefore, the contin- ual learning approaches are required to have the ability to adapt to the categories of freely increasing data, i.e., the model capacity should not have a category upper limit. In order to measure such ability of continual learning models, we propose a Cross-Datasets Continual Learning (CDCL) setup, which verifies the classification performance of the model on long-sequence domain-shift tasks. The contribu- tions of this paper are summarized as follows: • We establish AttriCLIP, which is a prompt tuning ap- proach for continual learning based on CLIP. We train different prompts according to the attributes of images to avoid knowledge overwriting caused by training the same model in sequence of classes. • AttriCLIP contrasts the images and their descriptive texts based on the learned attributes. This approach avoids the memory data requirement for fine-tuning the classifier of increasing size. • In order to evaluate the performance of the model on long-sequence domain-shift tasks, we propose a Cross-Datasets Continual Learning (CDCL) experi- mental setup. AttriCLIP exhibits excellent perfor- mance and training efficiency on CDCL.
Wu_Pix2map_Cross-Modal_Retrieval_for_Inferring_Street_Maps_From_Images_CVPR_2023	Abstract Self-driving vehicles rely on urban street maps for au- tonomous navigation. In this paper, we introduce Pix2Map, a method for inferring urban street map topology directly from ego-view images, as needed to continually update and expand existing maps. This is a challenging task, as we need to infer a complex urban road topology directly from raw image data. The main insight of this paper is that this problem can be posed as cross-modal retrieval by learning a joint, cross-modal embedding space for images and ex- isting maps, represented as discrete graphs that encode the topological layout of the visual surroundings. We conduct our experimental evaluation using the Argoverse dataset and show that it is indeed possible to accurately retrieve street maps corresponding to both seen and unseen roads solely from image data. Moreover, we show that our re- trieved maps can be used to update or expand existing maps and even show proof-of-concept results for visual localiza- tion and image retrieval from spatial graphs.	1. Introduction We propose Pix2Map , a method for inferring road maps directly from images. More precisely, given camera im- ages, Pix2Map generates a topological map of the visible surroundings, represented as a spatial graph. Such maps en- code both geometric and semantic scene information such as lane-level boundaries and locations of signs [52] and serve as powerful priors in virtually all autonomous vehi- cle stacks. In conjunction with on-the-fly sensory measure- ments from lidar or camera, such maps can be used for lo- calization [3] and path planning [39]. As map maintenance and expansion to novel areas are challenging and expen- sive, often requiring manual effort [38, 50], automated map maintenance and expansion has been gaining interest in the community [9, 10, 27, 32, 33, 35, 38, 40]. Why is it hard? To estimate urban street maps, we need to *Work done while at Carnegie Mellon University. Global Street Map Camera Data Street Map Adjacency Matrix Figure 1. Illustration of our proposed Pix2Map for cross- modal retrieval. Given unseen 360◦ego-view images collected from seven ring cameras ( left), our Pix2Map predicts the lo- cal street map by retrieving from the existing street map library (right ), represented as an adjacency matrix. The local street maps can be further used for global high-definition map maintenance (top). learn to map continuous images from ring cameras to dis- crete graphs with varying numbers of nodes and topology in bird’s eye view (BEV). Prior works that estimate road topology from monocular images first process images using Convolutional Neural Networks or Transformers to extract road lanes and markings [9] or road centerlines [10] from images. These are used in conjunction with recurrent neural networks for the generation of polygonal structures [12] or heuristic post-processing [33] to estimate a spatial graph in BEV . This is a very difficult learning problem: such meth- ods need to jointly learn to estimate a non-linear mapping from image pixels to BEV , as well as to estimate the road layout and learn to generate a discrete spatial graph. Pix2Map. Instead, our core insight is to simply sidestep the problem of graph generation and 3D localization from monocular images by recasting Pix2Map as a cross-modal retrieval task: given a set of test-time ego-view images, we (i) compute their visual embedding and then (ii) retrieve a graph with the closest graph embedding in terms of co- sine similarity. Given recent multi-city autonomous vehi- cle datasets [13], it is straightforward to construct pairs of ego-view images and street maps, both for training and test- 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17514 ing. We train image and graph encoders to operate in the same embedding space, making use of recent techniques for cross-modal contrastive learning [44]. Our key techni- cal contribution is a novel but simple graph encoder , based on sequential transformers from the language community (i.e., BERT [18]) that extract fixed-dimensional embeddings from street maps of arbitrary size and topology. In fact, we find that even naive nearest-neighbor retrieval performs comparably to leading techniques for map gener- ation [9, 10], i.e., returning the graph paired with the best- matching image in the training set. Moreoever, we demon- strate that cross-modal retrieval via Pix2Map performs even better due to its ability to learn graph embeddings that reg- ularize the output space of graphs. In addition, cross-modal retrieval has the added benefit of allowing one to expand the retrieval graph library with unpaired graphs that lack camera data, leveraging the insight that retrieval need not be limited to the same (image, graph) training pairs used for learning the encoders. This suggests that Pix2Map can be further improved with augmented road graph topologies that capture potential road graph updates (for which paired visual data might not yet exist). Beyond mapping, we show pilot experiments for visual localization, and the inverse method, Map2Pix , which retrieves a close-matching image from an image library given a graph. While not our pri- mary focus, such approaches may be useful for generating photorealistic simulated worlds [40]. We summarize our main contributions as follows: We (i) show that dynamic street map construction from cam- eras can be posed as a cross-modal retrieval task and pro- pose an contrastive image-graph model based on this fram- ing. Building on recent advances in multimodal represen- tation learning, we train a graph encoder and an image en- coder with a shared latent space. We (ii) demonstrate em- pirically that this approach is effective and perform ablation studies to highlight the impacts of architectural decisions. Our approach outperforms existing graph generation meth- ods from image cues by a large margin. We (iii) further show that it is possible to retrieve similar graphs to those in previously unseen areas without access to the ground truth graphs for those areas, and demonstrate the generalization ability to novel observations.
Wu_SCoDA_Domain_Adaptive_Shape_Completion_for_Real_Scans_CVPR_2023	Abstract 3D shape completion from point clouds is a challeng- ing task, especially from scans of real-world objects. Con- sidering the paucity of 3D shape ground truths for real scans, existing works mainly focus on benchmarking this task on synthetic data, e.g. 3D computer-aided design mod- els. However, the domain gap between synthetic and real data limits the generalizability of these methods. Thus, we propose a new task, SCoDA, for the domain adaptation of real scan shape completion from synthetic data. A new dataset, ScanSalon, is contributed with a bunch of elaborate 3D models created by skillful artists according to scans. To address this new task, we propose a novel cross-domain feature fusion method for knowledge transfer and a novel volume-consistent self-training framework for robust learn- ing from real data. Extensive experiments prove our method is effective to bring an improvement of 6% ∼7% mIoU.	1. Introduction Shape completion and reconstruction from scans is a practical 3D digitization task that is of great significancein applications of virtual and augmented reality. It takes a scanned point cloud as input and aims to recover the 3D shape of the target object. The completion of real scans is challenging for the poor quality of point clouds and the deficiency of 3D shape ground truths. Existing methods ex- ploit synthetic data, e.g.3D computer-aided design (CAD) models to alleviate the demand for real object shapes. For example, authors of [18, 26, 35] simulate the scanning pro- cess to obtain point clouds from CAD models with paired ground truth shapes to train learning-based reconstruction models. However, there still exist distinctions between the simulated and real scan because the latter is scanned from a real object with complex scanning noise and occlusion, which limits the generalization quality. Considering the underexploration in this field, we propose a new task, SCoDA ,Domain Adaptive Shape Completion, that aims to transfer the knowledge from the synthetic domain with rich clean shapes (source domain) into the shape completion of real scans (target domain), as illustrated in Fig. 1. To this end, we, for the first time, build an object-centric dataset, ScanSalon , which consists of real Scan s with Shape manua lannotati ons. Our ScanSalon con- tains a bunch of 3D models that are paired with real scans of objects in six categories: chair, desk/table, sofa, bed, lamp, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17630 and car. The 3D models are manually created of high qual- ity by skillful artists for around 10% of real scans, which can serve as the evaluation ground truths of shape comple- tion or as the few labels for semi-supervised domain adapta- tion. See Fig. 1 for some examples, and details of ScanSa- lonare exposed in Sec. 4. The main challenge of the proposed SCoDA task lies in the domain gap between synthetic and real point clouds. Due to the intrinsic complexity of the scanning process, e.g., the scanner parameters and object materials, it is difficult to simulate the scans in terms of sparsity, noisy extent, etc. More importantly, real scans are usually incomplete result- ing from the scene layout and object occlusion during scan- ning, which can hardly be simulated. Thus, we propose a novel domain adaptive shape completion approach to trans- fer the rich knowledge from the synthetic domain to the real one. At first, the reconstruction module in our approach is based on an implicit function for continuous shape com- pletion [18, 54, 58]. For an effective transfer, we observe that although the local patterns of real scans ( e.g., noise, sparsity, and incompleteness) are distinct from the simu- lated synthetic ones, the global topology or structure in a same category is usually similar between the synthetic and real data, for example, a chair from either the synthetic or real domain usually consists of a seat, a backrest, legs, etc. (see Fig. 1). In other words, the global topology is more likely to be domain invariant, while the local patterns are more likely to be domain specific. Accordingly, we propose a cross-domain feature fusion (CDFF) module to combine the global features and local features learned in the syn- thetic and real domain, respectively, which helps recover both fine details and global structures in the implicit recon- struction stage. Moreover, a novel volume-consistent self- training (VCST) framework is developed to encourage self- supervised learning from the target data. Specifically, we create two views of real scans by dropping different clus- ters of points to produce incompleteness of different extent, and the model is forced to make consistent implicit predic- tions at each spatial volume, which encourages the model’s robustness to the incompleteness of real scans. To construct a benchmark on the proposed new task, we implement some existing solutions to related tasks as base- lines, and develop extensive experiments for the baselines and our method on ScanSalon . These experiments also demonstrate the effectiveness of the proposed method. In summary, our key contributions are four-fold: • We propose a new task, SCoDA , namely domain adaptive shape completion for real scans; A dataset ScanSalon that contains 800 elaborate 3D models paired with real scans in 6 categories is contributed. • A novel cross-domain feature fusion module is de- signed to combine the knowledge of global shapes and local patterns learned in the synthetic and real domain,respectively. Such a feature fusion manner may also inspire the works in the 2D domain adaption field. • A volume-consistent self-training framework is pro- posed to improve the robustness of shape completion to the complex incompleteness of real scans. • A benchmark with multiple methods evaluated is con- structed for the task of SCoDA based on ScanSalon ; Extensive experiments also demonstrate the superior- ity of the proposed method.
Xu_CXTrack_Improving_3D_Point_Cloud_Tracking_With_Contextual_Information_CVPR_2023	Abstract 3D single object tracking plays an essential role in many applications, such as autonomous driving. It remains a challenging problem due to the large appearance varia- tion and the sparsity of points caused by occlusion and lim- ited sensor capabilities. Therefore, contextual information across two consecutive frames is crucial for effective ob- ject tracking. However, points containing such useful in- formation are often overlooked and cropped out in existing methods, leading to insufficient use of important contextual knowledge. To address this issue, we propose CXTrack, a novel transformer-based network for 3D object tracking, which exploits ConteXtual information to improve the track- ing results. Specifically, we design a target-centric trans- former network that directly takes point features from two consecutive frames and the previous bounding box as in- put to explore contextual information and implicitly propa- gate target cues. To achieve accurate localization for ob- jects of all sizes, we propose a transformer-based local- ization head with a novel center embedding module to dis- tinguish the target from distractors. Extensive experiments on three large-scale datasets, KITTI, nuScenes and Waymo Open Dataset, show that CXTrack achieves state-of-the-art tracking performance while running at 34 FPS.	1. Introduction Single Object Tracking (SOT) has been a fundamental task in computer vision for decades, aiming to keep track of a specific target across a video sequence, given only its initial status. In recent years, with the development of 3D data acquisition devices, it has drawn increasing attention for using point clouds to solve various vision tasks such as object detection [7, 12, 14, 15, 18] and object tracking [20, 29, 31–33]. In particular, much progress has been made on point cloud-based object tracking for its huge potential in applications such as autonomous driving [11,30]. However, it remains challenging due to the large appearance variation *corresponding author 𝑃!"#𝑃!CXTrackParadigm (Ours) 𝐵!"#BackboneBackboneRelation ModelingLocalization𝐵!𝑃!"#+𝐵!"#𝑃! crop sample & cropBackboneBackboneFeature SimilaritySC3D Paradigm𝐵!argmax 𝑃!"#+𝐵!"#𝑃! cropRelation ModelingLocalizationBackboneBackboneP2B Paradigm𝐵! 𝑃!"#+𝐵!"#𝑃! Motion-Centric Paradigm cropSegmentationMotionModeling𝐵!Figure 1. Comparison of various 3D SOT paradigms. Previous methods crop the target from the frames to specify the region of interest, which largely overlook contextual information around the target. On the contrary, our proposed CXTrack fully exploits con- textual information to improve the tracking results. of the target and the sparsity of 3D point clouds caused by occlusion and limited sensor resolution. Existing 3D point cloud-based SOT methods can be cat- egorized into three main paradigms, namely SC3D, P2B and motion-centric, as shown in Fig. 1. As a pioneering work, SC3D [6] crops the target from the previous frame, and compares the target template with a potentially large number of candidate patches generated from the current frame, which consumes much time. To address the effi- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1084 ciency problem, P2B [20] takes the cropped target tem- plate from the previous frame as well as the complete search area in the current frame as input, propagates tar- get cues into the search area and then adopts a 3D re- gion proposal network [18] to predict the current bound- ing box. P2B reaches a balance between performance and speed. Therefore many follow-up works adopt the same paradigm [3, 8, 9, 22, 29, 31, 33]. However, both SC3D and P2D paradigms overlook the contextual information across two consecutive frames and rely entirely on the appear- ance of the target. As mentioned in previous work [32], these methods are sensitive to appearance variation caused by occlusions and tend to drift towards intra-class distrac- tors. To this end, M2-Track [32] introduces a novel motion- centric paradigm, which directly takes point clouds from two frames without cropping as input, and then segments the target points from their surroundings. After that, these points are cropped and the current bounding box is es- timated by explicitly modeling motion between the two frames. Hence, the motion-centric paradigm still works on cropped patches that lack contextual information in later lo- calization. In short, none of these methods could fully uti- lize the contextual information around the target to predict the current bounding box, which may degrade tracking per- formance due to the existence of large appearance variation and widespread distractors. To address the above concerns, we propose a novel transformer-based tracker named CXTrack for 3D SOT, which exploits contextual information across two consecu- tive frames to improve the tracking performance. As shown in Fig. 1, different from paradigms commonly adopted by previous methods, CXTrack directly takes point clouds from the two consecutive frames as input, specifies the tar- get of interest with the previous bounding box and predicts the current bounding box without any cropping, largely pre- serving contextual information. We first embed local geo- metric information of the two point clouds into point fea- tures using a shared backbone network. Then we integrate the targetness information into the point features accord- ing to the previous bounding box and adopt a target-centric transformer to propagate the target cues into the current frame while exploring contextual information in the sur- roundings of the target. After that, the enhanced point fea- tures are fed into a novel localization head named X-RPN to obtain the final target proposals. Specifically, X-RPN adopts a local transformer [25] to model point feature in- teractions within the target, which achieves a better balance between handling small and large objects compared with other localization heads. To distinguish the target from dis- tractors, we incorporate a novel center embedding module into X-RPN, which embeds the relative target motion be- tween two frames for explicit motion modeling. Extensive experiments on three popular tracking datasets demonstratethat CXTrack significantly outperforms the current state-of- the-art methods by a large margin while running at real-time (34 FPS) on a single NVIDIA RTX3090 GPU. In short, our contributions can be summarized as: (1) a new paradigm for the real-time 3D SOT task, which fully exploits contextual information across consecutive frames to improve the tracking accuracy; (2) CXTrack: a transformer-based tracker that employs a target-centric transformer architecture to propagate targetness informa- tion and exploit contextual information; and (3) X-RPN: a localization head that is robust to intra-class distractors and achieves a good balance between small and large targets.
Wang_Balancing_Logit_Variation_for_Long-Tailed_Semantic_Segmentation_CVPR_2023	Abstract Semantic segmentation usually suffers from a long-tail data distribution. Due to the imbalanced number of samples across categories, the features of those tail classes may get squeezed into a narrow area in the feature space. Towards a balanced feature distribution, we introduce category-wise variation into the network predictions in the training phase such that an instance is no longer projected to a feature point, but a small region instead. Such a perturbation is highly dependent on the category scale, which appears as assigning smaller variation to head classes and larger variation to tail classes. In this way, we manage to close the gap between the feature areas of different categories, resulting in a more balanced representation. It is note- worthy that the introduced variation is discarded at the inference stage to facilitate a confident prediction. Although with an embarrassingly simple implementation, our method manifests itself in strong generalizability to various datasets and task settings. Extensive experiments suggest that our plug-in design lends itself well to a range of state-of-the-art approaches and boosts the performance on top of them.1	1. Introduction The success of deep models in semantic segmenta- tion [12, 58, 71, 109] benefits from large-scale datasets. However, popular datasets for segmentation, such as PAS- CAL VOC [24] and Cityscapes [18], usually follow a long-tail distribution, where some categories may have far fewer samples than others. Considering the particularity of this task, which targets assigning labels to pixels instead of images, it is quite difficult to balance the distribution 1Code: https://github.com/grantword8/BLV . †This work was done during the internship at SenseTime Research. ‡Rui Zhao is also with Qing Yuan Research Institute, Shanghai Jiao Tong University. Category 1 Decision boundary (a) w/oLogit Variation (b) w/ Balanced Logit VariationCategory 2 Category 3 Category 4Figure 1. Illustration of logit variation from the feature space, where each point corresponds to an instance and different colors stand for different categories. (a) Without logit variation, the features of tail classes ( e.g., the blue one) may get squeezed into a narrow area. (b) After introducing logit variation, which is controlled by the category scale ( i.e., number of training samples belonging to a particular category), we expand each feature point to a feature region with random perturbation, resulting in a more category-balanced feature distribution. from the aspect of data collection. Taking the scenario of autonomous driving as an example, a bike is typically tied to a smaller image region ( i.e., fewer pixels) than a car, and trains appear more rarely than pedestrians in a city. Therefore, learning a decent model from long tail data distributions becomes critical. A common practice to address such a challenge is to make better use of the limited samples from tail classes. For this purpose, previous attempts either balance the sample quantity ( e.g., oversample the tail classes when organizing the training batch) [8, 28, 36, 46, 75], or balance the per- sample importance ( e.g., assign the training penalties re- garding tail classes with higher loss weights) [19,51,67,69]. Given existing advanced techniques, however, performance degradation can still be observed in those tail categories. This works provides a new perspective on improving long-tailed semantic segmentation. Recall that, in modern pipelines based on neural networks [12,14,58,92,108,110], instances are projected to representative features before This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19561 categorized to a certain class. We argue that the features of tail classes may get squeezed into a narrow area in the feature space, as the blue region shown in Fig. 1a, because of miserly samples. To balance the feature distribution, we propose a simple yet effective approach via introducing bal- ancing logit variation (BLV) into the network predictions. Concretely, we perturb each predicted logit with a randomly sampled noise during training. That way, each instance can be seen as projected to a feature region, as shown in Fig. 1b, whose radius is dependent on the noise variance. We then propose to balance the variation by applying smaller variance to head classes and larger variance to tail classes so as to close the feature area gap between different categories. This newly introduced variation can be viewed as a special augmentation and discarded in the inference phase to ensure a reliable prediction. We evaluate our approach on three different settings of semantic segmentation, including fully supervised [21, 90, 101, 109], semi-supervised [13, 87, 111, 115], and unsu- pervised domain adaptation [23, 52, 107, 113], where we improve the baselines consistently. We further show that our method works well with various state-of-the-art frame- works [2, 41, 42, 87, 92, 108] and boosts their performance, demonstrating its strong generalizability.
Wei_Inferring_and_Leveraging_Parts_From_Object_Shape_for_Improving_Semantic_CVPR_2023	Abstract Despite the progress in semantic image synthesis, it re- mains a challenging problem to generate photo-realistic parts from input semantic map. Integrating part segmen- tation map can undoubtedly beneﬁt image synthesis, but is bothersome and inconvenient to be provided by users. To improve part synthesis, this paper presents to inferParts from Object ShapE(iPOSE) and leverage it for improving semantic image synthesis. However, albeit several part seg- mentation datasets are available, part annotations are still not provided for many object categories in semantic image synthesis. To circumvent it, we resort to few-shot regime to learn a PartNet for predicting the object part map with the guidance of pre-deﬁned support part maps. PartNet can be readily generalized to handle a new object cate- gory when a small number (e.g., 3) of support part maps for this category are provided. Furthermore, part semantic modulation is presented to incorporate both inferred part map and semantic map for image synthesis. Experiments show that our iPOSE not only generates objects with rich part details, but also enables to control the image synthe- sis ﬂexibly. And our iPOSE performs favorably against the state-of-the-art methods in terms of quantitative and qual- itative evaluation. Our code will be publicly available at https://github.com/csyxwei/iPOSE .	1. Introduction Semantic image synthesis allows to generate an image with input semantic map, which provides signiﬁcant ﬂexi- bility for controllable image synthesis. Recently, it has at- tracted intensive attention due to its wide applications, e.g., content generation and image editing [28,36], and extensive beneﬁts for many vision tasks [33]. Albeit rapid progress has been made in semantic im- age synthesis [10, 25, 28, 30, 33, 36, 43–45], it is still chal- lenging to generate photo-realistic parts from the semantic map. Most methods [28, 33] tackle semantic image syn- thesis with a spatially adaptive normalization architecture, while other frameworks are also explored, such as Style- GAN [19] and diffusion model [43, 45], etc. However, due to the lack of ﬁne-grained guidance ( e.g., object part infor- mation), these methods only exploited the object-level se- mantic information for image synthesis, and usually failed to generate photo-realistic object parts (see the top row of Fig. 1(c)). SAFM [25] adopted the shape-aware position descriptor to exploit the pixel’s position feature inside the object. However, as illustrated in Fig. 1(a), the obtained de- scriptor tends to be only region-aware instead of part-aware, leading to limited improvement in synthesized results (see the middle row of Fig. 1(c)). To improve image parts synthesis, one straightforward solution is to ask the user to provide part segmentation map and integrate it into semantic image synthesis dur- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11248 ing training and inference. However, it is bothersome and inconvenient for users to provide it, especially during in- ference. Fortunately, with the existing part segmentation datasets [7], we propose a method to inferParts from Object ShapE(iPOSE), and leverage it for improving semantic im- age synthesis. Speciﬁcally, based on these datasets, we ﬁrst construct an object part dataset that consists of paired (ob- ject shape, object part map) to train a part prediction net- work (PartNet). Besides, although the part dataset con- tains part annotations for several common object categories, many object categories in semantic image synthesis are still not covered. To address this issue, we introduce the few shot mechanism to our proposed PartNet. As shown in Fig. 1(b), for each category, a few annotated object part maps are se- lected as pre-deﬁned supports to guide the part prediction. Cross attention block is adopted to aggregate the part infor- mation from support part maps. Beneﬁted from the few shot setting, our PartNet can be readily generalized to handle a new object category not in the object part dataset. Partic- ularly, we can manually label kobject part maps for this category ( e.g., 3) as supports, and use them to infer the part map without ﬁne-tuning the part prediction model. By processing each object in the semantic map, we ob- tain the part map. With that, we further present a part se- mantic modulation (PSM) residual block to incorporate the part map with semantic map to improve the image syn- thesis. Speciﬁcally, the part map is ﬁrst used to modulate the normalized activations spatially to inject the part struc- ture information. Then, to inject the semantic texture in- formation, the semantic map and a randomly sampled 3D noise are further used to modulate the features with the SPADE module [28]. We ﬁnd that performing part modula- tion and semantic modulation sequentially disentangles the structure and texture for image synthesis, and is also beneﬁ- cial for generating images with realistic parts. Additionally, to facilitate model training, a global adversarial loss and an object-level CLIP style loss [54] are further introduced to encourage model to generate photo-realistic images. Experiments show that our iPOSE can generate more photo-realistic parts from the given semantic map, while having the ﬂexibility to control the generated objects (as illustrated in Fig. 1 (c)(d)). Both quantitative and quali- tative evaluations on different datasets further demonstrate that our method performs favorably against the state-of-the- art methods. The contributions of this work can be summarized as: • We propose a method iPOSE to infer parts from object shape and leverage them to improve semantic image synthesis. Particularly, a PartNet is proposed to predict the part map based on a few support part maps, which can be easily generalized to new object categories. • A part semantic modulation Resblock is presented to incorporate the predicted part map and semanticmap for image synthesis. And global adversarial and object-level CLIP style losses are further introduced to generate photo-realistic images. • Experimental results show that our iPOSE performs fa- vorably against state-of-the-art methods and can gen- erate more photo-realistic results with rich part details.
Yang_Panoptic_Video_Scene_Graph_Generation_CVPR_2023	Abstract Towards building comprehensive real-world visual per- ception systems, we propose and study a new problem called panoptic scene graph generation (PVSG). PVSG is related to the existing video scene graph generation (VidSGG) problem, which focuses on temporal interactions between humans and objects localized with bounding boxes in videos. However, the limitation of bounding boxes in detecting non-rigid objects and backgrounds often causes VidSGG systems to miss key details that are crucial for comprehensive video understanding. In contrast, PVSG re- quires nodes in scene graphs to be grounded by more pre- cise, pixel-level segmentation masks, which facilitate holis- tic scene understanding. To advance research in this new area, we contribute a high-quality PVSG dataset, which consists of 400 videos (289 third-person + 111 egocentric videos) with totally 150K frames labeled with panoptic seg- mentation masks as well as fine, temporal scene graphs. Wealso provide a variety of baseline methods and share useful design practices for future work.	1. Introduction In the last several years, scene graph generation has re- ceived increasing attention from the computer vision com- munity [15, 16, 24, 48–51]. Compared with object-centric labels like “person” or “bike,” or precise bounding boxes commonly seen in object detection, scene graphs provide far richer information in images, such as “a person riding a bike,” which capture both objects and the pairwise rela- tionships and/or interactions. A recent trend in the scene graph community is the shift from static, image-based scene graphs to temporal, video-level scene graphs [1, 41, 49]. This has marked an important step towards building more comprehensive visual perception systems. Compared with individual images, videos clearly contain more information due to the additional temporal dimension, 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18675 Table 1. Comparison between the PVSG dataset and some related datasets . Specifically, we choose three video scene graph generation (VidSGG) datasets, three video panoptic segmentation (VPS) datasets, and two egocentric video datasets—one for short-term action antic- ipation (STA) while the other for video object segmentation (VOS). Our PVSG dataset is the first long-video dataset with rich annotations of panoptic segmentation masks and temporal scene graphs. Dataset Task #Video Video Hours Avg. Len. View #ObjCls #RelCls Annotation # Seg Frame Year Source ImageNet-VidVRD [35] VidSGG 1,000 - - 3rd 35 132 Bounding Box - 2017 ILVSRC2016-VID [33] Action Genome [15] VidSGG 10,000 99 35s 3rd 80 50 Bounding Box - 2019 YFCC100M [42] VidOR [34] VidSGG 10,000 82 30s 3rd 35 25 Bounding Box - 2020 Charades [36] Cityscapes-VPS [17] VPS 500 - - vehicle 19 - Panoptic Seg. 3K 2020 - KITTI-STEP [45] VPS 50 - - vehicle 19 - Panoptic Seg. 18K 2021 - VIP-Seg [28] VPS 3,536 5 5s 3rd 124 - Panoptic Seg. 85K 2022 - Ego4D-STA [12] STA 1,498 111 264s ego - - Bounding Box - 2022 - VISOR [8] VOS 179 36 720s ego 257 2 Semantic Seg. 51K 2022 EPIC-KITCHENS [7] PVSG PVSG 400 9 77s 3rd + ego 126 57 Panoptic Seg. 150K 2023 VidOR + Ego4D + EPIC-KITCHENS which largely facilitates high-level understanding of tempo- ral events (e.g., actions [14]) and is useful for reasoning [59] and identifying causality [10] as well. However, we ar- gue that current video scene graph representations based on bounding boxes still fall short of human visual perception due to the lack of granularity —which can be addressed with panoptic segmentation masks . This is echoed by the evo- lutionary path in visual perception research: from image- level labels (i.e., classification) to spatial locations (i.e., ob- ject detection) to more fine-grained, pixel-wise masks (i.e., panoptic segmentation [20]). In this paper, we take scene graphs to the next level by proposing panoptic video scene graph generation (PVSG) , a new problem that requires each node in video scene graphs to be grounded by a pixel-level segmentation mask. Panop- tic video scene graphs can solve a critical issue exposed in bounding box-based video scene graphs: both things and stuff classes (i.e., amorphous regions containing water, grass, etc.) can be well covered—the latter are crucial for understanding contexts but cannot be localized with bound- ing boxes. For instance, if we switch from panoptic video scene graphs to bounding box-based scene graphs for the video in Figure 1, some nontrivial relations useful for con- text understanding like “adult-1 standing on/in ground” and “adult-2 standing on/in water” will be missing. It is also worth noting that bounding box-based video scene graph annotations, at least in current research [15], often miss small but important details, such as the “candles” on cakes. To help the community progress in this new area, we contribute a high-quality PVSG dataset, which consists of 400 videos among which 289 are third-person videos and 111 are egocentric videos. Each video contains an average length of 76.5 seconds. In total, 152,958 frames are labeled with fine panoptic segmentation and temporal scene graphs. There are 126 object classes and 57 relation classes. A more detailed comparison between our PVSG dataset and some related datasets is shown in Table 1. To solve the PVSG problem, we propose a two-stage framework: the first stage produces a set of features for eachmask-based instance tracklet while the second stage gener- ates video-level scene graphs based on tracklets’ features. We study two design choices for the first stage: 1) a panop- tic segmentation model + a tracking module; 2) an end-to- end video panoptic segmentation model. For the second scene graph generation stage, we provide four different im- plementations covering both convolution and Transformer- based methods. In summary, we make the following contributions to the scene graph community: 1.A new problem : We identify several issues associated with current research in scene graph generation and propose a new problem, which combines video scene graph generation with panoptic segmentation for holis- tic video understanding. 2.A new dataset : A high-quality dataset with fine, tem- poral scene graph annotations and panoptic segmenta- tion masks is proposed to advance the area of PVSG. 3.New methods and a benchmark : We propose a two- stage framework to address the PVSG problem and benchmark a variety of design ideas, from which valu- able insights on good design practices are drawn for future work.
Yang_Language_in_a_Bottle_Language_Model_Guided_Concept_Bottlenecks_for_CVPR_2023	Abstract Concept Bottleneck Models (CBM) are inherently inter- pretable models that factor model decisions into human- readable concepts. They allow people to easily understand why a model is failing, a critical feature for high-stakes ap- plications. CBMs require manually speciﬁed concepts and often under-perform their black box counterparts, preventing their broad adoption. We address these shortcomings and are ﬁrst to show how to construct high-performance CBMs without manual speciﬁcation of similar accuracy to black box models. Our approach, Language Guided Bottlenecks (LaBo), leverages a language model, GPT-3, to deﬁne a large space of possible bottlenecks. Given a problem domain, LaBo uses GPT-3 to produce factual sentences about cate- gories to form candidate concepts. LaBo efﬁciently searches possible bottlenecks through a novel submodular utility that promotes the selection of discriminative and diverse informa- tion. Ultimately, GPT-3’s sentential concepts can be aligned to images using CLIP , to form a bottleneck layer. Experi- ments demonstrate that LaBo is a highly effective prior for concepts important to visual recognition. In the evaluation with 11 diverse datasets, LaBo bottlenecks excel at few-shot classiﬁcation: they are 11.7% more accurate than black box linear probes at 1 shot and comparable with more data. Overall, LaBo demonstrates that inherently interpretable models can be widely applied at similar, or better, perfor- mance than black box approaches.1	1. Introduction As deep learning systems improve, their applicability to critical domains is hampered because of a lack of trans- parency. Efforts to address this have largely focused on post-hoc explanations [ 47,54,72]. Such explanations can be problematic because they may be incomplete or unfaith- ful with respect to the model’s computations [ 49]. Models 1Code and data are available at https://github.com/YueYANG1996/LaBo Input Image x !" Human Designed Concepts ... has nape color :: grey has bill shape :: cone ... has head pattern :: eyebow Ours: LLM Generated Concepts ... grayish brown back and wings black throat with a white boarder brown head with white stripes ... !" GPT-3 prompt: describe what the black-throated sparrow looks like: Figure 1. Our proposed high-performance Concept Bottleneck Model alleviates the need for human-designed concepts by prompt- ing large language models (LLMs) such as GPT-3 [ 4]. can also be designed to be inherently interpretable, but it is believed that such models will perform more poorly than their black box alternatives [ 16]. In this work, we provide evidence to the contrary. We show how to construct high- performance interpretable-by-design classiﬁers by combin- ing a language model, GPT-3 [ 4], and a language-vision model, CLIP [ 44]. Our method builds on Concept Bottleneck Models (CBM) [ 25], which construct predictors through a linear combination of human-designed concepts. For example, as seen in Figure 1, a qualiﬁed person can design concepts, such as “nape color,” as intermediate targets for a black box model before classifying a bird. CBMs provide abstractions that people can use to understand errors or intervene on, contributing to increased trust. Application of CBMs is limited because they require costly attribute annotations by domain experts and often under-perform their black box counterparts. In contexts where CBM performance is competitive with black box al- ternatives, interpretability properties are sacriﬁced [ 34,70]. To address both of these challenges, we propose to build systems that automatically construct CBMs. OurLanguage Model Guided Concept Bottleneck Model (LaBo ), Figure 2, allows for the automatic construction of high-performance CBMs for arbitrary classiﬁcation prob- lems without concept annotations. Large language models This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19187 dot product softmax 𝜎 class 2-red panda LLM candidates S2 submodular optimization ℱ class N-tree frog LLM candidates SN submodular optimization ℱ Generate concepts: Sec 3.4 prompt: describe what the axolotl looks like: LLM: The axolotl's limbs are delicate, and the tail is long and thin. Extract concept using LM and delete class names: Candidate concepts: limbs are delicate; tail is long and thin Select concepts: Sec 3.2 c1,1 c1,2 c1,k ... k concepts limbs are delicate tail is long and thin gills are bright pink ... c2,1 c2,2 c2,k ... k concepts thick, soft fur reddish brown fur ears are also red ... k concepts cN,1 cN,2 cN,k ... toes are long green body smooth bumpy skin ... ... ... ... ... ... ... ... ... ... ... ... text encoder Concept Space 𝑬𝐶∈ℝ𝑁𝐶×𝑑 concept scores: 𝑔(𝒙,𝐶)= 𝒙⋅𝑬𝐶T∈ℝ𝑁𝑐 (1,𝑁𝐶) Class-Concept Weight Matrix W ∈ℝ𝑁×𝑁𝑐 image encoder x ∈ℝ𝑑 test image 𝑦̂=argmax(𝑔(𝒙,𝐶)⋅𝜎(𝑾)T) (1,𝑁) Predict the label with concepts: Sec 3.3 Bottleneck-C (NC concepts) N classes submodular optimization LLM class 1-axolotl candidates S1 ℱ Figure 2. We present an overview of our Language-Model-Guided Concept Bottleneck Model ( LaBo ), which is interpretable by design image classiﬁcation system. First, we prompt the large language model (GPT-3) to generate candidate concepts (Sec 3.4). Second, we employ a submodular function to select concepts from all candidates to construct the bottleneck (Sec 3.2). Third, we apply a pretrained alignment model (CLIP) to obtain the embeddings of concepts and images, which is used to compute concept scores. Finally, we train a linear function in which the weight Wdenotes the concept-class association user to predict targets based on concept scores (Sec 3.3). (LLMs) contain signiﬁcant world knowledge [ 21,42,61], that can be elicited by inputting a string preﬁx and allowing LLMs to complete the string (prompting). For example, in Figure 1, GPT-3 is prompted about sparrows and completes with information such as “brown head with white stripes.” LaBo leverages this by constructing bottlenecks where the concepts are such GPT-3 generated sentences. Since our concepts are textual, we use CLIP to score their presence in an image and form a bottleneck layer out of these scores. A key advantage of LaBo is the ability to control the selection of concepts in the bottleneck by generating can- didates from the language model. We develop selection principles targeting both interpretability and classiﬁcation accuracy. For example, we prefer smaller bottlenecks that include shorter sentences that do not include class names. Furthermore, to maximize performance, we prefer attributes that CLIP can easily recognize and are highly discriminative. To account for appearance variation, we select attributes that cover a variety of information and are not repetitive. We formulate these factors into a novel sub-modular criterion that allows us to select good bottlenecks efﬁciently [ 38]. We have evaluated LaBo-created bottlenecks on 11 di- verse image classiﬁcation tasks, spanning recognition of common objects [ 11,26] to skin tumors [ 64]. ﬁne-grained types [ 3,32,39,67], textures [ 10], actions [ 59], skin tu- mors [ 64], and satellite photographed objects [ 8].2Our 2The only dataset specialization we perform is prompt tuning for GPT-main ﬁnding is that LaBo is a highly effective prior for what concepts to look for, especially in low data regimes. In eval- uations comparing with linear probes, LaBo outperforms by as much as 11.7% at 1-shot and marginally underperforms given larger data settings. Averaged over many dataset sizes, LaBo bottlenecks are 1.5% more accurate than linear probes. In comparison to modiﬁcations of CBMs that improve per- formance by circumventing the bottleneck [ 70], we achieve similar or better results without breaking the CBM abstrac- tion. In extensive ablations, we study key trade-offs in bot- tleneck design and show our selection criteria are crucial and highlight several other critical design choices. Human evaluations indicate that our bottlenecks are largely understandable, visual, and factual. Finally, anno- tators ﬁnd our GPT-3 sourced bottlenecks are more factual and groundable than those constructed from WordNet or Wikipedia sentences. Overall, our experiments demonstrate that automatically designed CBMs can be as effective as black box models while maintaining critical factors con- tributing to their interpretability.
Xu_Zero-Shot_Dual-Lens_Super-Resolution_CVPR_2023	Abstract The asymmetric dual-lens conﬁguration is commonly available on mobile devices nowadays, which naturally stores a pair of wide-angle and telephoto images of the same scene to support realistic super-resolution (SR). Even on the same device, however , the degradation for model-ing realistic SR is image-speciﬁc due to the unknown ac-quisition process (e.g., tiny camera motion). In this paper , we propose a zero-shot solution for dual-lens SR (ZeDuSR), where only the dual-lens pair at test time is used to learn animage-speciﬁc SR model. As such, ZeDuSR adapts itself to the current scene without using external training data, and thus gets rid of generalization difﬁculty. However , there aretwo major challenges to achieving this goal: 1) dual-lens alignment while keeping the realistic degradation, and 2) effective usage of highly limited training data. To overcomethese two challenges, we propose a degradation-invariantalignment method and a degradation-aware training strat- egy to fully exploit the information within a single dual-lenspair . Extensive experiments validate the superiority of Ze- DuSR over existing solutions on both synthesized and real- world dual-lens datasets. The implementation code is avail- able at https://github.com/XrKang/ZeDuSR .	1. Introduction Mobile devices such as smartphones are generally equipped with an asymmetric camera system consisting of multiple lenses with different focal lengths. As a commonconﬁguration, with a wide-angle lens and a telephoto lens,one can capture the same scene with different ﬁeld-of-views(FoVs). Within the overlapped FoV , the wide-angle and telephoto images naturally store low-resolution (LR) and high-resolution (HR) counterparts for learning a realistic super-resolution (SR) model. This provides a feasible way to obtain an HR image with a large FoV at the same time,which is beyond the capability of the original device. There are a few pioneer works along this direction, which are referred to as dual-lens/camera/zoomed SR inter- ∗Equal contribution.†Corresponding author. Figure 1. Degradation kernel (estimated by KernelGAN [ 1]) clus- tering on wide-angle images from iPhone11 [ 43]. changeably. Beyond traditional methods that simply trans- fer the telephoto content to the wide-angle view in theoverlapped FoV , recent deep-learning-based methods en- able resolution enhancement of the full wide-angle image. Speciﬁcally, Wang et al. [43] utilize the overlapped FoV of dual-lens image pairs to learn a reference-based SR model,where the wide-angle view is super-resolved by using the telephoto view as a reference. However, the external train-ing data are synthesized with the predeﬁned bicubic degra- dation, which leads to generalization difﬁculty when the trained SR model is applied to real devices. To narrow the domain gap between the training and inference stages, theyfurther adopt a self-supervised adaptation strategy to ﬁne- tune the pretrained model on real devices. On the otherhand, Zhang et al. [53] adopt self-supervised learning to train an SR model directly on a dual-lens dataset to avoidthe predeﬁned degradation, with the assumption of consis- tent degradation on the same device. In practice, however, the degradation kernel for model- ing realistic SR is inﬂuenced by not only the camera opticsbut also tiny camera motion during the acquisition processon mobile devices [ 1,6,36], resulting in the image-speciﬁc degradation for each dual-lens pair, even on the same de- vice. Figure 1gives an exemplar analysis, where we es- timate the degradation kernels of 146 wide-angle images captured by the dual-lens device iPhone11 [ 43] and cluster them using t-SNE [ 41]. Although these images are captured by the same device, they exhibit notably different degrada- tion kernels due to the unknown acquisition process. This limits the performances of previous dual-lens SR methods for practical applications, since they assume that the realis- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9130 Wideʉangle Telephoto Input Dual ʉlensSRmethods ResultPredefined degradation Extra dualʉlensdataset Supervised learningDCSRDCSR SelfDZSR ZeDuSRDeviceʉconsistent degradation Extradualʉlensdataset Selfʉsupervised learningSelfDZSR Imageʉspecificdegradation No extra dualʉlensdataset ZeroʉshotlearningZeDuSR Figure 2. Overview of dual-lens SR methods and visual compari- son of 2×SR on the real-world data. tic degradation is at least device-consistent. In this paper, we propose a zero-shot learning solution for realistic SR on dual-lens devices, termed as ZeDuSR,which learns an image-speciﬁc SR model solely from thedual-lens pair at test time. As such, ZeDuSR adapts itselfto the current scene under the unknown acquisition process and gets rid of the generalization difﬁculty when using ex- ternal training data. Figure 2summarizes the main differ- ences of ZeDuSR from existing dual-lens SR methods with a real-world reconstruction example. ZeDuSR gives a visu-ally improved result compared with its competitors, thanksto the image-speciﬁc degradation assumption. There are two major challenges to achieving the success of ZeDuSR. First, as also noticed in previous works [ 43,53], it is non-trivial to exploit the information of the dual-lensimage pair, due to the spatial misalignment caused by thephysical offset between the two lenses. Moreover, we ﬁnd that the alignment process for generating the training data will introduce additional frequency information, which in-evitably changes the realistic degradation. To overcome this challenge, we propose to constrain the alignment pro-cess in spatial, frequency, and feature domains simulta-neously to keep the degradation. Speciﬁcally, we pro- pose a degradation-invariant alignment method by leverag- ing adversarial and contrastive learning. The other chal- lenge is how to effectively use the highly limited data forlearning an image-speciﬁc SR model. To this end, we design a degradation-aware training strategy to fully ex- ploit the information within a single dual-lens pair. Thatis, we calculate the probability of each location for patchcropping according to the degradation similarity between the images before and after alignment. As such, samplesthat keep the degradation are assigned a higher probabil-ity to be selected during training, while the contribution ofdegradation-variant samples is reduced. We evaluate the performance of ZeDuSR on both syn- thesized and real-world dual-lens datasets. For quantita- tive evaluation with the HR ground-truth, we adopt a stereodataset (Middlebury2021 [ 35]) and a light ﬁeld dataset (HCI new [ 12], with two views selected) to simulate dual- lens image pairs with different baselines between the twolenses, by applying image-speciﬁc degradation on one of the two views. Besides, we also perform the qualita- tive evaluation on two real-world datasets, i.e., CameraFu- sion [ 43] captured by iPhone11 and RealMCVSR [ 18] cap- tured by iPhone12, where no additional degradation is in- troduced beyond the realistic one between the two views. Extensive experiments on both synthesized and real-worlddatasets demonstrate the superiority of our ZeDuSR overexisting solutions including single-image SR, reference-based SR, and dual-lens SR. Contributions of this paper are summarized as follows:•We propose a zero-shot learning solution for realis- tic SR on dual-lens devices, which assumes image-speciﬁc degradation and adapts itself to the current scene under the unknown acquisition process. •We propose a degradation-invariant alignment method by leveraging adversarial and contrastive learning to con-strain the alignment process in spatial, frequency, and fea-ture domains simultaneously. •We design a degradation-aware training strategy to ef- fectively exploit the information within the highly limitedtraining data, i.e., a single dual-lens pair at test time. •We conduct extensive experiments on both synthesized and real-world dual-lens datasets to validate the superiorityof our zero-shot solution.
Yang_Reconstructing_Animatable_Categories_From_Videos_CVPR_2023	Abstract Building animatable 3D models is challenging due to the need for 3D scans, laborious registration, and rigging. Re- cently, differentiable rendering provides a pathway to ob- tain high-quality 3D models from monocular videos, but these are limited to rigid categories or single instances. We present RAC, a method to build category-level 3D models from monocular videos, disentangling variations over in- stances and motion over time. Three key ideas are intro- duced to solve this problem: (1) specializing a category- level skeleton to instances, (2) a method for latent space regularization that encourages shared structure across a category while maintaining instance details, and (3) us- ing 3D background models to disentangle objects from the background. We build 3D models for humans, cats and dogs given monocular videos. Project page: https://gengshan- y.github.io/rac-www/ .	1. Introduction We aim to build animatable 3D models for deformable object categories. Prior work has done so for targeted cat- egories such as people (e.g., SMPL [ 1,30]) and quadruped animals (e.g., SMAL [ 4]), but such methods appear chal- lenging to scale due to the need of 3D supervision and reg- istration. Recently, test-time optimization through differ-entiable rendering [ 40,41,44,56,69] provides a pathway to generate high-quality 3D models of deformable objects and scenes from monocular videos. However, such models are typically built independently for each object instance or scene. In contrast, we would like to build category models that can generate different instances along with deforma- tions, given causally-captured video collections . Though scalable, such data is challenging to leverage in practice. One challenge is how to learn the morpho- logical variation of instances within a category. For ex- ample,husky s andcorgi s are bothdogs, but have dif- ferent body shapes, skeleton dimensions, and texture ap- pearance. Such variations are difﬁcult to disentangle from the variations within a single instance, e.g., as a dog artic- ulates, stretches its muscles, and even moves into differ- ent illumination conditions. Approaches for disentangling such factors require enormous efforts in capture and reg- istration [ 1,5], and doing so without explicit supervision remains an open challenge. Another challenge arises from the impoverished nature of in-the-wild videos: objects are often partially observ- able at a limited number of viewpoints, and input signals such as segmentation masks can be inaccurate for such “in- the-wild” data. When dealing with partial or impoverished video inputs, one would want the model to listen to the com- mon structures learned across a category – e.g., dogs have two ears. On the other hand, one would want the model to This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16995 Morphology βArticulation θ Figure 2. Disentangling morphologies βand articulation θ.We show different morphologies (body shape and clothing) given the same rest pose ( top) and bouncing pose ( bottom ). stay faithful to the input views. Our approach addresses these challenges by exploiting three insights: (1) We learn skeletons with constant bone lengths within a video, allowing for better disentanglement of between-instance morphology and within-instance artic- ulation. (2) We regularize the unobserved body parts to be coherent across instances while remaining faithful to the in- put views with a novel code-swapping technique. (3) We make use of a category-level background model that, while not 3D accurate, produces far better segmentation masks. We learn animatable 3D models of cats, dogs, and humans which outperform prior art. Because our models regis- ter different instances with a canonical skeleton, we also demonstrate motion transfer across instances.
Zhang_Structural_Multiplane_Image_Bridging_Neural_View_Synthesis_and_3D_Reconstruction_CVPR_2023	Abstract The Multiplane Image (MPI), containing a set of fronto- parallel RGB αlayers, is an effective and efficient represen- tation for view synthesis from sparse inputs. Yet, its fixed structure limits the performance, especially for surfaces im- aged at oblique angles. We introduce the Structural MPI (S- MPI), where the plane structure approximates 3D scenes concisely. Conveying RGB αcontexts with geometrically- faithful structures, the S-MPI directly bridges view synthe- sis and 3D reconstruction. It can not only overcome the critical limitations of MPI, i.e., discretization artifacts from sloped surfaces and abuse of redundant layers, and can also acquire planar 3D reconstruction. Despite the intu- ition and demand of applying S-MPI, great challenges are introduced, e.g., high-fidelity approximation for both RGB α layers and plane poses, multi-view consistency, non-planar regions modeling, and efficient rendering with intersected planes. Accordingly, we propose a transformer-based net- work based on a segmentation model [4]. It predicts com- pact and expressive S-MPI layers with their corresponding masks, poses, and RGB αcontexts. Non-planar regions are inclusively handled as a special case in our unified frame- work. Multi-view consistency is ensured by sharing global proxy embeddings, which encode plane-level features cov- ering the complete 3D scenes with aligned coordinates. In- tensive experiments show that our method outperforms both previous state-of-the-art MPI-based view synthesis methods and planar reconstruction methods.	1. Introduction Novel view synthesis [30, 51] aims to generate new im- ages from specifically transformed viewpoints given one or multiple images. It finds wide applications in augmented or mixed reality for immersive user experiences. The advance of neural networks mostly drives the re- cent progress. NeRF-based methods [28, 30] achieve im- pressive results but are limited in rendering speed and gen- *This work was done when Mingfang Zhang was an intern at MSRA. PlaneFormerStandardMPIOurs Single-/sparse-view inputStandard MPIStructural MPI3D ReconstructionView synthesis (D)(R) Ours(a) Standard MPI v.s. Structural MPI(b) View-synthesis result Input(c) Reconstruction result Figure 1. We propose the Structural Multiplane Image (S-MPI) representation to bridge the tasks of neural view synthesis and 3D reconstruction. It consists of a set of posed RGB αimages with ge- ometries approximating the 3D scene. The scene-adaptive S-MPI overcomes the critical limitations of standard MPI [20], e.g., dis- cretization artifacts (D) andrepeated textures (R) , and achieves a better depth map compared with the previous planar reconstruc- tion method, PlaneFormer [1]. eralizability. The multiplane image (MPI) representation [42, 50] shows superior abilities in these two aspects, es- pecially given extremely sparse inputs. Specifically, neural networks are utilized to construct MPI layers, containing a set of fronto-parallel RGB αplanes regularly sampled in a reference view frustum. Then, novel views are rendered in real-time through simple homography transformation and integral over the MPI layers. Unlike NeRF models, MPI models do not need another training for a new scene. Nevertheless, standard MPI has underlying limitations. 1) It is sensitive to discretization due to slanted surfaces in scenes. As all layered planes are parallel to the source im- age plane, slanted surfaces will be distributed to multiple MPI layers causing discretization artifacts in novel views, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16707 as shown in Fig. 1 (b). Increasing the number of layers can improve the representation capability [38] but also increase memory and computation costs. 2) It easily introduces re- dundancy. It tends to distribute duplicated textures into dif- ferent layers to mimic the lighting field [21], which can in- troduce artifacts with repeated textures as shown in Fig. 1 (b). The essential reason causing the above issues is that the MPI construction is dependent on source views but neglects the explicit 3D geometry of the scenes. Intuitively, we raise a question: Is it possible to construct MPIs adaptive to the scenes, considering both depths and orientations? Slanted planes are smartly utilized in 3D reconstruction to save stereo matching costs [11] or represent the scene compactly [12,17], especially for man-made environments. Recent advanced neural networks reach end-to-end planar reconstruction from images by formulating it as instance segmentation [25, 26], where planes are directly detected and segmented from the input image. However, non-planar regions are often not well-modeled or neglected [1, 25, 40], resulting in holes or discontinuities in depth maps, as shown in Fig. 1 (c). In this paper, we aim to bridge the neural view synthe- sis and the planar 3D reconstruction, that is, to construct MPIs adaptive to 3D scenes with planar reconstruction and achieve high-fidelity view synthesis. The novel representa- tion we propose is called Structural MPI (S-MPI), which is fully flexible in both orientations and depth offsets to approximate the scene geometry, as shown in Fig. 1 (a). Although our motivation is straightforward, there are great challenges in the construction of an S-MPI. (1) The network not only needs to predict RGB αvalues but also the planar approximation of the scene. (2) It is difficult to correspond the plane projections across views since they may cover dif- ferent regions and present different appearances. Recent plane reconstruction works [1, 18] build matches of planes after they are detected in each view independently, which may increase costs and accumulate errors. (3) Non-planar regions are challenging to model even with free planes. Pre- vious plane estimation methods [1,40,46] cannot simultane- ously handle planar and non-planar regions well. (4) In the rendering process, as an S-MPI contains planes intersecting with each other, an efficient rendering pipeline needs to be designed so that the rendering advantages of MPI can be inherited. To address these challenges, we propose to build an S-MPI with an end-to-end transformer-based model for both planar geometry approximation and view synthesis, in which planar and non-planar regions are processed jointly. We follow the idea [25] of formulating plane detection as instance segmentation and leverage the segmentation net- work [4]. Our S-MPI transformer uniformly takes planar and non-planar regions as two structure classes and predicts their representative attributes, which are for reconstruction(structure class, plane pose and plane mask) and view syn- thesis (RGB αimage). We term each instance with such attributes as a proxy . Note that non-planar layers are inclu- sively handled as fronto-parallel planes with adaptive depth offsets and the total number of the predicted proxy instances is adaptive to the scene. Our model can manipulate both single-view and multi- view input. It aims to generate a set of proxy embed- dings in the full extent of the scene, covering all planar and non-planar regions aligned in a global coordinate frame. For multi-view input, the proxy embeddings progressively evolve to cover larger regions and refine plane poses as the number of views increases. In this way, the predicted proxy instances are directly aligned, which avoids the so- phisticated matching in two-stage methods [1, 18]. The global proxy embeddings are effectively learned with the ensembled supervision from all local view projections. Our model achieves state-of-the-art performance for single-view view synthesis ( 10%↑PSNR) and planar reconstruction (20%↑recall) in datasets [6, 36] of man-made scenes and also achieve encouraging results for multi-view input com- pared to NeRF-based methods with high costs. In summary, our main contributions are as follows: • We introduce the Structural MPI representation, con- sisting of geometrically-faithful RGB αimages to the 3D scene, for both neural view synthesis and 3D re- construction. • We propose an end-to-end network to construct S-MPI, where planar and non-planar regions are uniformly handled with high-fidelity approximations for both ge- ometries and light filed. • Our model ensures multi-view consistency of planes by introducing the global proxy embeddings compre- hensively encoding the full 3D scene, and they effec- tively evolve with the ensembled supervision from all views.
Yao_Visual-Language_Prompt_Tuning_With_Knowledge-Guided_Context_Optimization_CVPR_2023	Abstract Prompt tuning is an effective way to adapt the pretrained visual-language model (VLM) to the downstream task us- ing task-related textual tokens. Representative CoOp-based work combines the learnable textual tokens with the class tokens to obtain speciﬁc textual knowledge. However, the speciﬁc textual knowledge is worse generalization to the unseen classes because it forgets the essential general tex- tual knowledge having a strong generalization ability. To tackle this issue, we introduce a novel Knowledge-guided Context Optimization (KgCoOp) to enhance the generaliza- tion ability of the learnable prompt for unseen classes. The key insight of KgCoOp is that the forgetting about essen- tial knowledge can be alleviated by reducing the discrep- ancy between the learnable prompt and the hand-crafted prompt. Especially, KgCoOp minimizes the discrepancy be- tween the textual embeddings generated by learned prompts and the hand-crafted prompts. Finally, adding the Kg- CoOp upon the contrastive loss can make a discriminative prompt for both seen and unseen tasks. Extensive eval- uation of several benchmarks demonstrates that the pro- posed Knowledge-guided Context Optimization is an efﬁ- cient method for prompt tuning, i.e.,achieves better perfor- mance with less training time. code.	1. Introduction With the help of the large scale of the image-text as- sociation pairs, the trained visual-language model (VLM) contains essential general knowledge, which has a bet- ter generalization ability for the other tasks. Recently, many visual-language models have been proposed, such as Contrastive Language-Image Pretraining (CLIP) [29], Flamingo [1], ALIGN [18], etc. Although VLM is an ef- fective model for extracting the visual and text description, training VLM needs a large scale of high-quality datasets. However, collecting a large amount of data for training a task-related model in real visual-language tasks is dif- ﬁcult. To address the above problem, the prompt tun-Table 1. Compared to existing methods, the proposed KgCoOp is an efﬁcient method, obtaining a higher performance with less training time . Methods PromptsAccuracyTraining-timeBase New H CLIP hand-crafted 69.34 74.22 71.70 - CoOp textual 82.63 67.99 74.60 6ms/image ProGrad textual 82.48 70.75 76.16 22ms/image CoCoOp textual+visual 80.47 71.69 75.83 160ms/image KgCoOp textual 80.73 73.6 77.0 6ms/image ing [4] [10] [19] [22] [28] [30] [33] [38] has been proposed to adapt the pretrained VLM to downstream tasks, achiev- ing a fantastic performance on various few-shot or zero-shot visual recognization tasks. The prompt tuning1usually applies task-related textual tokens to embed task-speciﬁc textual knowledge for pre- diction. The hand-crafted template “a photo of a [Class]” in CLIP [29] is used to model the textual-based class em- bedding for zero-shot prediction. By deﬁning the knowl- edge captured by the ﬁxed (hand-crafted) prompts as the general textual knowledge2, it has a high generalization capability on unseen tasks. However, the general textual knowledge is less able to describe the downstream tasks due to not consider the speciﬁc knowledge of each task. To obtain discriminative task-speciﬁc knowledge, Context Optimization(CoOp) [41], Conditional Context Optimiza- tion(CoCoOp) [40], and ProGrad [42] replace the hand- crafted prompts with a set of learnable prompts inferred by the labeled few-shot samples. Formally, the discrimina- tive knowledge generated by the learned prompts is deﬁned as the speciﬁc textual knowledge . However, CoOp-based methods have a worse generalization to the unseen classes with the same task, e.g.,obtaining a worse performance than CLIP for the unseen classes ( New), shown in Table 1. As the speciﬁc textual knowledge is inferred from the 1In this work, we only consider the textual prompt tuning and do not involve the visual prompt tuning. 2Inspired from [17], ‘knowledge’in this work denotes the information contained in the trained model. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6757 00.050.10.150.20.250.30.350.40.450.5 0510152025 ImageNet Caltech101 OxfordPets StandfordCars Flowers Food101 FGVCAircraft SUN397 DTD EuroSAT UCF101∇=(−)/= − 22 Figure 1. For the CoOp-based prompt tuning, the degree of per- formance degradation Onewon the New classes is consistent with the distance between the learnable textual embedding wcoop and the hand-crafted textual embedding wclip. The larger distance, the more severe the performance degradation. clipandcoop are the accuracy of New classes for CLIP and CoOp, respectively. labeled few-shot samples, it is discriminative for the seen classes and biased away from the unseen classes, lead- ing to worse performance on the unseen domain. For example, non-training CLIP obtains a higher New accu- racy on the unseen classes than CoOp-based methods, e.g., 74.22%/63.22%/71.69% for CLIP/CoOP/CoCoOp. The su- perior performance of CLIP on unseen classes veriﬁes that its general textual knowledge has a better generalization for unseen classes. However, the speciﬁc textual knowl- edge inferred by the CoOp-based methods always forgets the essential general textual knowledge, called catastrophic knowledge forgetting, i.e.,the more serve catastrophic for- getting, the larger performance degradation. To address this issue, we introduce a novel prompt tun- ing method Knowledge-guided Context Optimization (Kg- CoOp) to boost the generality of the unseen class by reduc- ing the forgetting of the general textual knowledge. The key insight of KgCoOp is that the forgetting about gen- eral textual knowledge can be alleviated by reducing the discrepancy between the learnable prompt and the hand- crafted prompt. The observation of relationship between the discrepancy of two prompts and the performance drop also verify the insight. As shown in Figure 1, the larger the distance between textual embeddings generated by the learnable prompt and the hand-crafted prompt, the more severe the performance degradation. Formally, the hand- crafted prompts “a photo of a [Class]” are fed into the text encoder of CLIP to generate the general textual embedding, regarded as the general textual knowledge. Otherwise, a set of learnable prompts is optimized to generate task-speciﬁc textual embedding. Furthermore, Knowledge-guided Con- text Optimization(KgCoOp) minimizes the euclidean dis- tance between general textual embeddings and speciﬁc tex- tual embeddings for remembering the essential general tex- tual knowledge. Similar to the CoOp and CoCoOp, the con- trastive loss between the task-speciﬁc textual and visual em-beddings is used to optimize the learnable prompts. We conduct comprehensive experiments under base-to- new generalization setting, few-shot classiﬁcation, and do- main generalization over 11 image classiﬁcation datasets and four types of ImageNets. The evaluation shows that the proposed KgCoOp is an efﬁcient method: using the less training time obtains a higher performance, shown in Ta- ble 1. In summary, the proposed KgCoOp obtains: 1) higher performance: KgCoOp obtains a higher ﬁnal performance than existing methods. Especially, KgCoOp obtains a clear improvement on the New class upon the CoOp, CoCoOp, and ProGrad, demonstrating the rationality and necessity of considering the general textual knowledge. 2) less training time: the training time of KgCoOp is the same as CoOp, which is faster than CoCoOp and ProGrad.
Zhang_Revisiting_the_Stack-Based_Inverse_Tone_Mapping_CVPR_2023	Abstract Current stack-based inverse tone mapping (ITM) meth- ods can recover high dynamic range (HDR) radiance by predicting a set of multi-exposure images from a single low dynamic range image. However, there are still some limitations. On the one hand, these methods estimate a fixed number of images (e.g., three exposure-up and three exposure-down), which may introduce unnecessary com- putational cost or reconstruct incorrect results. On the other hand, they neglect the connections between the up- exposure and down-exposure models and thus fail to fully excavate effective features. In this paper, we revisit the stack-based ITM approaches and propose a novel method to reconstruct HDR radiance from a single image, which only needs to estimate two exposure images. At first, we de- sign the exposure adaptive block that can adaptively adjust the exposure based on the luminance distribution of the in- put image. Secondly, we devise the cross-model attention block to connect the exposure adjustment models. Thirdly, we propose an end-to-end ITM pipeline by incorporating the multi-exposure fusion model. Furthermore, we propose and open a multi-exposure dataset that indicates the opti- mal exposure-up/down levels. Experimental results show that the proposed method outperforms some state-of-the-art methods.	1. Introduction The luminance distribution in nature spans a wide range, from the starlight ( 10−5cd/m2) to the direct sunlight (108cd/m2). The low dynamic range (LDR) devices can not cover the full range of luminance of the real scene, and thus fail to reproduce the realistic visual experience. High dynamic range imaging (HDRI) technology [1] [26] can solve this problem, which takes multiple LDR images of the same scene with different shutter time, and then gener- ates the high dynamic range (HDR) image via the multi- exposure fusion (MEF) method [4] [19] [32]. However, HDRI cannot handle the images that have already been cap- input LDR+1 +2 +3 -1 -2 -3 + EV - EVground truthPrevious stack-based method Proposed methodestimating six images estimating only two images Fixed Fusion tone mapping Q-score 59.6612 Q-score 63.9265Up-exposure Down-exposure Up-exposure Down-exposureCross modelLearning- based Fusion tone mapping61.4861.9262.4162.84 62.73 62.7263.12 62.94 61.38 59.6360.4761.2161.89 60.96 60.42 59.560.561.562.563.5 T=3 T=5 T=7 T=9 T=11HDR-VDP Score Length of Multi Exposure StackThe influences of different stack lengths DrTMO [7]Deep Recursive [12]Deep Chain [11] (a) The influences of different stack lengths (b) Comparison on the ITM pipelineFigure 1. (a) The influences of different stack lengths demonstrate that the stack length is important to the quality of reconstructed HDR image. However, the previous ITM methods [7] [11] [12] [10] simply set a fixed length, which cannot be the optimal choice for every scene. (b) The comparison between the MES predicted by the state-of-the-art stack-based method [10] and the proposed method. Our method only needs to estimate two exposure im- ages to recover more realistic details in highlights and shadows and achieves a higher HDR-VDP-2.2 Q-score. The HDR images are tone mapped by [15] for LDR display. tured, such as a large number of LDR images and videos on the Internet. The inverse tone mapping (ITM) technique is thus designed to recover the HDR radiance from a single LDR image, which is an ill-posed problem because the de- tails in the highlights and shadows are almost lost and dif- ficult to be restored. Fortunately, the development of deep learning [6] [7] provides a solution by learning and predict- ing the distribution of the lost information from the huge amount of training examples. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9162 There are two main deep-learning-based ITM ap- proaches, i.e., direct mapping methods [6] [17] [22] [28] [34] and stack-based methods [7] [10] [11] [12] [13]. The direct mapping methods learn an end-to-end model to re- cover the HDR radiance from the LDR input straightfor- wardly. By contrast, the stack-based methods simulate the HDRI technology by increasing and decreasing the expo- sure value (EV) of the input image to obtain the multi- exposure stack (MES). Compared to the direct mapping methods, the stack-based methods simulate the generation process of HDR images in the real world and perform the learning stage in the same LDR space, which avoids the sophisticated changes between the LDR and HDR domain [10] [11] [13]. The process of the previous stack-based based ITM methods can be roughly summarized in the following three main steps: (1) Training an up-exposure model and a down- exposure model separately and independently. (2) Gen- erating a fixed number of exposure adjusted images (e.g., three up-exposure and three down-exposure images) with the trained models. (3) Merging these images by a clas- sic multi-exposure fusion (MEF) approach [4]. However, each of these three steps has limitations that may cause in- accurate results. Specifically, (1) the process of increasing and decreasing the exposure should not be independent. For instance, when decreasing the exposure, some useful infor- mation of the under-exposed regions may become subtle, while the features in the opposite increasing process can compensate for it. (2) The times of exposure adjustment are important to the quality of reconstructed HDR images and a fixed length of MES will cause incomplete information recovery or introduce an unnecessary computational cost, as shown in Fig. 1 (a). (3) The classic MEF approaches cannot be integrated into the entire end-to-end training pro- cess. Although Kim et al. [10] proposed a differentiable HDR synthesis layer, it is time-consuming and highly de- pends on the shutter time and therefore cannot be applied to general scenes. In this paper, we propose a novel HDR reconstruction method with adaptive exposure adjustment, which provides an effective solution to the existing limitations in the field of stack-based ITM. At first, we design an efficient encoder equipped with the luminance-guided convolution (LGC) and cross-model attention block (CMAB) to extract useful information from local and cross-model features. With the help of CMAB, valid information on the entire up-exposure and down-exposure process can be fully explored to help their reconstruction. Secondly, the proposed up-exposure and down-exposure models can adjust the input LDR im- age only once to obtain the corresponding optimal expo- sure adjusted result. In this way, we can avoid the difficulty of determining the length of the MES and get the desired exposure adjustment directly. For this purpose, appropri-ate ground truth is needed to indicate the optimal exposure level. Therefore, we improve the SICE [2] dataset to form a new MES dataset with optimal exposure labels. On the other hand, since the exposure levels of the labels are dif- ferent, the models need to be able to generate different re- sults adaptively based on different inputs. Consequently, we devise the exposure adaptive block (EAB) to extract the global information and remap the features of the decoder. The features extracted from EAB are used to normalize the features in the decoder, which results in the image-adaptive capability. Thirdly, we propose a lightweight and fast multi- exposure fusion model (MEFM), which can merge the ex- posure adjusted results with the input image into the de- sired HDR image and thus make the whole pipeline end-to- end. Furthermore, we propose progressive reconstruction loss and mask-aware generative adversarial loss to avoid the artifacts in the restored textures of over/under-exposed re- gions. As Fig. 1 (b) shows, the proposed method only needs to estimate two exposure values to recover the lost infor- mation in the shadows and highlights respectively, which is more concise and effective. Experiments show that our ITM algorithm outperforms the state-of-the-art ITM methods in both quantitative and qualitative evaluations. This paper has the following main contributions: (1). We propose a novel stack-based ITM framework, which only needs to estimate two exposure images to form the MES. In this way, the lost information can be recov- ered more efficiently and precisely. Moreover, the exposure adaptive block is designed to adaptively adjust the exposure based on LDR inputs with different luminance distributions. (2). We connect the up-exposure and down-exposure models with the designed cross-model attention block, which can fully extract the effective features of the image regions with different luminance. (3). A lightweight and fast multi-exposure fusion net- work is proposed that can merge the generated results and makes the entire training pipeline end-to-end. (4). A more concise MES dataset is proposed and opened based on the SICE dataset [2], which contains the optimal exposure-up/down labels to train the adaptive exposure ad- justment networks.
Zara_AutoLabel_CLIP-Based_Framework_for_Open-Set_Video_Domain_Adaptation_CVPR_2023	Abstract Open-set Unsupervised Video Domain Adaptation (OU- VDA) deals with the task of adapting an action recognition model from a labelled source domain to an unlabelled tar- get domain that contains “target-private” categories, which are present in the target but absent in the source. In this work we deviate from the prior work of training a spe- cialized open-set classifier or weighted adversarial learn- ing by proposing to use pre-trained Language and Vision Models (CLIP). The CLIP is well suited for OUVDA due to its rich representation and the zero-shot recognition ca- pabilities. However, rejecting target-private instances with the CLIP’s zero-shot protocol requires oracle knowledge about the target-private label names. To circumvent the impossibility of the knowledge of label names, we propose AutoLabel that automatically discovers and generates object-centric compositional candidate target-private class names. Despite its simplicity, we show that CLIP when equipped with AutoLabel can satisfactorily reject the target-private instances, thereby facilitating better align- ment between the shared classes of the two domains. The code is available1.	1. Introduction Recognizing actions in video sequences is an important task in the field of computer vision, which finds a wide range of applications in human-robot interaction, sports, surveillance, and anomalous event detection, among others. Due to its high importance in numerous practical applica- tions, action recognition has been heavily addressed using deep learning techniques [32, 44, 58]. Much of the success in action recognition have noticeably been achieved in the supervised learning regime [6, 17, 49], and more recently shown to be promising in the unsupervised regime [20, 34, 56] as well. As constructing large scale annotated and cu- rated action recognition datasets is both challenging and ex- pensive, focus has shifted towards adapting a model from a source domain, having a labelled source dataset, to an un- 1https://github.com/gzaraunitn/autolabellabelled target domain of interest. However, due to the dis- crepancy (or domain shift ) between the source and target domains, naive usage of a source trained model in the target domain leads to sub-optimal performance [51]. To counter the domain shift and and improve the trans- fer of knowledge from a labelled source dataset to an unla- belled target dataset, unsupervised video domain adaptation (UVDA) methods [7, 9, 42] have been proposed in the liter- ature. Most of the prior literature in UVDA are designed with the assumption that the label space in the source and target domain are identical. This is a very strict assumption, which can easily become void in practice, as the target do- main may contain samples from action categories that are not present in the source dataset [43]. In order to make UVDA methods more useful for practical settings, open- set unsupervised video domain adaptation (OUVDA) meth- ods have recently been proposed [5, 8]. The main task in OUVDA comprise in promoting the adaptation between the shared (orknown ) classes of the two domains by excluding the action categories that are exclusive to the target domain, also called as target-private (orunknown ) classes. Exist- ing OUVDA prior arts either train a specialized open-set classifier [5] or weighted adversarial learning strategy [8] to exclude the target-private classes. Contrarily, we address OUVDA by tapping into the very rich representations of the open-sourced foundation Lan- guage and Vision Models (LVMs). In particular, we use CLIP (Contrastive Language-Image Pre-training) [45], a foundation model that is trained on web-scale image-text pairs, as the core element of our framework. We argue that the LVMs ( e.g., CLIP) naturally lend themselves well to OUVDA setting due to: (i)the representation learned by LVMs from webly supervised image-caption pairs comes encoded with an immense amount of prior about the real- world, which is (un)surprisingly beneficial in narrowing the shift in data distributions, even for video data; (ii)the zero- shot recognition capability of such models facilitates identi- fication and separation of the target-private classes from the shared ones, which in turn ensures better alignment between the known classes of the two domains. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11504 ride horseclimbfencingshoot ball{Oracle zero-shot classesKnownclasses{A video of {action}TextEncoderZero-shot videoVisionEncoderV1T1T2T3T4V1/uni22C5T1V1/uni22C5T2V1/uni22C5T3V1/uni22C5T4A video of {ride horse}(a) Zero-shot prediction using CLIP climbfencingKnownclasses{A video of {action}TextEncoder Target-private videoVisionEncoderV1T1T2T3T4V1/uni22C5T1V1/uni22C5T2V1/uni22C5T3V1/uni22C5T4A video of {horse person}Target DatasetAutoLabelhorse personperson ballCandidatetarget-privateclasses (b) Rejecting target-private instances with our proposed AutoLabel Figure 1. Comparison of AutoLabel with CLIP [45] for zero-shot prediction on target-private instances. (a) CLIP assumes the knowledge about the oracle zero-shot classes names ( ride horse, shoot ball ); (b) Our proposed AutoLabel discovers automatically the candidate target-private classes ( horse person, person ball ) and extends the known classes label set Zero-shot inference using CLIP requires multi-modal in- puts, i.e.,a test video and a set of allpossible prompts “ A video of {label }”, where label is a class name, for computing the cosine similarity (see Fig. 1a). However in the OUVDA scenario, except the shared classes, a priori knowledge about the target-private classes label names are not available (the target dataset being unlabelled). Thus, ex- ploiting zero-shot capability of CLIP to identify the target- private instances in an unconstrained OUVDA scenario be- comes a bottleneck. To overcome this issue we propose AutoLabel , an automatic labelling framework that con- structs a set of candidate target-private class names, which are then used by CLIP to potentially identify the target- private instances in a zero-shot manner. In details, the goal of AutoLabel is to augment the set of shared class names (available from the source dataset) with a set of candidate target-private class names that best represent the true target-private class names in the tar- get dataset at hand (see Fig. 1b). To this end, we use an external pre-trained image captioning model ViLT [31] to extract a set of attribute names from every frame in a video sequence (see Sec. 3.2.1 for details). This is motivated by the fact that actions are often described by the constituent objects and actors in a video sequence. As an example, a video with the prompt “ A video of{chopping onion }” can be loosely described by the proxy prompt “ A video of {knife },{onion } and{arm}” crafted from the predicted attribute names. In other words, the attributes “ knife ”, “onion ” and “ arm” when presented to CLIP in a prompt can elicit similar re- sponse as the true action label “ chopping onion ”. Naively expanding the label set using ViLT predicted at- tributes can introduce redundancy because: (i)ViLT pre- dicts attributes per frame and thus, there can be a lot of dis- tractor object attributes in a video sequence; and (ii)ViLT predicted attributes for the shared target instances will be duplicates of the true source action labels. Redundancy in the shared class names will lead to ambiguity in target- private instance rejection. Our proposed framework AutoLabel reduces the re- dundancy in the effective label set in the following manner.First, it uses unsupervised clustering ( e.g., k-means [36]) on the target dataset to cluster the target samples, and then con- structs the top- kmost frequently occurring attributes among the target samples that are assigned to each cluster. This step gets rid of the long-tailed set of attributes, which are inconsequential for predicting an action (see 3.2.2 for de- tails). Second, AutoLabel removes the duplicate sets of attributes that bear resemblance with the source class names (being the same shared underlying class) by using a set matching technique. At the end of this step, the effective label set comprises the shared class names and the candi- date sets of attribute names that represent the target-private class names (see Sec. 3.2.3 for details). Thus, AutoLabel unlocks the zero-shot potential of the CLIP, which is very beneficial in unconstrained OUVDA. Finally, to transfer knowledge from the source to the tar- get dataset, we adopt conditional alignment using a sim- plepseudo-labelling mechanism. In details, we provide to the CLIP-based encoder the target samples and the extended label set containing the shared and candidate target-private classes. Then we take the top- kpseudo-labelled samples for each predicted class and use them for optimizing a super- vised loss (see Sec. 3.2.4 for details). Unlike many open-set methods [5, 8] that reject all the target-private into a single unknown category, AutoLabel allows us to discriminate even among the target-private classes. Thus, the novelty of ourAutoLabel lies not only in facilitating the rejection of target-private classes from the shared ones, but also opens doors to open world recognition [1]. In summary, our contributions are: (i)We demonstrate that the LVMs like CLIP can be harnessed to address OU- VDA, which can be excellent replacement to complicated alignment strategies; (ii)We propose AutoLabel , an au- tomatic labelling framework that discovers candidate target- private classes names in order to promote better separation of shared and target-private instances; and (iii)We conduct thorough experimental evaluation on multiple benchmarks and surpass the existing OUVDA state-of-the-art methods.
Yi_NAR-Former_Neural_Architecture_Representation_Learning_Towards_Holistic_Attributes_Prediction_CVPR_2023	Abstract With the wide and deep adoption of deep learning mod- els in real applications, there is an increasing need to model and learn the representations of the neural networks them- selves. These models can be used to estimate attributes of different neural network architectures such as the accu- racy and latency, without running the actual training or inference tasks. In this paper, we propose a neural ar- chitecture representation model that can be used to esti- mate these attributes holistically. Specifically, we first pro- pose a simple and effective tokenizer to encode both the operation and topology information of a neural network into a single sequence. Then, we design a multi-stage fu- sion transformer to build a compact vector representation from the converted sequence. For efficient model train- ing, we further propose an information flow consistency augmentation and correspondingly design an architecture consistency loss, which brings more benefits with less aug- mentation samples compared with previous random aug- mentation strategies. Experiment results on NAS-Bench- 101, NAS-Bench-201, DARTS search space and NNLQP show that our proposed framework can be used to pre- dict the aforementioned latency and accuracy attributes of both cell architectures and whole deep neural networks, and achieves promising performance. Code is available at https://github.com/yuny220/NAR-Former.	1. Introduction As an ever increasing variety of deep neural network models are widely adopted in academic research and real applications, neural architecture representation is emerg- ing as an universal need to predict model attributes holis- tically. For example, modern neural architecture search (NAS) methods can depend on the neural architecture rep- 1This work was done while Yun Yi was an intern at Intellifusion. ∗Corresponding authors.resentation to build good model accuracy predictors [5, 23, 24, 26, 31, 34], which estimate model accuracies without running the expensive training procedure. In order to find faster execution graphs when deploying models to neural network accelerators, neural network compilers [3,20] need it to build network latency predictors, which estimate the real time cost without running it on the corresponding real hardware. As a straightforward approach to solve these holistic prediction tasks, a neural architecture representa- tion model should be built to take the symbolic description of the network as input, and generate its numerical repre- sentation which can be easily used with existing modeling tools to perform the desired downstream tasks. There are some neural architecture representation mod- els proposed in solving the individual tasks mentioned above, but we find they are rather application specific and have obvious drawbacks when used in new tasks. For ex- ample, the early MLP, LSTM based accuracy prediction approaches [6, 18, 24, 34] improve the efficiency and per- formance of NAS, while there is a limitation of predic- tion performance resulting from the the nonexistent or im- plicit topology encoding. Some later proposed GCN based methods [4, 17, 36] achieve better performance of accuracy prediction than the above methods. However, due to the use of adjacency matrix, the encoding dimension of these methods scales quadratically with the depth of the input ar- chitecture, making them difficult to model large networks. NNLQP [20] implements latency prediction at the model level, but GNN-based encoding scheme makes it inadequate in modeling long range interactions between nodes. Inspired by the progress in natural language understand- ing, our network uses a hand designed yet general and ex- tensible token encoding approach to encode the topology information and key parameters of the input neural net- work. Specifically, we design a generic real value tokenizer to encode neural network nodes’ operation type, location, and their inputs into vectors using the positional embed- ding approach that is widely used in transformers [33] and NERF [25]. This fixed length node encoding scheme makes This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7715 the network encoding scale linearly with the size of the in- put network rather than quadratically as in models that rely on taking adjacency matrix as input. Based on the tokenized input sequence, we design a transformer based model to further encode and fuse the net- work descriptions to generate a compact representation, e.g. a fixed length feature vector. Therefore, long range depen- dencies between nodes can be established by transformer. Besides the conventional multi-head self attention based transformers, we propose to use a multi-stage fusion trans- former structure at the deep stages of the model to further fuse and refine the representation in a cascade manner. For the network representation learning task, we also propose a data augmentation method called information flow consistency augmentation. The augmentation per- mutes the encoding order of nodes in the original network under our specified conditions without changing the struc- ture of the network. We find it empirically improves the performance of our method for downstream tasks. The main contributions of this paper can be summarized as the following points: 1. We propose a simple and effective neural network en- coding approach which tokenizes both operation and topology information of a neural network node into a sequence. When it is used to encode the entire net- work, the tokenized network encoding scheme scales better than adjacency matrix based ones, and builds the foundation for seamlessly using transformer structures for network representation learning. 2. We design a multi-stage fusion transformer to learn feature representations. Benefiting from the proposed tokenizer, a concise pure transformer based neural ar- chitecture representation learning framework (NAR- Former) is proposed for the first time. Our NAR- Former makes full use of the capacity of transformers in handling sequence inputs, and gets promising per- formance. 3. To facilitate efficient model training, we propose an information flow consistency augmentation and cor- respondingly design an architecture consistency loss, which brings more benefits with less augmentation samples compared with the existing random augmen- tation strategy. We conduct extensive experiments on accuracy predic- tion, neural architecture search as well as latency prediction. Experiments demonstrate the effectiveness of the proposed representation model on processing both cell leveled net- work components and whole deep neural networks. Specif- ically, our accuracy predictor based on this representation achieves highly competitive accuracy performance on cell- based structures in NAS-Bench-101 [40] and NAS-Bench-201 [10] datasets. Compared with other predictor-based NAS methods [23, 26], we efficiently find the architecture with 97.52% accuracy in DARTS [19] by only querying 100 neural architectures. We also conduct latency predic- tion experiments on neural networks deeper than 200 layers to demonstrate the universality of our model.
Yu_Video_Probabilistic_Diffusion_Models_in_Projected_Latent_Space_CVPR_2023	Abstract Despite the remarkable progress in deep generative models, synthesizing high-resolution and temporally co- herent videos still remains a challenge due to their high- dimensionality and complex temporal dynamics along with large spatial variations. Recent works on diffusion models have shown their potential to solve this challenge, yet they suffer from severe computation- and memory-inefficiency that limit the scalability. To handle this issue, we propose a novel generative model for videos, coined projected la- tent video diffusion model (PVDM), a probabilistic dif- fusion model which learns a video distribution in a low- dimensional latent space and thus can be efficiently trained with high-resolution videos under limited resources. Specifi- cally, PVDM is composed of two components: (a) an autoen- coder that projects a given video as 2D-shaped latent vectors that factorize the complex cubic structure of video pixels and (b) a diffusion model architecture specialized for our new fac- torized latent space and the training/sampling procedure to synthesize videos of arbitrary length with a single model. Ex- periments on popular video generation datasets demonstrate the superiority of PVDM compared with previous video syn- thesis methods; e.g., PVDM obtains the FVD score of 639.7 on the UCF-101 long video (128 frames) generation bench- mark, which improves 1773.4 of the prior state-of-the-art.	1. Introduction Recent progresses of deep generative models have shown their promise to synthesize high-quality, realistic samples in various domains, such as images [ 9,27,41], audio [ 8,31,32], 3D scenes [ 6,38,48], natural languages [ 2,5],etc. As a next step forward, several works have been actively focusing on the more challenging task of video synthe- sis [12,18,21,47,55,67]. In contrast to the success in other domains, the generation quality is yet far from real-world videos, due to the high-dimensionality and complexity of videos that contain complicated spatiotemporal dynamics in high-resolution frames.Inspired by the success of diffusion models in handling complex and large-scale image datasets [ 9,40], recent ap- proaches have attempted to design diffusion models for videos [ 16,18,21,22,35,66]. Similar to image domains, these methods have shown great potential to model video distribution much better with scalability (both in terms of spatial resolution and temporal durations), even achieving photorealistic generation results [ 18]. However, they suffer from severe computation and memory inefficiency, as diffu- sion models require lots of iterative processes in input space to synthesize samples [ 51]. Such bottlenecks are much more amplified in video due to a cubic RGB array structure. Meanwhile, recent works in image generation have pro- posed latent diffusion models to circumvent the computation and memory inefficiency of diffusion models [ 15,41,59]. Instead of training the model in raw pixels, latent diffusion models first train an autoencoder to learn a low-dimensional latent space succinctly parameterizing images [ 10,41,60] and then models this latent distribution. Intriguingly, the ap- proach has shown a dramatic improvement in efficiency for synthesizing samples while even achieving state-of-the-art generation results [ 41]. Despite their appealing potential, however, developing a form of latent diffusion model for videos is yet overlooked. Contribution. We present a novel latent diffusion model for videos, coined projected latent video diffusion model (PVDM). Specifically, it is a two-stage framework (see Fig- ure1for the overall illustration): •Autoencoder: We introduce an autoencoder that repre- sents a video with three 2D image-like latent vectors by factorizing the complex cubic array structure of videos. Specifically, we propose 3D →2D projections of videos at each spatiotemporal direction to encode 3D video pix- els as three succinct 2D latent vectors. At a high level, we design one latent vector across the temporal direction to parameterize the common contents of the video ( e.g., background), and the latter two vectors to encode the mo- tion of a video. These 2D latent vectors are beneficial for achieving high-quality and succinct encoding of videos, as well as enabling compute-efficient diffusion model architecture design due to their image-like structure. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18456 Denoising autoencoderVideo encoder Video decoderLatent space Diffusion process ⋯ ⋯ Upsample residual block (shared)Attention layerDownsample residual block (shared): Concatenation Generate the future clip Generate the initial clip Figure 1. Overall illustration of our projected latent video diffusion model (PVDM) framework. PVDM is composed of two components: (a) (left) an autoencoder that maps a video into 2D image-like latent space (b) (right) a diffusion model operates in this latent space. •Diffusion model: Based on the 2D image-like latent space built from our video autoencoder, we design a new diffusion model architecture to model the video distribution. Since we parameterize videos as image- like latent representations, we avoid computation-heavy 3D convolutional neural network architectures that are conventionally used for handling videos. Instead, our architecture is based on 2D convolution network diffu- sion model architecture that has shown its strength in handling images. Moreover, we present a joint training of unconditional and frame conditional generative mod- eling to generate a long video of arbitrary lengths. We verify the effectiveness of our method on two popu- lar datasets for evaluating video generation methods: UCF- 101 [ 54] and SkyTimelapse [ 64]. Measured with Inception score (IS; higher is better [ 44]) on UCF-101, a represen- tative metric of evaluating unconditional video generation, PVDM achieves the state-of-the-art result of 74.40 on UCF- 101 in generating 16 frames, 256 ×256 resolution videos. In terms of Fréchet video distance (FVD; lower is better [ 58]) on UCF-101 in synthesizing long videos (128 frames) of 256×256 resolution, it significantly improves the score from 1773.4 of the prior state-of-the-art to 639.7. Moreover, com- pared with recent video diffusion models, our model shows a strong memory and computation efficiency. For instance, on a single NVIDIA 3090Ti 24GB GPU, a video diffusion model [ 21] requires almost full memory ( ≈24GB) to train at 128×128 resolution with a batch size of 1. On the other hand, PVDM can be trained with a batch size of 7 at most per this GPU with 16 frames videos at 256 ×256 resolution. To our knowledge, the proposed PVDM is the first latent diffusion model designed for video synthesis. We believe our work would facilitate video generation research towards efficient real-time, high-resolution, and long video synthesis under the limited computational resource constraints.
Zhang_MOTRv2_Bootstrapping_End-to-End_Multi-Object_Tracking_by_Pretrained_Object_Detectors_CVPR_2023	Abstract In this paper, we propose MOTRv2, a simple yet effective pipeline to bootstrap end-to-end multi-object tracking with a pretrained object detector. Existing end-to-end methods, e.g. MOTR [43] and TrackFormer [20] are inferior to their tracking-by-detection counterparts mainly due to their poor detection performance. We aim to improve MOTR by ele- gantly incorporating an extra object detector. We first adopt the anchor formulation of queries and then use an extra ob- ject detector to generate proposals as anchors, providing detection prior to MOTR. The simple modification greatly eases the conflict between joint learning detection and asso- ciation tasks in MOTR. MOTRv2 keeps the query propoga- tion feature and scales well on large-scale benchmarks. MOTRv2 ranks the 1st place ( 73.4% HOTA on DanceTrack) in the 1st Multiple People Tracking in Group Dance Chal- lenge. Moreover, MOTRv2 reaches state-of-the-art perfor- mance on the BDD100K dataset. We hope this simple and effective pipeline can provide some new insights to the end- to-end MOT community. Code is available at https: //github.com/megvii-research/MOTRv2 .	1. Introduction Multi-object tracking (MOT) aims to predict the trajec- tories of all objects in the streaming video. It can be divided into two parts: detection and association. For a long time, the state-of-the-art performance on MOT has been domi- nated by tracking-by-detection methods [4, 36, 44, 45] with good detection performance to cope with various appear- ance distributions. These trackers [44] first employ an ob- ject detector (e.g., YOLOX [11]) to localize the objects in each frame and associate the tracks by ReID features or IoU matching. The superior performance of those methods par- tially results from the dataset and metrics biased towards detection performance. However, as revealed by the Dance- * The work was done during internship at MEGVII Technology and supported by National Key R&D Program of China (2020AAA0105200) and Beijing Academy of Artificial Intelligence (BAAI). DanceTrack BDD100K 54.273.473.583.7 32.343.644.856.5 HOTA mMOTA DetA mIDF1 Figure 1. Performance comparison between MOTR (grey bar) and MOTRv2 (orange bar) on the DanceTrack and BDD100K datasets. MOTRv2 improves the performance of MOTR by a large margin under different scenarios. Track dataset [27], their association strategy remains to be improved in complex motion. Recently, MOTR [43], a fully end-to-end framework is introduced for MOT. The association process is performed by updating the tracking queries while the new-born ob- jects are detected by the detect queries. Its association per- formance on DanceTrack is impressive while the detection results are inferior to those tracking-by-detection methods, especially on the MOT17 dataset. We attribute the infe- rior detection performance to the conflict between the joint detection and association processes. Since state-of-the-art trackers [6,9,44] tend to employ extra object detectors, one natural question is how to incorporate MOTR with an ex- tra object detector for better detection performance. One direct way is to perform IoU matching between the predic- tions of track queries and extra object detector (similar to TransTrack [28]). In our practice, it only brings marginal improvements in object detection while disobeying the end- to-end feature of MOTR. Inspired by tracking-by-detection methods that take the detection result as the input, we wonder if it is possible to feed the detection result as the input and reduce the learning of MOTR to the association. Recently, there are some ad- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22056 YOLOXMOTR YOLOXMOTR …Input @ 𝑡=0Prediction @ 𝑡=0Input @ 𝑡=1Prediction @ 𝑡=1… ProposalQTrackQProposalQTrackQFigure 2. The overall architecture of MOTRv2. The proposals produced by state-of-the-art detector YOLOX [11] are used to generate the proposal queries, which replaces the detect queries in MOTR [43] for detecting new-born objects. The track queries are transferred from previous frame and used to predict the bounding boxes for tracked objects. The concatenation of proposal queries and track queries as well as the image features are input to MOTR to generate the predictions frame-by-frame. vances [18, 35] for anchor-based modeling in DETR. For example, DAB-DETR initializes object queries with the center points, height, and width of anchor boxes. Sim- ilar to them, we modify the initialization of both detect and track queries in MOTR. We replace the learnable po- sitional embedding (PE) of detect query in MOTR with the sine-cosine PE [30] of anchors, producing an anchor-based MOTR tracker. With such anchor-based modeling, propos- als generated by an extra object detector can serve as the anchor initialization of MOTR, providing local priors. The transformer decoder is used to predict the relative offsets w.r.t. the anchors, making the optimization of the detection task much easier. The proposed MOTRv2 brings many advantages com- pared to the original MOTR. It greatly benefits from the good detection performance introduced by the extra object detector. The detection task is implicitly decoupled from the MOTR framework, easing the conflict between the de- tection and association tasks in the shared transformer de- coder. MOTRv2 learns to track the instances across frames given the detection results from an extra detector. MOTRv2 achieves large performance improvements on the DanceTrack, BDD100K, and MOT17 datasets com- pared to the original MOTR (see Fig. 1). On the DanceTrack dataset, MOTRv2 surpasses the tracking-by- detection counterparts by a large margin ( 14.8% HOTA compared to OC-SORT [6]), and the AssA metric is 18.8% higher than the second-best method. On the large-scale multi-class BDD100K dataset [42], we achieved 43.6% mMOTA, which is 2.4% better than the previous best solu- tion Unicorn [41]. MOTRv2 also achieves state-of-the-art performance on the MOT17 dataset [15, 21]. We hope our simple and elegant design can serve as a strong baseline forfuture end-to-end multi-object tracking research.
Ye_AccelIR_Task-Aware_Image_Compression_for_Accelerating_Neural_Restoration_CVPR_2023	Abstract Recently, deep neural networks have been successfully applied for image restoration (IR) (e.g., super-resolution, de-noising, de-blurring). Despite their promising perfor- mance, running IR networks requires heavy computation. A large body of work has been devoted to addressing this issue by designing novel neural networks or pruning their parameters. However, the common limitation is that while images are saved in a compressed format before being en- hanced by IR, prior work does not consider the impact of compression on the IR quality. In this paper, we present AccelIR, a framework that optimizes image compression considering the end-to-end pipeline of IR tasks. AccelIR encodes an image through IR-aware compression that optimizes compression levels across image blocks within an image according to the im- pact on the IR quality. Then, it runs a lightweight IR net- work on the compressed image, effectively reducing IR com- putation, while maintaining the same IR quality and image size. Our extensive evaluation using nine IR networks shows that AccelIR can reduce the computing overhead of super- resolution, de-nosing, and de-blurring by 49%, 29%, and 32% on average, respectively.	1. Introduction Image restoration (IR) is a class of techniques that re- covers a high-quality image from a lower-quality counter- part (e.g., super-resolution, de-noising, de-blurring). With the advances of deep learning, IR has been widely deployed in various applications such as satellite/medical image en- hancement [9,42,43,51,52], facial recognition [7,44,65,69], and video streaming/analytics [15, 27, 66, 67, 70]. Mean- while, the resolution of images used in these applica- tions has been rapidly increasing along with the evolution in client devices (e.g., smartphones [58, 61], TV moni- tors [57]). Thus, deep neural networks (DNNs) used for IR need to support higher-resolution images such as 4K (4096×2160) and even 8K (7680 ×4320). However, because the computing and memory overheadJPEG + IR AccelIRTaskFLOPs PSNR FLOPs PSNR Super-resolution 1165G 25.28dB 298G 25.30dB De-noising 1701G 30.49dB 1132G 30.70dB De-blurring 2590G 30.57dB 1718G 30.58dB Table 1. Computing overhead and quality of IR under the same compression ratio (1.2bpp). AccelIR reduces computation by 34- 74% while providing the same IR quality and image size. grow quadratically to the input resolution, applying IR net- works to such a large image is computationally expen- sive [32, 34]. Prior work addresses this issue using three different approaches: 1) designing efficient feature extrac- tion or up-scaling layers [2, 16, 29, 33, 59, 75], 2) adjusting network complexities within an image according to the IR difficulty [32, 34], and 3) pruning network parameters con- sidering their importance [50]. The common limitation in the prior studies is that they do not consider the detrimental impact of image compression on the IR quality, despite the fact that images in real-world applications are commonly saved in a compressed format before being enhanced by IR. In this work, we observe that there is a large opportunity in optimizing image compression considering the end-to- endpipeline of IR tasks. Compression loss has a signifi- cant impact on the IR quality, while its impact also greatly varies according to the image content, even within the same image. Such heterogeneity offers room for IR-aware image compression that optimizes compression levels across im- age blocks within an image according to the impact on the IR quality. IR-awareness allows us to use a lighter-weight IR network because the quality enhancement from IR-aware image compression can compensate the quality loss due to the reduced network capacity. Based on this observation, we present AccelIR, the first IR-aware image compression framework that considers the end-to-end pipeline of IR tasks, including image compres- sion. AccelIR aims to reduce IR computation while main- taining the same IR quality and image size. To enable this, AccelIR develops a practical IR-aware compression algo- rithm and adopts a lightweight IR network. AccelIR oper- ates in two phases: offline profiling and online compression. In the offline phase, AccelIR clusters image blocks in the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18216 representative datasets [1, 21] into groups. For each group, it constructs profiles that describe the impact of compres- sion level on the resulting IR quality and image size. In ad- dition to the profiles, a lightweight CNN is trained to guide the best-fit group for unseen image blocks. In the online phase, AccelIR retrieves the profiles for each block within an image by running the CNN. Our framework then refers to the block-level profiles to select the optimal compression level for each block, maximizing the IR quality at the same image size. Finally, the lightweight IR network is applied. We evaluate AccelIR using a full system implementation using JPEG [53] and WebP [68], the most widely used im- age compression standards. As shown in Table 1, our evalu- ation using five different super-resolution [2, 16, 36, 37, 56], two de-noising [71, 73], and two de-blurring networks [12, 72] shows that AccelIR consistently delivers a significant benefit in a wide range of settings. Compared to applying IR to images encoded by the standard JPEG and WebP, Ac- celIR reduces the computing cost of super-resolution, de- noising, and de-blurring by 35-74%, 24-34%, and 24-34%, respectively, under the same IR quality and image size. In addition, AccelIR can support any type of image codec and is well-fit to serve new IR tasks and networks that are not shown in the training phase. Thus, AccelIR can be easily integrated with the existing IR applications.
Yang_BEVFormer_v2_Adapting_Modern_Image_Backbones_to_Birds-Eye-View_Recognition_via_CVPR_2023	Abstract We present a novel bird’s-eye-view (BEV) detector with perspective supervision, which converges faster and bet- ter suits modern image backbones. Existing state-of-the- art BEV detectors are often tied to certain depth pre- trained backbones like VoVNet, hindering the synergy be- tween booming image backbones and BEV detectors. To address this limitation, we prioritize easing the optimization of BEV detectors by introducing perspective view supervi- sion. To this end, we propose a two-stage BEV detector, where proposals from the perspective head are fed into the bird’s-eye-view head for final predictions. To evaluate the effectiveness of our model, we conduct extensive ablation studies focusing on the form of supervision and the gener- ality of the proposed detector. The proposed method is ver- ified with a wide spectrum of traditional and modern image backbones and achieves new SoTA results on the large-scale nuScenes dataset. The code shall be released soon.	1. Introduction Bird’s-eye-view(BEV) recognition models [17,21,25,27, 29, 35, 42] are a class of camera-based models for 3D ob- ject detection. They have attracted interest in autonomous driving as they can naturally integrate partial raw observa- tions from multiple sensors into a unified holistic 3D out- put space. A typical BEV model is built upon an image backbone, followed by a view transformation module that lifts perspective image features into BEV features, which are further processed by a BEV feature encoder and some task-specific heads. Although much effort is put into de- signing the view transformation module [17, 27, 42] and in- corporating an ever-growing list of downstream tasks [9,27] into the new recognition framework, the study of image *: Equal contribution. B: Corresponding author, email: [email protected] in BEV models receives far less attention. As a cutting-edge and highly demanding field, it is natural to in- troduce modern image backbones into autonomous driving. Surprisingly, the research community chooses to stick with V oVNet [13] to enjoy its large-scale depth pre-training [26]. In this work, we focus on unleashing the full power of mod- ern image feature extractors for BEV recognition to unlock the door for future researchers to explore better image back- bone design in this field. However, simply employing modern image backbones without proper pre-training fails to yield satisfactory results. For instance, an ImageNet [6] pre-trained ConvNeXt-XL [23] backbone performs just on par with a DDAD-15M pre- trained V oVNet-99 [26] for 3D object detection, albeit the latter has 3.5 ×parameters of the former. We owe the strug- gle of adapting modern image backbones to the following issues: 1) The domain gap between natural images and au- tonomous driving scenes. Backbones pre-trained on general 2D recognition tasks fall short of perceiving 3D scenes, es- pecially estimating depth. 2) The complex structure of cur- rent BEV detectors. Take BEVFormer [17] as an example. The supervision signals of 3D bounding boxes and object class labels are separated from the image backbone by the view encoder and the object decoder, each of which is com- prised of transformers of multiple layers. The gradient flow for adapting general 2D image backbones for autonomous driving tasks is distorted by the stacked transformer layers. In order to combat the difficulties mentioned above in adapting modern image backbones for BEV recognition, we introduce perspective supervision into BEVFormer, i.e. ex- tra supervision signals from perspective-view tasks and di- rectly applied to the backbone. It guides the backbone to learn 3D knowledge missing in 2D recognition tasks and overcomes the complexity of BEV detectors, greatly facili- tating the optimization of the model. Specifically, we build a perspective 3D detection head [26] upon the backbone, which takes image features as input and directly predicts This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17830 the 3D bounding boxes and class labels of target objects. The loss of this perspective head, denoted as perspective loss, is added to the original loss (BEV loss) deriving from the BEV head as an auxiliary detection loss. The two de- tection heads are jointly trained with their corresponding loss terms. Furthermore, we find it natural to combine the two detection heads into a two-stage BEV detector, BEV- Former v2 . Since the perspective head is full-fledged, it could generate high-quality object proposals in the perspec- tive view, which we use as first-stage proposals. We encode them into object queries and gather them with the learn- able ones in the original BEVFormer, forming hybrid object queries, which are then fed into the second-stage detection head to generate the final predictions. We conduct extensive experiments to confirm the effec- tiveness and necessity of our proposed perspective super- vision. The perspective loss facilitates the adaptation of the image backbone, resulting in improved detection per- formance and faster model convergence. While without this supervision, the model cannot achieve comparable re- sults even if trained with a longer schedule. Consequently, we successfully adapt modern image backbones to the BEV model, achieving 63.4% NDS on nuScenes [2] test-set. Our contributions can be summarized as follows: • We point out that perspective supervision is key to adapting general 2D image backbones to the BEV model. We add this supervision explicitly by a detec- tion loss in the perspective view. • We present a novel two-stage BEV detector, BEV- Former v2. It consists of a perspective 3D and a BEV detection head, and the proposals of the former are combined with the object queries of the latter. • We highlight the effectiveness of our approach by com- bining it with the latest developed image backbones and achieving significant improvements over previous state-of-the-art results on the nuScenes dataset.
Xu_MEDIC_Remove_Model_Backdoors_via_Importance_Driven_Cloning_CVPR_2023	Abstract We develop a novel method to remove injected backdoors in deep learning models. It works by cloning the benign behaviors of a trojaned model to a new model of the same structure. It trains the clone model from scratch on a very small subset of samples and aims to minimize a cloning loss that denotes the differences between the activations of im- portant neurons across the two models. The set of important neurons varies for each input, depending on their magni- tude of activations and their impact on the classification result. We theoretically show our method can better recover benign functions of the backdoor model. Meanwhile, we prove our method can be more effective in removing back- doors compared with fine-tuning. Our experiments show that our technique can effectively remove nine different types of backdoors with minor benign accuracy degradation, outper- forming the state-of-the-art backdoor removal techniques that are based on fine-tuning, knowledge distillation, and neuron pruning.1	1. Introduction Backdoor attack is a prominent threat to applications of Deep Learning models. Misbehaviors, e.g., model mis- classification, can be induced by an input stamped with a backdoor trigger , such as a patch with a solid color or a pattern [14, 29]. A large number of various backdoor attacks have been proposed, targeting different modalities and fea- turing different attack methods [2, 22, 37, 44, 49, 55]. An important defense method is hence to remove backdoors from pre-trained models. For instance, Fine-prune proposed to remove backdoor by fine-tuning a pre-trained model on clean inputs [27]. Distillation [23] removes neurons that are not critical for benign functionalities. Model connectiv- ity repair (MCR) [56] interpolates weight parameters of a poisoned model and its finetuned version. ANP [50] adver- 1Code is available at https://github.com/qiulingxu/ MEDICsarially perturbs neuron weights of a poisoned model and prunes those neurons that are sensitive to perturbations (and hence considered compromised). While these techniques are very effective in their targeted scenarios, they may fall short in some other attacks. For example, if a trojaned model is substantially robust, fine-tuning may not be able to remove the backdoor without lengthy retraining. Distillation may be- come less effective if a neuron is important for both normal functionalities and backdoor behaviors. And pruning neu- rons may degrade model benign accuracy. More discussions of related work can be found in Section 2. In this paper, we propose a novel backdoor removal tech- nique MEDIC2as illustrated in Figure 1 (A). It works by cloning the benign functionalities of a pre-trained model to a new sanitized model of the same structure . Given a trojaned model and a small set of clean samples , MEDIC trains the clone model from scratch. The training is not only driven by the cross-entropy loss as in normal model training, but also by forcing the clone model to generate the same internal activation values as the original model at the correspond- ing internal neurons , i.e., steps 1⃝,2⃝, and 4⃝in Figure 1 (A). Intuitively, one can consider it essentially derives the weight parameters in the clone model by resolving the acti- vation equivalence constraints. There are a large number of such constraints even with just a small set of clean samples. However, such faithful cloning likely copies the backdoor behaviors as well as it tends to generate the same set of weight values. Therefore, our cloning is further guided by importance (step 3⃝). Specifically, for each sample x, we only force the important neurons to have the same activation values. A neuron is considered important if (1) it tends to be substantially activated by the sample, when compared to its activation statistics over the entire population (the activation criterion in red in Figure 1 (A)), and (2) the large activation value has substantial impact on the classification result (the impact criterion ) . The latter can be determined by analyz- ing the output gradient regarding the neuron activation. By constraining only the important neurons and relaxing the 2MEDIC stands for “Remove ModEl Backdoors via Importance DrIven Cloning ”. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20485 others, the model focuses on cloning the behaviors related to the clean inputs and precludes the backdoor. The set of weight values derived by such cloning are largely different from those in the original model. Intuitively, it is like solving the same set of variables with only a subset of equivalence constraints. The solution tends to differ substantially from that by solving the full set of constraints. Example. Figure 2 shows an example by visualizing the in- ternal activations and importance values for a trojaned model on a public benchmark [35]. Image (a) shows a right-turn traffic sign stamped with a polygon trigger in yellow, which induces the classification result of stop sign. Image (c) visu- alizes the activation values for the trojaned image whereas (d) shows those for its clean version. The differences be- tween the two, visualized in (e), show that the neurons fall into the red box are critical to the backdoor behaviors. A faithful cloning method would copy the behaviors for these neurons (and hence the backdoor). In contrast, our method modulates the cloning process using importance values and hence precludes the backdoor. The last image shows the importance values with the bright points denoting important neurons and the dark ones unimportant. Observe that it pre- vents copying the behaviors of the compromised neurons for this example as they are unimportant. One may notice that the unimportant compromised neurons for this example may become important for another example. Our method naturally handles this using per-sample importance values. □ Our technique is different from fine-tuning as it trains from scratch. It is also different from knowledge distilla- tion [25] shown in Figure 1 (B), which aims to copy be- haviors across different model structures and constrains log- its equivalence (and sometimes internal activation equiva- lence [53] as well). In contrast, by using the same model structure, we have a one-to-one mapping for individual neu- rons in the two models such that we can enforce strong constraints on internal equivalence. This allows us to copy behaviors with only a small set of clean samples. We evaluate our technique on nine different kinds of backdoor attacks. We compare with five latest backdoor removal methods (details in related work). The results show our method can reduce the attack success rate to 8.5% on average with only 2% benign accuracy degradation. It con- sistently outperforms the other baselines by 25%. It usually takes 15 mins to sanitize a model. We have also conducted an adaptive attack in which the trigger is composed of ex- isting benign features in clean samples. It denotes the most adversary context for MEDIC . Our ablation study shows all the design choices are critical. For example, without using importance values, it can only reduce ASR to 36% on average. In summary, our main contributions include the follow- ing.•We propose a novel importance driven cloning method to remove backdoors. It only requires a small set of clean samples. •We theoretically analyze the advantages of the cloning method. •We empirically show that MEDIC outperforms the state- of-the-art methods.
Yu_Phase-Shifting_Coder_Predicting_Accurate_Orientation_in_Oriented_Object_Detection_CVPR_2023	Abstract With the vigorous development of computer vision, ori- ented object detection has gradually been featured. In this paper, a novel differentiable angle coder named phase- shifting coder (PSC) is proposed to accurately predict the orientation of objects, along with a dual-frequency version (PSCD). By mapping the rotational periodicity of differ- ent cycles into the phase of different frequencies, we pro- vide a unified framework for various periodic fuzzy prob- lems caused by rotational symmetry in oriented object de- tection. Upon such a framework, common problems in oriented object detection such as boundary discontinuity and square-like problems are elegantly solved in a unified form. Visual analysis and experiments on three datasets prove the effectiveness and the potentiality of our approach. When facing scenarios requiring high-quality bounding boxes, the proposed methods are expected to give a com- petitive performance. The codes are publicly available at https://github.com/open-mmlab/mmrotate.	1. Introduction As a fundamental task in computer vision, object de- tection has been extensively studied. Early researches are mainly focused on horizontal object detection [33], on the ground that objects in natural scenes are usually oriented upward due to gravity. However, in other domains such as aerial images [2,12,20,23,25], scene text [7,10,13,14,34], and industrial inspection [11, 19], oriented bounding boxes are considered more preferable. Upon requirements in these scenarios, oriented object detection has gradually been fea- tured. At present, several solutions around oriented object de- *Corresponding author is Feipeng Da. This work is supported by Spe- cial Project on Basic Research of Frontier Leading Technology of Jiangsu Province of China (BK20192004C).tection have been developed, among which the most intu- itive way is to modify horizontal object detectors by adding an output channel to predict the orientation angle. Such a solution faces two problems: 1) Boundary problem [24]: Boundary discontinuity problem is often caused by angular periodicity. Assuming orientation −π/2is equivalent to π/2, the network output is sometimes expected to be −π/2and sometimes π/2when facing the same input. Such a situation makes the network confused about in which way it should perform regression. 2) Square-like problem [27]: Square-like problem usu- ally occurs when a square bounding box cannot be uniquely defined. Specifically, a square box should be equivalent to a90◦rotated one, but the regression loss between them is high due to the inconsistency of angle parameters. Such ambiguity can also seriously confuse the network. A more comprehensive introduction to these problems can be found in previous researches [25, 27]. Also, sev- eral methods have been proposed to address these problems, which will be reviewed in Sec. 2. Through rethinking the above problems, we find that they can inherently be unified as rotationally symmetric problems (boundary under 180◦and square-like under 90◦ rotation), which is quite similar to the periodic fuzzy prob- lem of the absolute phase acquisition [37] in optical mea- surement. Inspired by this, we come up with an idea to utilize phase-shifting coding, a technique widely used in optical measurement [36], for angle prediction in oriented object detection. The technique has the potential to solve both boundary discontinuity and square-like problems: 1) Phase-shifting encodes the measured distance (or par- allax) into the periodic phase in optical measurement. The orientation angle can also be encoded into the periodic phase, and boundary discontinuity is thus inherently solved. 2) Phase-shifting also has the periodic fuzzy problem, which is similar to the square-like problem, and many solu- tions exist. For example, the dual-frequency phase-shifting This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13354 technique solves the periodic fuzzy problem by mixing phases of different frequencies (also known as phase un- wrapping [37]). Motivations of this paper: Based on the above analysis, we believe that the phase- shifting technique can be modified and adapted to oriented object detection. What is the principle of the phase-shifting angle coder? How to integrate this module into a deep neural network? Will this technique result in better perfor- mance? These questions are what this paper is for. Contributions of this paper: 1) We are the first to utilize the phase-shifting coder to cope with the angle regression problem in the deep learning area. An integral and stable solution is elaborated on in this paper. Most importantly, the codes are well-written, publicly available, and with reproducible results. 2) The performance of the proposed methods is evalu- ated through extensive experiments. The experimental re- sults are of high quality—All the listed results are retested on identical environments to ensure fair comparisons (in- stead of copied from other papers). The rest of this paper is organized as follows: Section 2 reviews the related methods around oriented object detection. Section 3 describes the principles of the phase-shifting coder in detail. Section 4 conducts exper- iments on several datasets to evaluate the performance of the proposed methods. Section 5 concludes the paper.
Zhang_VQACL_A_Novel_Visual_Question_Answering_Continual_Learning_Setting_CVPR_2023	Abstract Research on continual learning has recently led to a va- riety of work in unimodal community, however little atten- tion has been paid to multimodal tasks like visual question answering (VQA). In this paper, we establish a novel VQA Continual Learning setting named VQACL, which contains two key components: a dual-level task sequence where vi- sual and linguistic data are nested, and a novel composi- tion testing containing new skill-concept combinations. The former devotes to simulating the ever-changing multimodal datastream in real world and the latter aims at measuring models’ generalizability for cognitive reasoning. Based on our VQACL, we perform in-depth evaluations of ﬁve well- established continual learning methods, and observe that they suffer from catastrophic forgetting and have weak gen- eralizability. To address above issues, we propose a novel representation learning method, which leverages a sample- speciﬁc and a sample-invariant feature to learn represen- tations that are both discriminative and generalizable for VQA. Furthermore, by respectively extracting such repre- sentation for visual and textual input, our method can ex- plicitly disentangle the skill and concept. Extensive exper- imental results illustrate that our method signiﬁcantly out- performs existing models, demonstrating the effectiveness and compositionality of the proposed approach. The code is available at https://github.com/zhangxi1997/VQACL.	1. Introduction Continual learning [43] has recently gained a lot of at- tention in the deep learning community because it enables models to learn continually on a sequence of non-stationary tasks and is close to the human learning process [2, 36]. However, the vibrant research in continual learning mainly focuses on unimodal tasks such as image classiﬁcation [37, 46, 51] and sequence tagging [4, 48], and the demand of multimodal tasks is ignored. In recent years, the volume of multimodal data has grown tremendously [8, 56, 57]. For example, tens of millions of texts, images, and videos are uploaded to social media platforms every day, such as Face- Figure 1. The illustration of real-world scenario for VQA system, which may continuously receive new types of questions, fresh visual concepts, and novel skill-concept compositions. book and Twitter. To cope with such constantly emerging real-world data, a practical AI system should be capable of continually learning from multimodal sources while allevi- ating forgetting previously learned knowledge. Visual Question Answering (VQA) is a typical multi- modal task and has drawn increasing interest over the past few years [12, 49, 60], which can automatically generate a textual answer given a question and an image. To deal with ever-changing questions and visual scenes in real life, applying continual learning to VQA is essential. However, it is not easy to set up a suitable continual learning setting for this task. We identify that two vital issues need to be considered. First, the VQA input comes from both vision and linguistic modalities, thus the task setting should si- multaneously tackle continuous data from both modalities in a holistic manner. For example, as shown in Fig. 1, the AI system might deal with new types of questions (e.g., Where ... ,Why ... ) as well as fresh visual concepts (e.g., Lo- quat,Deer ). Second, compositionality [24], a vital property of cognitive reasoning, should be considered in the VQA continual learning. The compositionality here denotes the model’s generalization ability towards novel combinations ofreasoning skills (i.e., question type) and visual concepts (i.e., image object). As illustrated in Fig. 1, if the system has been trained with question type Count (e.g., How many ) with a variety of objects (e.g., Person, Cat, and Dount ), as well as another question type (e.g., What color ) about a new object (e.g., Truck ). Then, it is expected to answer a novel This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19102 question like ‘How many trucks are there?’ , even if the composition of skill Count and concept Truck has yet to be seen. Such ability is very crucial when deploying a model in the real world because it is infeasible to view all possible skill-concept compositions. Remarkably, several works has addressed continual learning with VQA [14, 16, 25]. How- ever, they still apply a classic unimodal-like continual learn- ing setting for the task by devising a set of VQA tasks sim- ply based on question type or image scene, which ignores above two crucial issues: handling continuous multimodal data simultaneously and testing model’s compositionality. To achieve these two keypoints, in this paper, we propose a novel generative VQA continual learning setting named VQACL based on two well-known datasets: VQA v2 [13] and NExT-QA [49]. Speciﬁcally, as shown in Fig. 2(a), our VQACL setting consists of a dual-level task sequence. In the outer level, we set up a sequence of linguistic-driven tasks to evaluate models’ ability for the ever-changing ques- tion types. Moreover, to process the continuously shifted visual contents, for each outer level task, we further con- struct a series of randomly ordered visual-driven subtasks according to image object categories in the inner level. Such dual-level setting is similar to the cognitive process of chil- dren, who master a skill by trying it on various objects. For example, when learning to recognize colors, a baby usually asks all the things surrounding him ‘ what color ’ they are. Besides, to evaluate models’ compositionality, we construct a novel composition split. As shown in Fig 2(b), we remove a visual-driven subtask from each task in the outer level during training and utilize it for testing. In this way, the testing data contain novel skill-concept combinations that are not seen at the training time. In conclusion, on the one hand, our VQACL setting requires models to perform effective multimodal knowledge transfer from old tasks to new tasks while mitigating catastrophic forgetting [31]. On the other hand, the model should be capable of generalizing to novel compositions for cognitive reasoning. Using the proposed VQACL setting, we establish an initial set of baselines by adapting several well-known and state-of-the-art continual learning methods [1, 3, 7, 22, 45] from image classiﬁcation to the generative VQA tasks. The baselines are implemented on an advanced vision-and- language transformer [9] without pre-training. After bench- marking these baseline models, we ﬁnd that few of them can do well in the novel composition testing, which limits their wide applications in practice. To enhance the model’s compositionality, it is critical to learn an excellent repre- sentation that is discriminative for seen skills or concepts, and is generalizable to novel skill-concept compositions. To achieve it, recent static VQA methods [27, 47, 59] always ﬁrst learn joint representations for visual and textual inputs, and then utilize contrastive learning to implicitly disentan- gle the skill and concept within the joint feature. How-ever, such implicit disentangling makes existing models still dogged by the interference between the skill and concept, leading to suboptimal generalization results. Moreover, the complex contrastive sample building process makes these works tough to be applied to continual learning. Inspired by above discussions, we propose a novel repre- sentation learning method for VQACL, which introduces a sample-speciﬁc (SS) and a sample-invariant (SI) feature to learn better representations that are both discriminative and generalizable. To explicitly decouple the reasoning skills and visual concepts, we learn the SS and SI representation for visual and textual input separately. Speciﬁcally, the SS feature for each modality is learned through a trans- former encoder that stacks multiple self-attention layers, which can encode the most attractive and salient contents into the SS feature to make it discriminative. For the SI feature, we resort to prototype learning to aggregate the object class or question type information into it. Because the category knowledge is stable and representative across different scenarios, the SI feature can possess strong gen- eralizability. Besides, to ﬁt the continual learning setting, we constantly update the SI feature in training. In this way, it can capture new typical knowledge while retaining his- torical experience, helping alleviate the forgetting problem. In conclusion, combining the SS and SI features, we can obtain the representation that is conducive to the model’s compositional discriminability and generalizability. In summary, the major contributions of our work are threefold: (1) We introduce a new continual learning set- ting VQACL to simulate real-world generative VQA. It can not only simultaneously tackle the continuous data from vision and linguistic modality, but also test models’ com- positionality for cognitive reasoning. (2) We propose a sim- ple but effective representation learning method for contin- ual VQA, which novelly deploys a discriminative sample- speciﬁc feature and a generalizable sample-invariant feature to alleviate forgetting and enhance the models’ composition ability. (3) We re-purpose and evaluate ﬁve well-established methods on our VQACL, and observe that they struggle to obtain satisfactory results. Remarkably, our model con- sistently achieves the best performance, demonstrating the effectiveness and compositionality of our approach.
Yang_Behavioral_Analysis_of_Vision-and-Language_Navigation_Agents_CVPR_2023	Abstract To be successful, Vision-and-Language Navigation (VLN) agents must be able to ground instructions to ac- tions based on their surroundings. In this work, we develop a methodology to study agent behavior on a skill-specific basis – examining how well existing agents ground instruc- tions about stopping, turning, and moving towards speci- fied objects or rooms. Our approach is based on gener- ating skill-specific interventions and measuring changes in agent predictions. We present a detailed case study analyz- ing the behavior of a recent agent and then compare multi- ple agents in terms of skill-specific competency scores. This analysis suggests that biases from training have lasting ef- fects on agent behavior and that existing models are able to ground simple referring expressions. Our comparisons be- tween models show that skill-specific scores correlate with improvements in overall VLN task performance.	1. Introduction Following navigation instructions requires coordinating observations and actions in accordance with the natural lan- guage. Stopping when told to stop. Turning when told to turn. And appropriately grounding referring expressions when an action is conditioned on some aspect of the envi- ronment. All three of these example are required when fol- lowing the instruction “Turn left then go down the hallway until you see a desk. Walk towards the desk and then stop. ” In this work, we examine how well current instruction- following agents can execute different types of these sub- behaviors which we will refer to as skills . We situate our study in the popular Vision-and-Language Navigation (VLN) paradigm [2]. In a VLN episode, an agent is spawned in a never-before-seen environment and must navigate to a goal location specified by a natural language navigation instruction. An agent’s instruction- following capabilities are typically measured at the episode level – examining whether an agent reaches near the goal (success rate [2]), how efficiently it does so (SPL [1]), or how well its trajectory matches the ground truth path whichthe human-generated instruction was based on (nDTW [15]). These metrics are useful for comparing agents in ag- gregate, but take a perspective that has little to say about an agent’s fine grained competencies or what sub-instructions it is able to ground appropriately. In this work, we design an experimental paradigm based on controlled interventions to analyze fine-grained agent be- haviors. We focus our study on an agent’s ability to ex- ecute unconditional instructions like stopping or turning, as well as, conditional instructions that require more vi- sual grounding like moving towards specified objects and rooms. Our approach leverages annotations from RxR [13] to produce truncated trajectory-instruction pairs that can then be augmented with an additional skill-specific sub- instruction. We carefully filter these trajectories and gen- erate template-based sub-instructions to build non-trivial intervention episodes that evaluate an agent’s ability to ground skill-specific language to the appropriate actions. To demonstrate the value of this approach, we present a case study analyzing the behavior of a contemporary VLN model [6]. While we find evidence that the model can reli- ably ground some skill-specific language, our analysis also reveals that its errors are not random. But rather, they re- flect a systematic bias towards forward actions learned dur- ing training. For object- or room-seeking skills, we find only modest relationships between instructions and agent actions. Finally, we derive aggregate skill-specific scores and compare across VLN models with different overall task performance. We find that higher skill-specific scores cor- relate with higher task performance; however, not all skills share the same scale of improvement between weaker and stronger VLN models – suggesting that improvements in VLN may be fueled by some skills more than others. Contributions. To summarize this work, we: - Develop an intervention-based behavioral analysis paradigm for evaluating the behavior of VLN agents,1 - Provide a case study on a contemporary VLN agent [6], evaluating fine-grained competencies and biases, and - Examine the relationships between skill-specific metrics 1https://github.com/Yoark/vln-behave This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2574 and overall VLN task performance.
Yin_Gloss_Attention_for_Gloss-Free_Sign_Language_Translation_CVPR_2023	Abstract Most sign language translation (SLT) methods to date require the use of gloss annotations to provide additional supervision information, however, the acquisition of gloss is not easy. To solve this problem, we first perform an anal- ysis of existing models to confirm how gloss annotations make SLT easier. We find that it can provide two aspects of information for the model, 1) it can help the model im- plicitly learn the location of semantic boundaries in contin- uous sign language videos, 2) it can help the model under- stand the sign language video globally. We then propose gloss attention , which enables the model to keep its atten- tion within video segments that have the same semantics lo- cally, just as gloss helps existing models do. Furthermore, we transfer the knowledge of sentence-to-sentence similar- ity from the natural language model to our gloss atten- tion SLT network (GASLT) to help it understand sign lan- guage videos at the sentence level. Experimental results on multiple large-scale sign language datasets show that our proposed GASLT model significantly outperforms existing methods. Our code is provided in https://github. com/YinAoXiong/GASLT .	1. Introduction Sign languages are the primary means of communication for an estimated 466 million deaf and hard-of-hearing peo- ple worldwide [52]. Sign language translation (SLT), a so- cially important technology, aims to convert sign language videos into natural language sentences, making it easier for deaf and hard-of-hearing people to communicate with hear- ing people. However, the grammatical differences between sign language and natural language [5, 55] and the unclear semantic boundaries in sign language videos make it diffi- cult to establish a mapping relationship between these two kinds of sequences. Existing SLT methods can be divided into three cate- gories, 1) two-stage gloss-supervised methods, 2) end-to- *Both authors contributed equally to this research. †Corresponding author. 0 10 20 30 40 500 10 20 30 40 50 0.050.100.150.200.250.30(a) self-attention +gloss-supervised 0 10 20 30 40 500 10 20 30 40 50 0.020.040.060.080.100.120.140.16(b) self-attention +gloss-free 0 10 20 30 40 500 10 20 30 40 50 0.20.40.60.8(c) gloss-attention +gloss-free Figure 1. Visualization of the attention map in the shallow encoder layer of three different SLT models. As shown in (a), an essential role of gloss is to provide alignment information for the model so that it can focus on relatively more important local areas. As shown in (b), the traditional attention calculation method is diffi- cult to converge to the correct position after losing the supervision signal of the gloss. However, our proposed method (c) can still flexibly maintain the attention in important regions (just like (a)) due to the injection of inductive bias, which can partially replace the role played by gloss. end gloss-supervised methods, and 3) end-to-end gloss-free methods. The first two approaches rely on gloss annota- tions, chronologically labeled sign language words, to assist the model in learning alignment and semantic information. However, the acquisition of gloss is expensive and cumber- some, as its labeling takes a lot of time for sign language ex- perts to complete [5]. Therefore, more and more researchers have recently started to turn their attention to the end-to- end gloss-free approach [42, 51]. It learns directly to trans- late sign language videos into natural language sentences without the assistance of glosses, which makes the approach more general while making it possible to utilize a broader range of sign language resources. The gloss attention SLT network (GASLT) proposed in this paper is a gloss-free SLT method, which improves the performance of the model and removes the dependence of the model on gloss supervision by injecting inductive bias into the model and transferring knowledge from a powerful natural language model. A sign language video corresponding to a natural lan- guage sentence usually consists of many video clips with complete independent semantics, corresponding one-to-one with gloss annotations in the semantic and temporal order. Gloss can provide two aspects of information for the model. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2551 On the one hand, it can implicitly help the model learn the location of semantic boundaries in continuous sign lan- guage videos. On the other hand, it can help the model understand the sign language video globally. In this paper, the GASLT model we designed obtain information on these two aspects from other channels to achieve the effect of replacing gloss. First, we observe that the semantics of sign language videos are temporally local- ized, which means that adjacent frames have a high prob- ability of belonging to the same semantic unit. The visu- alization results in Figure 1a and the quantitative analysis results in Table 1 support this view. Inspired by this, we design a new dynamic attention mechanism called gloss at- tention to inject inductive bias [49] into the model so that it tends to pay attention to the content in the local same se- mantic unit rather than others. Specifically, we first limit the number of frames that each frame can pay attention to, and set its initial attention frame to frames around it so that the model can be biased to focus on locally closer frames. However, the attention mechanism designed in this way is static and not flexible enough to handle the information at the semantic boundary well. We then calculate an offset for each attention position according to the input query so that the position of the model’s attention can be dynamically ad- justed on the original basis. It can be seen that, as shown in Figure 1c, our model can still focus on the really important places like Figure 1a after losing the assistance of gloss. In contrast, as shown in Figure 1b, the original method fails to converge to the correct position after losing the supervision signal provided by the gloss. Second, to enable the model to understand the semantics of sign language videos at the sentence level and disam- biguate local sign language segments, we transfer knowl- edge from language models trained with rich natural lan- guage resources to our model. Considering that there is a one-to-one semantic correspondence between natural lan- guage sentences and sign language videos. We can in- directly obtain the similarity relationships between sign language videos by inputting natural language sentences into language models such as sentence bert [56]. Us- ing this similarity knowledge, we can enable the model to understand the semantics of sign language videos as a whole, which can partially replace the second aspect of the information provided by gloss. Experimental results on three datasets RWTH-PHOENIX-WEATHER-2014T (PHOENIX14T) [5], CSL-Daily [70] and SP-10 [66] show that the translation performance of the GASLT model ex- ceeds the existing state of the art methods, which proves the effectiveness of our proposed method. We also conduct quantitative analysis and ablation experiments to verify the accuracy of our proposed ideas and the effectiveness of our model approach. To summarize, the contributions of this work are as fol-lows: • We analyze the role of gloss annotations in sign lan- guage translation. • We design a novel attention mechanism and knowl- edge transfer method to replace the role of gloss in sign language translation partially. • Extensive experiments on three datasets show the ef- fectiveness of our proposed method. A broad range of new baseline results can guide future research in this field.
Zhang_Towards_Efficient_Use_of_Multi-Scale_Features_in_Transformer-Based_Object_Detectors_CVPR_2023	Abstract Multi-scale features have been proven highly effective for object detection but often come with huge and even prohibitive extra computation costs, especially for the re- cent Transformer-based detectors. In this paper, we pro- pose Iterative Multi-scale Feature Aggregation (IMFA) – a generic paradigm that enables efficient use of multi-scale features in Transformer-based object detectors. The core idea is to exploit sparse multi-scale features from just a few crucial locations, and it is achieved with two novel de- signs. First, IMFA rearranges the Transformer encoder- decoder pipeline so that the encoded features can be iter- atively updated based on the detection predictions. Second, IMFA sparsely samples scale-adaptive features for refined detection from just a few keypoint locations under the guid- ance of prior detection predictions. As a result, the sam- pled multi-scale features are sparse yet still highly ben- eficial for object detection. Extensive experiments show that the proposed IMFA boosts the performance of multiple Transformer-based object detectors significantly yet with only slight computational overhead.	1. Introduction Detecting objects of vastly different scales has always been a major challenge in object detection [28]. Fortu- nately, strong evidence [11, 22, 25, 48, 69, 72] shows that object detectors can significantly benefit from multi-scale features while dealing with large scale variation. For ConvNet-based object detectors like Faster R-CNN [42] and FCOS [49], Feature Pyramid Network (FPN) [25] and its variants [12, 18, 19, 30, 48, 69, 70] have become the go-to components for exploiting multi-scale features. Other than ConvNet-based object detectors, the recently proposed DEtection TRansformer (DETR) [4] has estab- lished a fully end-to-end object detection paradigm with *marks corresponding author.†marks equal technical contribution. Project Page: https://github.com/ZhangGongjie/IMFA. 80 100 120 140 160 180 200 GFLOPs4142434445COCO val2017 AP (%)SMCA-DETR(MS)Deform-DETR(MS) Cond-DETR(SS)Cond-DETR(DC)Cond-DETR + IMFA (Ours) Anchor-DETR(SS)Anchor-DETR(DC)Anchor-DETR + IMFA (Ours) DAB-DETR(SS)DAB-DETR(DC)DAB-DETR + IMFA (Ours) Single-Scale (SS) Dilated-Conv (DC) Multi-Scale (MS) Multi-Scale w/ our IMFAFigure 1. The proposed Iterative Multi-scale Feature Aggregation (IMFA) is a generic approach for efficient use of multi-scale fea- tures in Transformer-based object detectors. It boosts detection accuracy on multiple object detectors at minimal costs of addi- tional computational overhead. Results are obtained with ResNet- 50. Best viewed in color. promising performance. However, naively incorporating multi-scale features using FPN in these Transformer-based detectors [4,11,20,29,35,55,66,72] often brings enormous and even unfeasible computation costs, primarily due to the poor efficiency of the attention mechanism in processing high-resolution features. Concretely, to handle a feature map with a spatial size of H×W, ConvNet requires a com- putational cost of O(HW), while the complexity of the at- tention mechanism in Transformer-based object detectors is O(H2W2). To mitigate this issue, Deformable DETR [72] and Sparse DETR [43] replace the original global dense attention with sparse attention. SMCA-DETR [11] re- stricts most Transformer encoder layers to be scale-specific, with only one encoder layer to integrate multi-scale fea- tures. However, as the number of tokens increases quadrati- cally w.r.t. feature map size (typically 20x ∼80x compared to single-scale), these methods are still costly in computation and memory consumption, and rely on special operators like This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6206 deformable attention [72] that introduces extra complexity for deployment. To the best of our knowledge, there is yet no generic approach that can efficiently exploit multi-scale features for Transformer-based object detectors. In this paper, we present Iterative Multi-scale Feature Aggregation (IMFA) , a concise and effective technique that can serve as a generic paradigm for efficient use of multi- scale features in Transformer-based object detectors. The motivation comes from two key observations: (i)the com- putation of high-resolution features is highly redundant as the background usually occupies most of the image space, thus only a small portion of high-resolution features are useful to object detection; (ii)unlike ConvNet, the Trans- former’s attention mechanism does not require grid-shaped feature maps, which offers the feasibility of aggregating multi-scale features only from some specific regions that are likely to contain objects of interest. The two observa- tions motivate us to sparsely sample multi-scale features from just a few informative locations and then aggregate them with encoded image features in an iterative manner. Concretely, IMFA consists of two novel designs in the Transformer-based detection pipelines. First , IMFA rear- ranges the encoder-decoder pipeline so that each encoder layer is immediately connected to its corresponding de- coder layer. This design enables iterative update of en- coded image features along with refined detection predic- tions. Second , IMFA sparsely samples multi-scale features from the feature pyramid generated by the backbone, with the sampling process guided by previous detection predic- tions. Specifically, motivated by the spatial redundancy of high-resolution features, IMFA only focuses on a few promising regions with high likelihood of object occurrence based on prior predictions. Furthermore, inspired by the significance of objects’ keypoints for recognition and lo- calization [39, 59, 66, 71], IMFA first searches several key- points within each promising region, and then samples use- ful features around these keypoints at adaptively selected scales. The sampled features are finally fed to the subse- quent encoder layer along with the image features encoded by the previous layer. With the two new designs, the pro- posed IMFA aggregates only the most crucial multi-scale features from those informative locations. Since the number of the aggregated features is small, IMFA introduces min- imal computational overhead while consistently improving the detection performance of Transformer-based object de- tectors. It is noteworthy that IMFA is a generic paradigm for efficient use of multi-scale features: (i)as shown in Fig. 1, it can be easily integrated with multiple Transformer-based object detectors with consistent performance boosts; (ii)as discussed in Section 5.4, IMFA has the potential to boost DETR-like models on tasks beyond object detection. To summarize, the contributions of this work are threefold. • We propose a novel DETR-based detection pipeline,where encoded features can be iteratively updated along with refined detection predictions. This new pipeline al- lows to leverage intermediate predictions as guidance for robust and efficient multi-scale feature encoding. • We propose a sparse sampling strategy for multi-scale features, which first identifies several promising regions under the guidance of prior detections, then searches sev- eral keypoints within each promising region, and finally samples their features at adaptively selected scales. We demonstrate that such sparse multi-scale features can sig- nificantly benefit object detection. • Based on the two contributions above, we propose Iter- ative Multi-scale Feature Aggregation (IMFA) – a sim- ple and generic paradigm that enables efficient use of multi-scale features in Transformer-based object detec- tors. IMFA consistently boosts detection performance on multiple object detectors, yet remains computationally efficient. This is the pioneering work that investigates a generic approach for exploiting multi-scale features effi- ciently in Transformer-based object detectors.
Yi_A_Simple_Framework_for_Text-Supervised_Semantic_Segmentation_CVPR_2023	Abstract Text-supervised semantic segmentation is a novel re- search topic that allows semantic segments to emerge with image-text contrasting. However, pioneering methods could be subject to speciﬁcally designed network architectures. This paper shows that a vanilla contrastive language-image pre-training (CLIP) model is an effective text-supervised se- mantic segmentor by itself. First, we reveal that a vanilla CLIP is inferior to localization and segmentation due to its optimization being driven by densely aligning visual and language representations. Second, we propose the locality-driven alignment (LoDA) to address the problem, where CLIP optimization is driven by sparsely aligning lo- cal representations. Third, we propose a simple segmenta- tion (SimSeg) framework. LoDA and SimSeg jointly amelio- rate a vanilla CLIP to produce impressive semantic segmen- tation results. Our method outperforms previous state-of- the-art methods on PASCAL VOC 2012, PASCAL Context and COCO datasets by large margins. Code and models are available at github.com/muyangyi/SimSeg .	1. Introduction Semantic segmentation is a fundamental task in com- puter vision, with the purpose of allocating semantic classes to the corresponding pixels. Most existing methods for se- mantic segmentation are restricted by the scale of datasets. The quantity or category is insufﬁcient due to the high cost of annotating segmentation masks. Text-supervised seman- tic segmentation makes a breakthrough for this challenge, where models are pre-trained with image-text pairs and zero-shot transferred to semantic segmentation. Figure 1illustrates an abstraction of text-supervised semantic segmentation in comparison with existing task paradigms. The base domain is denoted as DB, which contains the manually labeled samples. The target do- * Corresponding authors. †Work was done when the ﬁrst author was an intern at ByteDance.!!!"seg mapsOpen-Vocabulary!!!#!"Zero-ShotWeakly SupervisedText-Supervised (Ours) image-text pairs + seg annotations GroupViTTextEncoderimage-text pairsseg mapsnon-universal backboneImageEncoderTextEncoder ImageEncoderTextEncoderimage-text pairsseg maps GroupViTLSeg SimSeg (Ours)fine-tune moduleimage-text pairs + seg annotationsImageEncoderTextEncoderseg maps OpenSegfine-tune module!!!#!!!" Figure 1. A comparison of our proposed approach with existing paradigms, where DB,DT,DOdenote base domain, target do- main and open domain, respectively. The components in red are those missing in SimSeg. Illustration inspired by [ 50]. main is denoted as DT, which contains test samples. And open domain DOinvolves a large variety of linguistic in- formation. It can provide additional textual descriptions when segmenting the images. Open-vocabulary meth- ods ( e.g., LSeg [ 22], OpenSeg [ 17]) use pre-trained vision- and-language models [ 20,33], but still need annotated sam- ples to ﬁne-tune. Weakly supervised methods [ 1,2] are free from mask labels but require image-level class la- bels ( DT✓D O). Text-supervision is an annotation-free scheme, eliminating the need for mask annotations ( DB) or image-level labels ( i.e.,DT*DO). Text-supervision lever- ages massive web image-text pairs and enables to generate segmentation masks in a zero-shot manner. GroupViT [ 44] is the ﬁrst work of text-supervision, yet the non-universal backbone design hinders its ﬂexibility ( e.g., novel backbone This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7071 adaptation and multi-task joint learning). We could improve current methods by creating a simple framework for text- supervised semantic segmentation. To this end, we target on the vanilla CLIP [ 33] architecture, a neat dual-stream contrastive language-image pre-training model. As the preliminary of this work, we explore the potential problems of a vanilla CLIP-based segmentor. We mainly study CLIP developed with Transformer-based encoders due to their intrinsic properties for segmentation [ 6] and su- perior performance. CLIP is originally driven by aligning vision and textual holistic vectors ( e.g.,[cls]tokens from Transformer-based encoders), and a simple revision facili- tates CLIP models for segmentation, i.e., densely aligning all image patches and caption words. A similarity map, which describes correlations between all image patches and one class word , is a coarse categorical segmentation mask per se . However, we observe two problems that greatly sup- press the ability of the CLIP-based segmentor: (1) Visual encoder of the learned CLIP model focuses on contextual pixels, and (2) image-text contrasting mainly relies on con- textual words. These problems jointly reveal that the opti- mization of CLIP is signiﬁcantly driven by contextual in- formation. As a consequence, the CLIP-based segmentor yields poor semantic segmentation results, due to an infe- rior ability to perceive both contextual and non-contextual information in complex natural images. In the following, we attempt to solve the above prob- lems. One practical strategy is avoiding optimization with contextual information. For a versatile segmentor, both con- textual and non-contextual information are essential. Con- textual and non-contextual pixels should be sparsely aligned to corresponding text entities. To this end, we propose a locality-driven alignment (LoDA) strategy for training CLIP models. Firstly, we propose to select partial features with the maximum responses, in both image and text modal- ities. Secondly, we propose to drive the image-text con- trasting with only selected features. Our proposal success- fully solves the problems from two aspects: (1) Vision en- coder perceives main objects, (2) main objects and context are equally signiﬁcant in the image-text contrasting. Cou- pled with LoDA, a simple but effective framework named SimSeg is proposed to do zero-shot semantic segmentation. Beneﬁting from our proposals, a simple CLIP framework is equipped with impressive zero-shot semantic segmentation performances. Our contributions are three-fold: •We reveal the problems of a vanilla CLIP attached with Transformer-based encoders when producing segmenta- tion masks. To solve the problems, we propose a training strategy named locality-driven alignment (LoDA). •We design a simple but effective text-driven zero-shot semantic segmentation framework named SimSeg. Our proposed LoDA and SimSeg jointly allow a simple CLIP to segment universal categories.•We achieve remarkable improvements over previous methods on PASCAL VOC, PASCAL Context and COCO zero-shot segmentation tasks. Moreover, we pro- vide extensive analyses and ablations of our proposals.
Yu_Range-Nullspace_Video_Frame_Interpolation_With_Focalized_Motion_Estimation_CVPR_2023	Abstract Continuous-time video frame interpolation is a funda- mental technique in computer vision for its flexibility in synthesizing motion trajectories and novel video frames at arbitrary intermediate time steps. Yet, how to infer accu- rate intermediate motion and synthesize high-quality video frames are two critical challenges. In this paper, we present a novel VFI framework with improved treatment for these challenges. To address the former, we propose focalized trajectory fitting, which performs confidence-aware motion trajectory estimation by learning to pay focus to reliable optical flow candidates while suppressing the outliers. The second is range-nullspace synthesis, a novel frame renderer cast as solving an ill-posed problem addressed by learning decoupled components in orthogonal subspaces. The pro- posed framework sets new records on 7 of 10 public VFI benchmarks.	1. Introduction Continuous-time Video Frame Interpolation (VFI) aims at upsampling the temporal resolution of low frame-rate videos steplessly by synthesizing the missing frames at ar- bitrary time steps. It is a fundamental technology for vari- ous downstream video applications such as streaming [34], stabilization [5], and compression [35]. A key challenge of this task is to find a continuous map- ping from arbitrary time step to the latent scene motion to correctly render the target frame, observing the low frame- rate video. Typically, it was realized via two stages: motion trajectory fitting and frame synthesis, e.g.[1,3,4,10,13,17, 21, 29, 36, 38, 39]. In the former, a parametric trajectory model is fitted from the optical flows extracted from input frames, which can be resampled at any time step to get in- termediate motion. For representing such a motion model, ∗This work is done during Zhiyang’s internship at SenseTime. Correspondence should be addressed to Yu Zhang ([email protected]) and Shunqing Ren ([email protected]). 𝑰0(a) Consecutive input frames and estimated optical flows (c) The proposed pipeline𝑰−1𝑰−1𝑰−1 𝑰2𝑰2𝑰2 𝑰1𝑰1𝑰1 𝑰0𝑰0𝑰0 Reconstruction Reconstruction Deep Range -Nullspace Solver (b) Conventional pipelineRendering Randering Rays𝑰0 𝑰𝑡 Flow Confidence 𝑰1𝑰𝑡 Resample 𝑭0→𝑡 Resample 𝑭1→𝑡𝑰1 Resample 𝑭0→𝑡 Resample 𝑭1→𝑡 ? 𝑰0 𝑰−1 𝑰1𝑰2 𝑾−1𝑾1𝑾2Figure 1. Motivation of the proposed method. Given input frames and extracted optical flows (a), common VFI pipeline (b) is to fit a continuous motion trajectory model, resample flows at the target time step, based on which rendering rays are predicted to synthe- size the intermediate frame by reorganizing pixels from the input frames. Our pipeline (c) proposes improved treatment in two as- pects: 1) our focalized motion estimation assigns dynamic weights to the extracted optical flows to suppress the outlier flows and im- prove fitting accuracy, and the novel range-nullspace solver treats intermediate frame synthesis as an inverse problem instead of di- rect rendering. Please see text for more details. the Taylor polynomial is often adopted with different or- ders [4,13,36]. However, fitting trajectory models are prone to optical flow errors, which are inevitable due to occlusion. Besides motion fitting, the frame synthesis step faces the challenges of correctly inferring scene geometry. Particu- larly, the intermediate flows resampled from the continuous motion model only depict partial correspondences between the intermediate frame with input frames, while existing methods predict complete rendering rays in a data-driven manner to facilitate full rendering, as shown in Fig. 1 (b). This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22159 Though powerful architectures [15] and rendering mecha- nisms [10] were proposed, it requires the network to fully encode the scene geometry to predict correct rays, which may raise the difficulty of architecture design and the bur- den of learning. To overcome these issues, we propose a novel framework with improved motion fitting and frame synthesis compo- nents. As shown in Fig. 1 (c), the first is focalized trajectory fitting, which extracts a set of optical flow candidates from the input video and assigns learned confidence weights to them. Confidence-aware, differentiable trajectory fitting is then followed, focalizing only the confident flows and pro- ducing improved parametric motion models. Such models, when resampled at the target time step, can produce accu- rate flows even at the occlusion boundary. For the latter, we follow the fact that intermediate frame synthesis is an ill-posed task and propose a deep solver based on the range nullspace theory [2]. Our solver decomposes the latent in- termediate frame into several orthogonal image components and adopts different networks to learn each of them. Such physics-guided design encourages learning decoupled sub- tasks for each network, relieving the burden of encoding scene geometry and benefiting the use of lightweight mod- els. In particular, our framework sets new records on 7 out of 10 public VFI benchmarks while being parameter- efficient. We summarize the contributions of this paper as follows. 1) A novel lightweight VFI framework is proposed, which refreshes the records of 7 out of 10 public VFI benchmarks. 2) The idea of focalized trajectory fitting, which improves parametric motion estimation in VFI and generates better resampling quality of intermediate flows. 3) A new perspec- tive that treats intermediate frame synthesis as an ill-posed problem, solved with a deep range nullspace solver that de- couples frame synthesis into several orthogonal tasks.
Yang_UniSim_A_Neural_Closed-Loop_Sensor_Simulator_CVPR_2023	Abstract Rigorously testing autonomy systems is essential for making safe self-driving vehicles (SDV) a reality. It requires one to generate safety critical scenarios beyond what can be collected safely in the world, as many scenarios happen rarely on our roads. To accurately evaluate performance, we need to test the SDV on these scenarios in closed-loop, where the SDV and other actors interact with each other at each timestep. Previously recorded driving logs provide *Indicates equal contribution.a rich resource to build these new scenarios from, but for closed loop evaluation, we need to modify the sensor data based on the new scene configuration and the SDV’s deci- sions, as actors might be added or removed and the tra- jectories of existing actors and the SDV will differ from the original log. In this paper, we present UniSim, a neural sensor simulator that takes a single recorded log captured by a sensor-equipped vehicle and converts it into a realistic closed-loop multi-sensor simulation. UniSim builds neural feature grids to reconstruct both the static background and dynamic actors in the scene, and composites them together This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1389 to simulate LiDAR and camera data at new viewpoints, with actors added or removed and at new placements. To better handle extrapolated views, we incorporate learnable priors for dynamic objects, and leverage a convolutional network to complete unseen regions. Our experiments show UniSim can simulate realistic sensor data with small domain gap on downstream tasks. With UniSim, we demonstrate, for the first time, closed-loop evaluation of an autonomy system on safety-critical scenarios as if it were in the real world.	1. Introduction While driving along a highway, a car from the left sud- denly swerves into your lane. You brake hard, avoiding an accident, but discomforting your passengers. As you replay the encounter in your mind, you consider how the scenario would have gone if the other vehicle had acceler- ated more, if you had slowed down earlier, or if you had instead changed lanes for a more comfortable drive. Hav- ing the ability to generate such “what-if” scenarios from a single recording would be a game changer for developing safe self-driving solutions. Unfortunately, such a tool does not exist and the self-driving industry primarily test their systems on pre-recorded real-world sensor data (i.e., log- replay), or by driving new miles in the real-world. In the former, the autonomous system cannot execute actions and observe their effects as new sensor data different from the original recording is not generated, while the latter is neither safe, nor scalable or sustainable. The status quo calls for novel closed-loop sensor simulation systems that are high fidelity and represent the diversity of the real world. Here, we aim to build an editable digital twin of the real world (through the logs we captured), where existing ac- tors in the scene can be modified or removed, new actors can be added, and new autonomy trajectories can be exe- cuted. This enables the autonomy system to interact with the simulated world, where it receives new sensor obser- vations based on its new location and the updated states of the dynamic actors, in a closed-loop fashion. Such a simulator can accurately measure self-driving performance, as if it were actually in the real world, but without the safety hazards, and in a much less capital-intensive manner. Compared to manually-created game-engine based virtual worlds [15, 62], it is a more scalable, cost-effective, realis- tic, and diverse way towards closed-loop evaluation. Towards this goal, we present UniSim, a realistic closed- loop data-driven sensor simulation system for self-driving. UniSim reconstructs and renders multi-sensor data for novel views and new scene configurations from a single recorded log. This setting is very challenging as the observations are sparse and often captured from constrained viewpoints ( e.g., straight trajectories along the roads). To better handle ex- trapolation from the observed views, we propose a series of enhancements over prior neural rendering approaches. Inparticular, we leverage multi-resolution voxel-based neural fields to represent and compose the static scene and dy- namic agents, and volume render feature maps. To better handle novel views and incorporate scene context to reduce artifacts, a convolutional network (CNN) renders the fea- ture map to form the final image. For dynamic agents, we learn a neural shape prior that helps complete the objects to render unseen areas. We use this sparse voxel-based rep- resentations to efficiently simulate both image and LiDAR observations under a unified framework. This is very useful as SDVs often use several sensor modalities for robustness. Our experiments show that UniSim realistically simu- lates camera and LiDAR observations at new views for large-scale dynamic driving scenes, achieving SoTA perfor- mance in photorealism. Moreover, we find UniSim reduces the domain gap over existing camera simulation methods on the downstream autonomy tasks of detection, motion fore- casting and motion planning. We also apply UniSim to aug- ment training data to improve perception models. Impor- tantly, we show, for the first time, closed-loop evaluation of an autonomy system on photorealistic safety-critical scenar- ios, allowing us to better measure SDV performance. This further demonstrates UniSim’s value in enabling safer and more efficient development of self-driving.
Yoshiyasu_Deformable_Mesh_Transformer_for_3D_Human_Mesh_Recovery_CVPR_2023	Abstract We present Deformable mesh trans Former (DeFormer), a novel vertex-based approach to monocular 3D human mesh recovery. DeFormer iteratively fits a body mesh model to an input image via a mesh alignment feedback loop formed within a transformer decoder that is equipped with efficient body mesh driven attention modules: 1) body sparse self-attention and 2) deformable mesh cross atten- tion. As a result, DeFormer can effectively exploit high- resolution image feature maps and a dense mesh model which were computationally expensive to deal with in pre- vious approaches using the standard transformer attention. Experimental results show that DeFormer achieves state-of- the-art performances on the Human3.6M and 3DPW bench- marks. Ablation study is also conducted to show the ef- fectiveness of the DeFormer model designs for leveraging multi-scale feature maps. Code is available at https: //github.com/yusukey03012/DeFormer .	1. Introduction Human mesh recovery from a single image is an im- portant problem in computer vision, with a wide range of applications like virtual reality, sports motion analysis and human-computer interactions. It is a challenging prob- lem as it requires modeling of complex nonlinear mappings from an image to 3D body shape and pose. In the past decade, remarkable progress has been accom- plished in the field of 3D human mesh recovery based on convolutional neural networks (CNNs) [41]. A common way used in this field is a parametric approach that employs statistical human models parameterized by shape and pose parameters [33]. Here, CNNs features are regressed with body shape and pose parameters with approximately 100 di- mensions. Then, a body mesh surface is obtained from these body parameter predictions by inputting them to the human body kinematics and statistical shape models. Leveraging multi-scale image feature maps and iteratively refining the body parameters based on global and local spatial contexts in an image, the pyramidal mesh alignment feedbacks (Py- Figure 1. Summary of this work. We propose DeFormer, a mem- ory efficient decoder-only transformer architecture for 3D human mesh recovery based on block sparse self-attention [16, 40, 51] and multi-scale (MS) deformable cross attention [56]. Leverag- ing MS feature maps efficiently in transformer, our method out- performs previous parametric approaches using MS feature maps [52, 53] and vertex transformer approaches using a single-scale feature [28, 29]. Further, by learning sparse self-attention patterns based on body mesh and skeleton connectivity, DeFormer can re- cover a dense mesh with 6.9K joint/vertex queries at interactive rates. The units of the MPJPE and PA-MPJPE scores are in mil- limeters. The lower is better. MAF) model [53] produces a 3D mesh recovery result well- aligned to an input image. Another paradigm for human mesh recovery is a vertex- based approach that directly regresses 3D vertex coordi- nates [14, 15, 26, 28, 29, 35]. Recently, transformer archi- tectures have been applied to vertex-based human mesh re- covery and show good reconstruction performances by cap- turing long-range interactions between body parts via self- attentions. Furthermore, as a vertex-based approach, they naturally possess potential in learning fine-grained effects like facial expressions, finger poses and clothing non-rigid deformations, if such supervisions are given. Despite their This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17006 promising performances, the main challenge of the current transformer-based approaches based on the standard trans- former attention is its memory cost, which limits the use of high-resolution image feature maps and a dense body mesh model within it for better 3D reconstruction. In fact, the cur- rent methods are limited to managing a single-scale coarse feature map with 7×7grids and a coarse mesh with around 400 vertices [28, 29]. In this paper, we propose a novel vertex-based trans- former approach to 3D human mesh recovery, named Deformable mesh trans Former (DeFormer). DeFormer it- eratively fits a human mesh model to an image using the mesh-alignment feedback loop formed in a transformer de- coder. In order to leverage multi-scale feature maps and a dense mesh model in the human mesh recovery trans- former model, we design a decoder network architecture us- ing block sparse attention and multi-scale (MS) deformable attention [56] as illustrated in Fig. 1. The block sparse self- attention module exploits the sparse connectivity patterns established from a human body mesh and its skeleton. This sparsifies self-attention access patterns and reduces mem- ory consumption. The MS deformable cross attention mod- ule aggregates multi-scale image features by attending only to a small set of key sampling points around the mesh ver- tex of the current reconstruction. DeFormer therefore can attend to visual feature maps in both coarse and fine levels, which contain not only global contexts but also local spa- tial information. It can then extract useful contextualized features for human mesh recovery and regress corrective displacements to refine the estimated mesh shape. Conse- quently, DeFormer can exploit high-resolution image fea- ture maps and a dense mesh model that are not accessible to previous approaches based on the standard transformer attention [14,28,29] due to time/memory computational de- mands. As a result, DeFormer achieves better performance than previous approaches in 3D human mesh recovery. The main contributions of this paper include: •DeFormer : A novel decoder-only transformer archi- tecture for monocular 3D human mesh recovery based on the new body-mesh-driven attention modules that leverage multi-scale visual features and a dense mesh reasonably efficiently. DeFormer achieves new SOTA results on the Human3.6M and 3DPW benchmarks. •Body Sparse Self-Attention that exploits the sparse connectivity patterns extracted from a human body mesh model and its skeleton to restrict self-attention access patterns, improving memory efficiency. •Deformable Mesh cross Attention (DMA) that effi- ciently aggregates and exploits multi-scale image fea- ture maps by attending only to a small set of key sam- pling points around the reference points obtained from the current body joint and mesh vertex reconstructions.
Yang_RILS_Masked_Visual_Reconstruction_in_Language_Semantic_Space_CVPR_2023	Abstract Both masked image modeling (MIM) and natural lan- guage supervision have facilitated the progress of trans- ferable visual pre-training. In this work, we seek the syn- ergy between two paradigms and study the emerging prop- erties when MIM meets natural language supervision. To this end, we present a novel masked visual Reconstruction In Language semantic Space (RILS) pre-training frame- work, in which sentence representations, encoded by the text encoder, serve as prototypes to transform the vision-only signals into patch-sentence probabilities as semantically meaningful MIM reconstruction targets. The vision models can therefore capture useful components with structured in- formation by predicting proper semantic of masked tokens. Better visual representations could, in turn, improve the text encoder via the image-text alignment objective, which is essential for the effective MIM target transformation. Ex- tensive experimental results demonstrate that our method not only enjoys the best of previous MIM and CLIP but also achieves further improvements on various tasks due to their mutual benefits. RILS exhibits advanced transferabil- ity on downstream classification, detection, and segmenta- tion, especially for low-shot regimes. Code is available at https://github.com/hustvl/RILS .	1. Introduction Learning transferable representation lies a crucial task in deep learning. Over the past few years, natural language processing (NLP) has achieved great success in this line of research [18, 45, 60]. To explore similar trajectories in the vision domain, researchers tend to draw upon the suc- cesses of NLP and have made tremendous progress: (i) Inspired by the advanced model architecture [60] as well as self-supervised learning paradigm [18] in NLP, vision Transformers (ViT) [21, 42] and masked image modeling *This work was done while S. Yang was an intern at ARC Lab, Tencent PCG.†X. Wang and Y . Ge are the corresponding authors. V-Enc V-Enc L-EncV-Dec ImageMM M M TextsContra//prediction targetRecon XXFigure 1. Overview of our RILS. Recon andContra represent masked reconstruction loss and image-text contrastive loss. Dur- ing pre-training, RILS learns to perform masked image modeling and image-text contrastive simultaneously. Masked predictions and corresponding targets are transformed to probabilistic distri- butions in language space by leveraging text features as semantic- rich prototypes. Under such a scheme, both objective are unified and achieve mutual benefits from each other. Vision encoder ob- tains the ability to capture meaningful and fine-grained context in- formation, sparking decent transfer capacity. (MIM) [3, 28] open a new era of self-supervised visual representation learning, and show unprecedented transfer- ability on various tasks, especially on fine-grained tasks such as object detection [22, 39]. (ii) Inspired by the great scalability brought by leveraging web-scale collections of texts as training data in NLP [5, 48, 49, 72], CLIP [47] and ALIGN [35] bring such a principle to vision and manifest the immense potential of leveraging natural language super- vision for scalable visual pre-training. Strong transferability of pre-trained visual models on low-shot regimes ensues. It also facilitates diverse applications by extracting contextu- alized image or text features [26, 50, 52]. The remarkable attainment achieved by these two re- search lines pushes us to ponder: Is it possible to unify both masked image modeling and natural language supervision to pursuit better visual pre-training? A straightforward way towards this goal is to simply combine masked image mod- eling (MIM) with image-text contrastive learning (ITC) for multi-task learning. Although the na ¨ıve combination is fea- sible to inherit the above advantages, we find it remains un- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23304 satisfactory due to the mutual benefits between MIM and ITC have not yet been fully explored. Motivated by this, we develop RILS, a tailored framework to seek the synergy of masked image modeling and language supervision. The core insight of RILS is to perform masked visual re- construction in language semantic space . Specifically, in- stead of reconstructing masked patches in the standalone vi- sion space ( e.g., raw pixel [28,66], low-level features [3,62] or high-level perceptions [12, 34, 63, 75]), we map patch features to a probabilistic distribution over a batch of text features as the reconstruction target, which is enabled by ITC that progressively aligns the image and text spaces. The text features serve as semantically rich prototypes and probabilistic distributions explicitly inject the semantic in- formation onto each image patch. The MIM objective is formulated as a soft cross-entropy loss to minimize the KL divergence between text-injected probabilistic distributions of masked vision tokens and their corresponding targets. The visual model optimized by our language-assisted re- construction objective, in turn, improves ITC with better vi- sual representations that capture fine-grained local contexts. Under such a working mechanism, the two objec- tives( i.e., MIM and ITC) complement each other and form a unified solution for transferable and scalable visual pre- training. Note that a lot of works [34, 63, 75] have mani- fested the importance of semantic information in the recon- struction target of MIM objectives. However, it is abstract to pursue such a semantically rich space with visual-only signals due to its unstructured characteristics [29]. Thanks to natural language supervision, this issue is alleviated by performing masked reconstruction in language space. Extensive experiments on various downstream tasks demonstrate that our design enjoys the best of both worlds. With a vanilla ViT-B/ 16as the vision model and25-epoch pre-training on 20million image-text pairs, RILS achieves 83.3%top-1accuracy when fine-tune on ImageNet- 1K [15], +1.2%and+0.6%better than the MAE [28] and CLIP [47] counterparts. Advanced perfor- mance can be consistently acquired when transfer to fine- grained tasks such as detection and segmentation. More- over, our approach exhibits promising out-of-the-box ca- pability under an extremely low-shot regime. RILS also demonstrates superior performance on zero-shot image classification and image-text retrieval. On ImageNet- 1K benchmark, RILS obtains 45.0%zero-shot classification ac- curacy, +4.7%/+ 3.4%higher than CLIP [47] /SLIP [43] under the same training recipe. Compelling results of RILS imply the promising capacity in the unification of MIM and language supervision.
Zhang_MP-Former_Mask-Piloted_Transformer_for_Image_Segmentation_CVPR_2023	Abstract We present a mask-piloted Transformer which improves masked-attention in Mask2Former for image segmenta- tion. The improvement is based on our observation that Mask2Former suffers from inconsistent mask predictions between consecutive decoder layers, which leads to incon- sistent optimization goals and low utilization of decoder queries. To address this problem, we propose a mask-piloted training approach, which additionally feeds noised ground- truth masks in masked-attention and trains the model to reconstruct the original ones. Compared with the predicted masks used in mask-attention, the ground-truth masks serve as a pilot and effectively alleviate the negative impact of inaccurate mask predictions in Mask2Former. Based on this technique, our MP-Former achieves a remarkable perfor- mance improvement on all three image segmentation tasks (instance, panoptic, and semantic), yielding +2.3AP and +1.6mIoU on the Cityscapes instance and semantic seg- mentation tasks with a ResNet-50 backbone. Our method also significantly speeds up the training, outperforming Mask2Former with half of the number of training epochs on ADE20K with both a ResNet-50 and a Swin-L backbones. Moreover, our method only introduces little computation dur- ing training and no extra computation during inference. Our code will be released at https://github.com/IDEA- Research/MP-Former .	1. Introduction Image segmentation is a fundamental problem in com- puter vision which includes semantic, instance, and panoptic *Equal contribution. †This work was done when Hao Zhang and Feng Li were interns at IDEA. ‡Corresponding author.segmentation. Early works design specialized architectures for different tasks, such as mask prediction models [2, 15] for instance segmentation and per-pixel prediction mod- els [25, 28] for semantic segmentation. Panoptic segmen- tation [18] is a unified task of instance and semantic seg- mentation. But universal models for panoptic segmentation such as Panoptic FPN [18] and K-net [34] are usually sub- optimal on instance and semantic segmentation compared with specialized models. Recently, Vision Transformers [29] has achieved tremen- dous success in many computer vision tasks, such as image classification [11, 24] and object detection [1]. Inspired by DETR’s set prediction mechanism, Mask2Former [7] pro- poses a unified framework and achieves a remarkable per- formance on all three segmentation tasks. Similar to DETR, it uses learnable queries in Transformer decoders to probe features and adopts bipartite matching to assign predictions to ground truth (GT) masks. Mask2Former also uses pre- dicted masks in a decoder layer as attention masks for the next decoder layer. The attention masks serve as a guidance to help next layer’s prediction and greatly ease training. Despite its great success, Mask2Former is still an initial attempt which is mainly based on a vanilla DETR model that has a sub-optimal performance compared to its later variants. For example, in each decoder layer, the predicted masks are built from scratch by dot-producting queries and feature map without performing layer-wise refinement as proposed in deformable DETR [37]. Because each layer build masks from scratch, the masks predicted by a query in different layers may change dramatically. To show this inconsistent predictions among layers, we build two metrics mIoU-LiandUtilito quantify this problem and analyze its consequences in Section 3. As a result, this problem leads to a low utilization rate of decoder queries, which is especially severe in the first few decoder layers. As the attention masks predicted from an early layer are usually inaccurate, when serving as a guidance for its next layer, they will lead to This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18074 sub-optimal predictions in the next layer. Since detection is a similar task to segmentation, both aiming to locate instances in images, we seek inspirations from detection to improve segmentation. For example, the recently proposed denoising training [19] for detection has shown very effective to stabilize bipartite matching between training epochs. DINO [33] achieves a new SOTA on COCO detection built upon an improved denoising training method. The key idea in denoising training is to feed noised GT boxes in parallel with learnable queries into Transformer decoders and train the model to denoise and recover the GT boxes. Inspired by DN-DETR [19], we propose a mask-piloted (MP) training approach. We divide the decoder queries into two parts, an MP part and a matching part. The matching part is the same as in Mask2Former, feeding learnable queries and adopting bipartite matching to assign predictions to GT instances. In the MP part, we feed class embeddings of GT categories as queries and GT masks as attention masks into a Transformer decoder layer and let the model to reconstruct GT masks and labels. To reinforce the effectiveness of our method, we propose to apply MP training for all decoder layers. Moreover, we also add point noises to GT masks and flipping noises to GT class embeddings to force the model to recover GT masks and labels from inaccurate ones. We summarize our contributions as follows. 1.We propose a mask-piloted training approach to im- proving masked-attention in Mask2Former, which suf- fers from inconsistent and inaccurate mask predictions among layers. We also develop techniques including multi-layer mask-piloted training with point noises and label-guided training to further improve the segmenta- tion performance. Our method is computationally ef- ficient, introducing no extra computation during infer- ence and little computation during training. 2.Through analyzing the predictions of Mask2Former, we discover a failure mode of Mask2Former—inconsistent predictions between consecutive layers. We build two novel metrics to measure this failure, which are layer- wise query utilization and mean IoU. We also show that our method improves the two metrics by a large margin, which validates its effectiveness on alleviating inconsistent predictions. 3.When evaluating on three datasets, including ADE20k, Cityscapes, and MS COCO, our MP-Former outper- forms Mask2Former by a large margin. Besides im- proved performance, our method also speeds up train- ing. Our model exceeds Mask2Former on all three seg- mentation tasks with only half of the number of train- ing steps on ADE20k. Also, our model trained for 36 epochs is comparable with Mask2Former trained for 50 epochs on instance and panoptic segmentation on MS COCO.
Yan_Universal_Instance_Perception_As_Object_Discovery_and_Retrieval_CVPR_2023	Abstract All instance perception tasks aim at finding certain ob- jects specified by some queries such as category names, lan- guage expressions, and target annotations, but this com- plete field has been split into multiple independent sub- tasks. In this work, we present a universal instance per- ception model of the next generation, termed UNINEXT . UNINEXT reformulates diverse instance perception tasks into a unified object discovery and retrieval paradigm and can flexibly perceive different types of objects by simply changing the input prompts. This unified formulation brings the following benefits: (1) enormous data from different tasks and label vocabularies can be exploited for jointly training general instance-level representations, which is es- pecially beneficial for tasks lacking in training data. (2) the unified model is parameter-efficient and can save re- dundant computation when handling multiple tasks simul- taneously. UNINEXT shows superior performance on 20 challenging benchmarks from 10 instance-level tasks in- cluding classical image-level tasks (object detection and instance segmentation), vision-and-language tasks (refer- ring expression comprehension and segmentation), and six video-level object tracking tasks. Code is available at https://github.com/MasterBin-IIAU/UNINEXT.	1. Introduction Object-centric understanding is one of the most essen- tial and challenging problems in computer vision. Over the years, the diversity of this field increases substantially. In this work, we mainly discuss 10 sub-tasks, distributed on the vertices of the cube shown in Figure 1. As the most fundamental tasks, object detection [8, 9, 30, 59, 82, 84, 91] and instance segmentation [6, 37, 64, 90, 97] require finding all objects of specific categories by boxes and masks re- spectively. Extending inputs from static images to dynamic *This work was performed while Bin Yan worked as an intern at ByteDance. Email: yan [email protected].†Corresponding authors: [email protected], [email protected]. Time ReferenceFormat CoarseFine NoneLanguage or AnnotationStaticDynamicObject Detection MOTS VIS RVOS VOSRES REC SOT MOTInstance Segmentation personskateboardpersonsheepsheepsheepsheepdogsheep a zebra to the left of the frame The fourth person from the left tennis player in whiteFigure 1. Task distribution on the Format-Time-Reference space. Better view on screen with zoom-in. videos, Multiple Object Tracking (MOT) [3, 74, 126, 128], Multi-Object Tracking and Segmentation (MOTS) [47, 93, 108], and Video Instance Segmentation (VIS) [43,100,103, 112] require finding all object trajectories of specific cate- gories in videos. Except for category names , some tasks provide other reference information. For example, Refer- ring Expression Comprehension (REC) [115,121,130], Re- ferring Expression Segmentation (RES) [116,119,121], and Referring Video Object Segmentation (R-VOS) [7, 86, 104] aim at finding objects matched with the given language ex- pressions like “The fourth person from the left”. Besides, Single Object Tracking (SOT) [5, 51, 106] and Video Ob- ject Segmentation (VOS) [18, 78, 107] take the target an- notations (boxes or masks) given in the first frame as the reference, requiring to predict the trajectories of the tracked objects in the subsequent frames. Since all the above tasks aim to perceive instances of certain properties, we refer to them collectively as instance perception . Although bringing convenience to specific applications, such diverse task definitions split the whole field into frag- mented pieces. As the result, most current instance percep- tion methods are developed for only a single or a part of sub-tasks and trained on data from specific domains. Such This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15325 fragmented design philosophy brings the following draw- backs: (1) Independent designs hinder models from learn- ing and sharing generic knowledge between different tasks and domains, causing redundant parameters. (2) The pos- sibility of mutual collaboration between different tasks is overlooked. For example, object detection data enables models to recognize common objects, which can naturally improve the performance of REC and RES. (3) Restricted by fixed-size classifiers, traditional object detectors are hard to jointly train on multiple datasets with different label vo- cabularies [36,61,88] and to dynamically change object cat- egories to detect during inference [22, 61, 74, 81, 112, 120]. Since essentially all instance perception tasks aim at find- ing certain objects according to some queries , it leads to a natural question: could we design a unified model to solve all mainstream instance perception tasks once and for all? To answer this question, we propose UNINEXT, a uni- versal instance perception model of the next generation. We first reorganize 10 instance perception tasks into three types according to the different input prompts: (1) cate- gory names as prompts (Object Detection, Instance Seg- mentation, VIS, MOT, MOTS). (2) language expressions as prompts (REC, RES, R-VOS). (3) reference annota- tions as prompts (SOT, VOS). Then we propose a uni- fied prompt-guided object discovery and retrieval formula- tion to solve all the above tasks. Specifically, UNINEXT first discovers Nobject proposals under the guidance of the prompts, then retrieves the final instances from the proposals according to the instance-prompt match- ing scores . Based on this new formulation, UNINEXT can flexibly perceive different instances by simply changing the input prompts. To deal with different prompt modalities, we adopt a prompt generation module, which consists of a reference text encoder and a reference visual encoder. Then an early fusion module is used to enhance the raw visual features of the current image and the prompt embeddings. This operation enables deep information exchange and pro- vides highly discriminative representations for the later in- stance prediction step. Considering the flexible query-to- instance fashion, we choose a Transformer-based object de- tector [131] as the instance decoder. Specifically, the de- coder first generates Ninstance proposals, then the prompt is used to retrieve matched objects from these proposals. This flexible retrieval mechanism overcomes the disadvan- tages of traditional fixed-size classifiers and enables joint training on data from different tasks and domains. With the unified model architecture, UNINEXT can learn strong generic representations on massive data from various tasks and solve 10 instance-level perception tasks using a single model with the same model parameters. Ex- tensive experiments demonstrate that UNINEXT achieves superior performance on 20 challenging benchmarks. The contributions of our work can be summarized as follows.• We propose a unified prompt-guided formulation for universal instance perception, reuniting previously fragmented instance-level sub-tasks into a whole. • Benefiting from the flexible object discovery and re- trieval paradigm, UNINEXT can train on different tasks and domains, in no need of task-specific heads. • UNINEXT achieves superior performance on 20 chal- lenging benchmarks from 10 instance perception tasks using a single model with the same model parameters.
Yao_Large-Scale_Training_Data_Search_for_Object_Re-Identification_CVPR_2023	Abstract We consider a scenario where we have access to the tar- get domain, but cannot afford on-the-fly training data an- notation, and instead would like to construct an alternative training set from a large-scale data pool such that a com- petitive model can be obtained. We propose a search and pruning (SnP) solution to this training data search prob- lem, tailored to object re-identification (re-ID), an appli- cation aiming to match the same object captured by differ- ent cameras. Specifically, the search stage identifies and merges clusters of source identities which exhibit similar distributions with the target domain. The second stage, subject to a budget, then selects identities and their im- ages from the Stage I output, to control the size of the re- sulting training set for efficient training. The two steps provide us with training sets 80% smaller than the source pool while achieving a similar or even higher re-ID accu- racy. These training sets are also shown to be superior to a few existing search methods such as random sampling and greedy sampling under the same budget on training data size. If we release the budget, training sets resulting from the first stage alone allow even higher re-ID accu- racy. We provide interesting discussions on the specificity of our method to the re-ID problem and particularly its role in bridging the re-ID domain gap. The code is available at https://github.com/yorkeyao/SnP	1. Introduction The success of a deep learning-based object re-ID relies on one of its critical prerequisites: the labeled training data. To achieve high accuracy, typically a massive amount of data needs to be used to train deep learning models. How- ever, creating large-scale object re-ID training data with manual labels is expensive. Furthermore, collecting training data that contributes to the high test accuracy of the trained model is even more challenging. Recent years have seen a large amount of datasets proposed and a significant increase in the data size of a single dataset. For example, the Rand- Person [37] dataset has 8000 identities, which is more than 1K 15K# identities 7K PersonXDukeMSMTSnP (Proposed) MarketSource pool Number of training samples (× 103)Rank-1 accuracy (%)Figure 1. We present a search and pruning (SnP) solution to the training data search problem in object re-ID. The source data pool is 1 order of magnitude larger than existing re-ID training sets in terms of the number of images and the number of identities. When the target is AlicePerson [1], from the source pool, our method (SnP) results in a training set 80% smaller than the source pool while achieving a similar or even higher re-ID accuracy. The searched training set is also superior to existing individual train- ing sets such as Market-1501 [54], Duke [55], and MSMT [45]. 6×larger than the previous PersonX [37] dataset. However, these datasets generally have their own dataset bias, making the model trained on one dataset unable to generalize well to another. For example, depending on the filming scenario, different person re-ID datasets generally have biases on camera positions, human races, and cloth- ing style. Such dataset biases usually lead to model bias, which results in the model’s difficulty performing well in an unseen filming scenario. To address this, many try to improve learning algorithms, including domain adaptation and domain generalization methods [2, 10, 27, 33, 35, 56]. Whereas these algorithms are well-studied and have proven successful in many re-ID applications, deciding what kind of data to use for training the re-ID model is still an open research problem, and has received relatively little attention in the community. We argue that this is a crucial problem This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15568 to be answered in light of the ever-increasing scale of the available datasets. In this paper, we introduce SnP, a search and pruning so- lution for sampling an efficient re-ID training set to a target domain. SnP is designed for the scenario in that we have a target dataset that does not have labeled training data. Our collected source pool, in replace, provides suitable labeled data to train a competitive model. Specifically, given a user- specified budget ( e.g., maximum desired data size), we sam- ple a subset of the source pool, which satisfies the budget re- quirement and desirably has high-quality data to train deep learning models. This scenario is especially helpful for de- ploying a re-ID system for unknown test environments, as it is difficult to manually label a training set for these new environments. We note that due to the absence of an in- distribution training set, the searched data are directly used for training the re-ID model rather than pre-training. In particular, we combine several popular re-ID datasets into a source pool, and represent each image in the pool with features. Those features are extracted from an Imagenet- pretrained model [39]. The images with features are stored on the dataserver to serve as a gallery. When there is a query from the target, we extract the feature of the query image, and search in the gallery for similar images on a feature level. Specifically, in the search stage, we calculate feature- level distance, i.e., Fréchet Inception Distance (FID) [16]. Given the constraint of a budget, we select the most repre- sentative samples in the pruning stage, based on the outputs from the search stage. This limits the size of the constructed training set and enables efficient training. Combining search and pruning, we construct a train- ing dataset that empirically shows significant accuracy im- provements on several object re-ID targets, compared to the baselines. Without budget constraints, our searched training sets allow higher re-ID accuracy than the complete source pool, due to its target-specificity. With budget constraints, the pruned training set still achieves comparable or better performance than the source pool. The proposed SnP is demonstrated to be superior to random or greedy sampling. We show in Fig. 1 that the training set constructed by SnP leads to the best performance on the target compared to the others. We provide discussions on the specificity of our method and its role in bridging the re-ID domain gap.
Zhang_Improving_the_Transferability_of_Adversarial_Samples_by_Path-Augmented_Method_CVPR_2023	Abstract Deep neural networks have achieved unprecedented suc- cess on diverse vision tasks. However, they are vulnerable to adversarial noise that is imperceptible to humans. This phe- nomenon negatively affects their deployment in real-world scenarios, especially security-related ones. To evaluate the robustness of a target model in practice, transfer-based at- tacks craft adversarial samples with a local model and have attracted increasing attention from researchers due to their high efficiency. The state-of-the-art transfer-based attacks are generally based on data augmentation, which typically augments multiple training images from a linear path when learning adversarial samples. However, such methods se- lected the image augmentation path heuristically and may augment images that are semantics-inconsistent with the target images, which harms the transferability of the gen- erated adversarial samples. To overcome the pitfall, we propose the Path-Augmented Method (PAM). Specifically, PAM first constructs a candidate augmentation path pool. It then settles the employed augmentation paths during ad- versarial sample generation with greedy search. Further- more, to avoid augmenting semantics-inconsistent images, we train a Semantics Predictor (SP) to constrain the length of the augmentation path. Extensive experiments confirm that PAM can achieve an improvement of over 4.8% on av- erage compared with the state-of-the-art baselines in terms of the attack success rates.	1. Introduction Deep neural networks (DNNs) appear to be the state- of-the-art solutions for a wide variety of vision tasks [21, *Corresponding author. Figure 1. Illustration of how SIM and our PAM augment images (red dots) during the generation of adversarial samples. SIM only considers one linear path from the target image Xto a baseline image X′. Besides, SIM may augment images that are semantics- inconsistent with the target image. In contrast, our PAM augments images along multiple augmentation paths. We also constrain the length of the path to avoid augmenting images that are semantics- inconsistent with the target one. 28]. However, DNNs are vulnerable to adversarial sam- ples [47], which are elaborately designed by adding human- imperceptible noise to the clean image to mislead DNNs into wrong predictions. The existence of adversarial sam- ples causes negative effects on security-sensitive DNN- based applications, such as self-driving [27] and medical diagnosis [6, 7]. Therefore, it is necessary to understand the DNNs [15, 32, 33, 40] and enhance attack algorithms to better identify the DNN model’s vulnerability, which is the first step to improve their robustness against adversarial samples [23, 24, 42]. There are generally two kinds of attacks in the litera- ture [9]. One is the white-box attacks, which consider the white-box setting where attackers can access the architec- tures and parameters of the victim models. The other is the black-box attacks, which focus on the black-box situ- ation where attackers fail to get access to the specifics of This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8173 the victim models [10, 39]. Black-box attacks are more ap- plicable than the white-box counterparts to real-world sys- tems. There are two basic black-box attack methodolo- gies: the query-based [1, 2, 37] and the transfer-based at- tacks [38, 41]. Query-based attacks interact with the victim model to generate adversarial samples, but they may incur excessive queries. In contrast, transfer-based attacks craft adversarial samples with a local source model and do not need to query the victim model. Therefore, transfer-based attacks have attracted more attention recently because of their high efficiency [10, 39]. However, transfer-based attacks generally craft adversar- ial samples by employing white-box strategies like the Fast Gradient Sign Method (FGSM) [11] to attack a local model, which often leads to limited transferability due to overfit- ting to the employed local model. Most existing solutions address the overfitting issue from the perspective of opti- mization and generalization, which regards the local model and the target image as the training data of the adversar- ial sample. Therefore, the transferability of the learned adversarial sample corresponds to its generalization abil- ity across attacking different models [20]. Such method- ologies to improve adversarial transferability can be cate- gorized into two groups. One is the optimizer-based ap- proaches [9, 20, 34, 36], which adopt more advanced opti- mizers to escape from poor local optima during the genera- tion of adversarial samples. The other is the augmentation- based methods [10,20,35,44], which resort to data augmen- tation and exploit multiple training images to learn a more transferable adversarial sample. Current state-of-the-art augmentation-based attacks gen- erally apply a heuristics-based augmentation method. For example, the Scale-Invariant attack Method (SIM) [20] aug- ments multiple scale copies of the target image, while Ad- mix [35] augments multiple scale copies of the mixtures of the target image and the images from other categories. SIM exponentially augments images along a linear path from the target image to a baseline image, which is the origin. Ad- mix, in contrast, first augments the target image with the mixture of the target image and the images from other cate- gories. Then it also exponentially augments images along a linear path from the mixture image to the origin. Therefore, such methods only consider the image augmentation path to one baseline image, i.e., the origin. Besides, although they attempt to augment images that are semantics-consistent to the target image [20, 35], they fail to constrain the length of the image augmentation path, which may result in augment- ing semantics-inconsistent images. To overcome the pitfalls of existing augmentation-based attacks, we propose a transfer-based attack called Path- Augmented Method (PAM). PAM proposes to augment im- ages from multiple image augmentation paths to improve the transferability of the learned adversarial sample. How-ever, due to the continuous space of images, the possible image augmentation paths starting from the target image are countless. In order to cope with the efficiency prob- lem, we first select representative path directions to con- struct a candidate augmentation path pool. Then we settle the employed augmentation paths during adversarial sam- ple generation with greedy search. Furthermore, to avoid augmenting semantics-inconsistent images, we train a Se- mantics Predictor, which is a lightweight neural network, to constrain the length of each augmentation path. The difference between our PAM and SIM is illustrated in Figure 1. During the generation of adversarial samples, PAM augments images along multiple image augmentation paths from the target image to different baseline images, while SIM only augments images along a single image aug- mentation path from the target image to the origin. Be- sides, PAM constrains the length of the image augmenta- tion path to avoid augmenting images that are far away from the target image and preserve the semantic meaning of the target image. In contrast, SIM may augment images that are semantics-inconsistent with the target image due to the overlong image augmentation path. To confirm the superiority of our PAM, we conduct ex- tensive experiments against both undefended and defended models on the ImageNet dataset. Experimental results show that our PAM can achieve an improvement of over 3.7% on average compared with the state-of-the-art baselines in terms of the attack success rates. We also evaluate the per- formance of the combination of PAM with other compatible attack methods. Again, experimental results confirm that our method can significantly outperform the state-of-the-art baselines by about 7.2% on average. In summary, our contributions in this paper are threefold: • We discover that the state-of-the-art augmentation- based attacks (SIM and Admix) actually augment training images from a linear path for learning adver- sarial samples. We argue that they suffer from limited and overlong augmentation paths. • To address their pitfalls, We propose the Path- Augmented Method (PAM). PAM augments images from multiple augmentation paths during the genera- tion of adversarial samples. Besides, to make the aug- mented images preserve the semantic meaning of the target image, we train a Semantics Predictor (SP) to constrain the length of each augmentation path. • We conduct extensive experiments to validate the ef- fectiveness of our methodologies. Experimental re- sults confirm that our approaches can outperform the state-of-the-art baselines by a margin of over 3.7% on average. Besides, when combined with other compati- ble strategies, our method can significantly surpass the state-of-the-art baselines by 7.2% on average. 8174
Xu_Where_Is_My_Wallet_Modeling_Object_Proposal_Sets_for_Egocentric_CVPR_2023	Abstract This paper deals with the problem of localizing objects in image and video datasets from visual exemplars. In par- ticular, we focus on the challenging problem of egocen- tric visual query localization. We first identify grave im- plicit biases in current query-conditioned model design and visual query datasets. Then, we directly tackle such bi- ases at both frame and object set levels. Concretely, our method solves these issues by expanding limited annota- tions and dynamically dropping object proposals during training. Additionally, we propose a novel transformer- based module that allows for object-proposal set context to be considered while incorporating query information. We name our module Conditioned Contextual Transformer or CocoFormer. Our experiments show the proposed adapta- tions improve egocentric query detection, leading to a better visual query localization system in both 2D and 3D configu- rations. Thus, we can improve frame-level detection perfor- mance from 26.28% to31.26% in AP , which correspond- ingly improves the VQ2D and VQ3D localization scores by significant margins. Our improved context-aware query object detector ranked first and second respectively in the VQ2D and VQ3D tasks in the 2nd Ego4D challenge. In ad- dition to this, we showcase the relevance of our proposed model in the Few-Shot Detection (FSD) task, where we also achieve SOTA results. Our code is available at https: //github.com/facebookresearch/vq2d_cvpr .	1. Introduction The task of Visual Queries Localization can be described as the question, ‘when was the last time that I saw X’, where X is an object query represented by a visual crop. In the Ego4D [24] setting, this task aims to retrieve objects from an ‘episodic memory’, supported by the recordings from an *Work done during an internship at Meta AI. visual queryvisual query 0.721 ->0.027 target object distractor query frameannotated trackmodeltrain /validate modelinference query framevisual query target objectdistractor easy moderate hard 0.954 ->0.7120.803 ->0.352Figure 1. Existing approaches to visual query localization are biased. Top: only positive frames containing query object are used to train and validate the query object detector, while the model is naturally tested on whole videos, where the query object is mostly absent. Training with background frames alleviates confu- sion caused by hard negatives during inference and helps predict fewer false positives in negative frames. Bottom : visual query, target object, and distractors of three query objects from easy to hard. The confidence scores in white from a biased model get sup- pressed in our model in cyan. Our method improves the visual query localization system with fewer false positives, even for the very challenging scenario shown in the last row. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2593 egocentric device, such as VR headsets or AR glasses. A real-world application of such functionality is localizing a user’s items via a pre-registered object-centered image of them. A functional visual query localization system will allow users to find their belongings by a short re-play, or via a 3D arrow pointing to their real-world localization. The current solutions to this problem [24, 64] rely on a so-called ‘Siam-detector’ that is trained on annotated response tracks but tested on whole video sequences, as shown in Fig. 1 (top, gray arrows). The Siam-detector model design allows the incorporation of query exemplars by independently comparing the query to all object propos- als. During inference on a given video, the visual query is fixed, and the detector runs over all the frames in the ego- centric video recording. Although existing methods offer promising results in query object detection performance, it still suffers from do- main andtask biases. The domain bias appears due to only training with frames with well-posed objects in clear view, while daily egocentric recordings are naturally out of fo- cus, blurry, and undergoing uncommon view angles. On the other hand, task bias refers to the issue of the query object always being available during training, while in re- ality, it is absent during most of the test time. Therefore, baseline models predict false positives on distractors, espe- cially when the query object is out of view. These issues are exacerbated by the current models’ independent scoring of each object proposal, as the baseline models learn to give high scores to superficially similar objects while disregard- ing other proposals to reassess those scores. In Fig. 1 (bottom), we show the visual query, target object, and distractors for three data samples from easy to hard. The confidence scores in white of the distract- ing objects are reported from a publicly-available ‘Siam- RCNN’ model [24]. Due to random viewpoints and the large number of possible object classes that are exhibited in egocentric recordings, the target object is hard to dis- cover and confused with high-confidence false positives. In this work, we show that these biases can be largely tack- led by training a conditioned contextual detector with aug- mented object proposal sets. Our detector is based on a hypernetwork architecture that allows incorporating open- world visual queries, and a transformer-based design that permits context from other proposals to take part in the de- tection reasoning. We call this model Conditioned Contex- tual Transformer or CocoFormer. CocoFormer has a condi- tional projection layer that generates a transformation ma- trix from the query. This transformation is then applied to the proposal features to create query-conditioned proposal embeddings. These query-aware proposal embeddings are then fed to a set-transformer [33], which effectively allows our model to utilize the global context of the correspond- ing frame. We show our proposed CocoFormer surpassesSiam-RCNN and is more flexible in different applications as a hyper-network, such as multimodal query object detec- tion and few-shot object detection (FSD). Our CocoFormer has the increased capability to model objects for visual query localization because it tackles the domain bias by incorporating conditional context, but it still suffers from task bias induced by the current training strat- egy. To alleviate this, we propose to sample proposal sets from both labeled and unlabeled video frames, which col- lects data closer to the real distribution. Concretely, we en- large the positive query-frame training pairs by the natural view-point change of the same object in the view, while we create negative query-frame training pairs that incorporate objects from background frames. Essentially, we sample the proposal set from those frames in a balanced manner. We collect all the objects in a frame as a set and decouple the task by two set-level questions: (1) does the query ob- ject exist? (2) what is the most similar object to the query? Since the training bias issue only exists in the first prob- lem, we sample positive/negative proposal sets to reduce it. Note that this is an independent process of frame sampling, as the domain bias impairs the understanding of objects in the egocentric view, while task bias implicitly hinders the precision of visual query detection. Overall, our experiments show diversifying the object query and proposal sets, and leveraging global scene infor- mation can evidently improve query object detection. For example, training with unlabeled frames can improve the detection AP from 27.55% to28.74%, while optimizing the conditional detector’s architecture can further push AP to 30.25%. Moreover, the visual query system can be evalu- ated in both 2D and 3D configurations. In VQ2D, we ob- serve a significant improvement from baseline 0.13to0.18 on the leaderboard. In VQ3D, we show consistent improve- ment across all metrics. In the end, our improved context- aware query object detector ranked first and second respec- tively in the VQ2D and VQ3D tasks in the 2nd Ego4D chal- lenge, and achieve SOTA in Few-Shot Detection (FSD) on the COCO dataset.
Ye_Decoupling_Human_and_Camera_Motion_From_Videos_in_the_Wild_CVPR_2023	Abstract We propose a method to reconstruct global human tra- jectories from videos in the wild. Our optimization method decouples the camera and human motion, which allows us to place people in the same world coordinate frame. Most existing methods do not model the camera motion; meth- ods that rely on the background pixels to infer 3D human motion usually require a full scene reconstruction, which is often not possible for in-the-wild videos. However, even when existing SLAM systems cannot recover accurate scene reconstructions, the background pixel motion still provides enough signal to constrain the camera motion. We show that relative camera estimates along with data-driven hu- man motion priors can resolve the scene scale ambiguity and recover global human trajectories. Our method robustly recovers the global 3D trajectories of people in challeng- ing in-the-wild videos, such as PoseTrack. We quantify our improvement over existing methods on 3D human dataset Egobody. We further demonstrate that our recovered camera scale allows us to reason about motion of multiple people in a shared coordinate frame, which improves performance of downstream tracking in PoseTrack.	1. Introduction Consider the video sequence in Figure 1. As human ob- servers, we can clearly perceive that the camera is following the athlete as he runs across the field. However, when this dynamic 3D scene is projected onto 2D images, because the camera tracks the athlete, the athlete appears to be at the center of the camera frame throughout the sequence — i.e. the projection only captures the netmotion of the underlying human and camera trajectory. Thus, if we rely only on the person’s 2D motion, as many human video reconstruction methods do, we cannot recover their original trajectory in the world (Figure 1 bottom left). To recover the person’s 3D motion in the world (Figure 1 bottom right), we must also reason about how much the camera is moving. We present an approach that models the camera motion to recover the 3D human motion in the world from videos in the wild. Our system can handle multiple people and re- constructs their motion in the same world coordinate frame, enabling us to capture their spatial relationships. Recovering the underlying human motion and their spatial relationships is a key step towards understanding humans from in-the-wild videos. Tasks, such as autonomous planning in environments with humans [50], or recognition of human interactions with This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21222 the environment [68] and other people [21, 37], rely on in- formation about global human trajectories. Current methods that recover global trajectories either require additional sen- sors, e.g. multiple cameras or depth sensors [11, 51], or dense 3D reconstruction of the environment [12, 31], both of which are only realistic in active or controlled capture settings. Our method acquires these global trajectories from videos in the wild, with no constraints on the capture setup, camera motion, or prior knowledge of the environment. Be- ing able to do this from dynamic cameras is particularly relevant with the emergence of large-scale egocentric video datasets [5, 9, 69]. To do this, given an input RGB video, we first estimate the relative camera motion between frames from the static scene’s pixel motion with a SLAM system [56]. At the same time, we estimate the identities and body poses of all detected people with a 3D human tracking system [45]. We use these estimates to initialize the trajectories of the humans and cameras in the shared world frame. We then optimize these global trajectories over multiple stages to be consistent with both the 2D observations in the video and learned priors about how human move in the world [46]. We illustrate our pipeline in Figure 2. Unlike existing works [11, 31], we optimize over human and camera trajectories in the world frame without requiring an accurate 3D reconstruction of the static scene. Because of this, our method operates on videos captured in the wild, a challenging domain for prior methods that require good 3D geometry, since these videos rarely contain camera viewpoints with sufficient baselines for reliable scene reconstruction. We combine two main insights to enable this optimization. First, even when the scene parallax is insufficient for accu- rate scene reconstruction, it still allows reasonable estimates of camera motion up to an arbitrary scale factor. In fact, in Figure 2, the recovered scene structure for the input video is a degenerate flat plane, but the relative camera motion still explains the scene parallax between frames. Second, human bodies can move realistically in the world in a small range of ways. Learned priors capture this space of realistic human motion well. We use these insights to parameterize the camera trajectory to be both consistent with the scene parallax and the 2D reprojection of realistic human trajecto- ries in the world. Specifically, we optimize over the scale of camera displacement, using the relative camera estimates, to be consistent with the human displacement. Moreover, when multiple people are present in a video, as is often the case in in-the-wild videos, the motions of all the people further constrains the camera scale, allowing our method to operate on complex videos of people. We evaluate our approach on EgoBody [69], a new dataset of videos captured with a dynamic (ego-centric) camera with ground truth 3D global human motion trajectory. Our approach achieves significant improvement upon the state-of-the-art method that also tries to recover the human mo- tion without considering the signal provided by the back- ground pixels [63]. We further evaluate our approach on PoseTrack [1], a challenging in-the-wild video dataset origi- nally designed for tracking. To demonstrate the robustness of our approach, we provide the results on all PoseTrack validation sequences on our project page. On top of qualita- tive evaluation, since there are no 3D ground-truth labels in PoseTrack, we test our approach through an evaluation on the downstream application of tracking. We show that the re- covered scaled camera motion trajectory can be directly used in the PHALP system [45] to improve tracking. The scaled camera enables more persistent 3D human registration in the 3D world, which reduces the re-identification mistakes. We provide video results and code at the project page.
Zhang_SINE_SINgle_Image_Editing_With_Text-to-Image_Diffusion_Models_CVPR_2023	Abstract Recent works on diffusion models have demonstrated a strong capability for conditioning image generation, e.g., text-guided image synthesis. Such success inspires many efforts trying to use large-scale pre-trained diffusion mod- els for tackling a challenging problem–real image editing. Works conducted in this area learn a unique textual token corresponding to several images containing the same ob- ject. However, under many circumstances, only one im- age is available, such as the painting of the Girl with a Pearl Earring. Using existing works on fine-tuning the pre- trained diffusion models with a single image causes severe overfitting issues. The information leakage from the pre- trained diffusion models makes editing can not keep the same content as the given image while creating new fea- tures depicted by the language guidance. This work aims to address the problem of single-image editing. We propose a novel model-based guidance built upon the classifier-free guidance so that the knowledge from the model trained on a single image can be distilled into the pre-trained diffu- sion model, enabling content creation even with one given image. Additionally, we propose a patch-based fine-tuning that can effectively help the model generate images of arbi- trary resolution. We provide extensive experiments to vali- date the design choices of our approach and show promis- ing editing capabilities, including changing style, content addition, and object manipulation. Our code is made pub- licly available here.	1. Introduction Automatic real image editing is an exciting direction, enabling content generation and creation with minimal ef- fort. Although many works have been conducted in this area, achieving high-fidelity semantic manipulation on an image is still a challenging problem for the generative mod- els, considering the target image might be out of the train- ing data distribution [4, 11, 25, 26, 32, 58, 59]. The recently introduced large-scale text-to-image models, e.g., DALL·E 2 [37], Imagen [42], Parti [55], and StableDiffusion [39], can perform high-quality and diverse image generation with “… sculpture”Source Image Source Image“…super-hero cape”“coffee machine…”“coffee maker…” “… in snow”“… blossom street”Source ImageSource Image“…covered by snow”Figure 1. With only onereal image, i.e., Source Image, our method is able to manipulate and generate the content in various ways, such as changing style, adding context, modifying the object, and enlarging the resolution, through guidance from the text prompt. natural language guidance. The success of these works has inspired many subsequent efforts to leverage the pre-trained large-scale models for real image editing [10, 15, 20, 40]. They show that, with properly designed prompts and a lim- ited number of fine-tuning steps, the text-to-image models can manipulate a given subject with text guidance. On the downside, the recent text-guided editing works that build upon the diffusion models suffer several limita- tions. First, the fine-tuning process might lead to the pre- trained large-scale model overfit on the real image, which degrades the synthesized images’ quality when editing. To tackle these issues, methods like using multiple images with the same content and applying regularization terms on the same object have been introduced [4, 40]. However, query- ing multiple images with identical content or object might not be an available choice; for instance, there is only one painting for Girl with a Pearl Earring . Directly editing the single image brings information leakage from the pre- trained large-scale models, generating images with differ- ent content (examples in Fig. 5); therefore, the application scenarios of these methods are greatly constrained. Second, these works lack a reasonable understanding of the object geometry for the edited image. Thus, generating images with different spatial size as the training data cause unde- sired artifacts, e.g., repeated objects, and incorrectly modi- fied geometry (examples in Fig. 5). Such drawbacks restrict applying these methods for generating images with an ar- bitrary resolution, e.g., synthesizing high-resolution images from a single photo of a castle (as in Fig. 4), again limiting This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6027 the usage of these methods. In this work, we present SINE, a framework utilizing pre-trained text-to-image diffusion models for SINgle im- ageEditing and content manipulation. We build our ap- proach based upon existing text-guided image generation approaches [10, 40] and propose the following novel tech- niques to solve overfitting issues on content and geometry, and language drift [28, 31]: • First, by appropriately modifying the classifier-free guid- ance [22], we introduce model-based classifier-free guid- ance that utilizes the diffusion model to provide the score guidance for content and structure. Taking advantage of the step-by-step sampling process used in diffusion mod- els, we use the model fine-tuned on a single image to plant a content “seed” at the early stage of the denois- ing process and allow the pre-trained large-scale text-to- image model to edit creatively conditioned with the lan- guage guidance at a later stage. • Second, to decouple the correlation between pixel posi- tion and content, we propose a patch-based fine-tuning strategy, enabling generation on arbitrary resolution. With a text descriptor describing the content that is aimed to be manipulated and language guidance depicting the desired output, our approach can edit the single unique image to the targeted domain with details preserved in arbi- trary resolution. The output image keeps the structure and background intact while having features well-aligned with the target language guidance. As shown in Fig. 1, trained on a painting Girl with a Pearl Earring with resolution as 512×512, we can sample an image of a sculpture of the girl at the resolution of 640×512with the identity features preserved. Moreover, our method can successfully handle various edits such as style transfer, content addition, and object manipulation (more examples in Fig. 3). We hope our method can further boost creative content creation by opening the door to editing arbitrary images.
Xu_Zero-Shot_Object_Counting_CVPR_2023	Abstract Class-agnostic object counting aims to count object in- stances of an arbitrary class at test time. Current methods for this challenging problem require human-annotated ex- emplars as inputs, which are often unavailable for novel categories, especially for autonomous systems. Thus, we propose zero-shot object counting (ZSC), a new setting where only the class name is available during test time. Such a counting system does not require human annotators in the loop and can operate automatically. Starting from a class name, we propose a method that can accurately iden- tify the optimal patches which can then be used as counting exemplars. Specifically, we first construct a class prototype to select the patches that are likely to contain the objects of interest, namely class-relevant patches. Furthermore, we introduce a model that can quantitatively measure how suit- able an arbitrary patch is as a counting exemplar. By ap- plying this model to all the candidate patches, we can se- lect the most suitable patches as exemplars for counting. Experimental results on a recent class-agnostic counting dataset, FSC-147, validate the effectiveness of our method. Code is available at https://github.com/cvlab- stonybrook/zero-shot-counting .	1. Introduction Object counting aims to infer the number of objects in an image. Most of the existing methods focus on counting objects from specialized categories such as human crowds [37], cars [29], animals [4], and cells [46]. These meth- ods count only a single category at a time. Recently, class- agnostic counting [28, 34, 38] has been proposed to count objects of arbitrary categories. Several human-annotated bounding boxes of objects are required to specify the ob- jects of interest (see Figure 1a). However, having humans in the loop is not practical for many real-world applications, such as fully automated wildlife monitoring systems or vi- *Work done prior to joining Amazon Be happy , Douya(a) Few-shot Counting (b) Zero-Shot Counting Figure 1. Our proposed task of zero-shot object counting (ZSC). Traditional few-shot counting methods require a few exemplars of the object category (a). We propose zero-shot counting where the counter only needs the class name to count the number of object instances. (b). Few-shot counting methods require human annota- tors at test time while zero-shot counters can be fully automatic. sual anomaly detection systems. A more practical setting, exemplar-free class-agnostic counting, has been proposed recently by Ranjan et al. [33]. They introduce RepRPN, which first identifies the objects that occur most frequently in the image, and then uses them as exemplars for object counting. Even though RepRPN does not require any annotated boxes at test time, the method simply counts objects from the class with the high- est number of instances. Thus, it can not be used for count- ing a specific class of interest. The method is only suitable for counting images with a single dominant object class, which limits the potential applicability. Our goal is to build an exemplar-free object counter where we can specify what to count. To this end, we in- troduce a new counting task in which the user only needs to provide the name of the class for counting rather than the exemplars (see Figure 1b). In this way, the counting model can not only operate in an automatic manner but also allow the user to define what to count by simply providing the class name. Note that the class to count during test time can be arbitrary. For cases where the test class is completely unseen to the trained model, the counter needs to adapt to the unseen class without any annotated data. Hence, we This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15548 name this setting zero-shot object counting (ZSC), inspired by previous zero-shot learning approaches [6, 57]. To count without any annotated exemplars, our idea is to identify a few patches in the input image containing the target object that can be used as counting exemplars. Here the challenges are twofold: 1) how to localize patches that contain the object of interest based on the provided class name, and 2) how to select good exemplars for counting. Ideally, good object exemplars are visually representative for most instances in the image, which can benefit the object counter. In addition, we want to avoid selecting patches that contain irrelevant objects or backgrounds, which likely lead to incorrect object counts. To this end, we propose a two-step method that first lo- calizes the class-relevant patches which contain the objects of interest based on the given class name, and then selects among these patches the optimal exemplars for counting. We use these selected exemplars, together with a pre-trained exemplar-based counting model, to achieve exemplar-free object counting. In particular, to localize the patches containing the ob- jects of interest, we first construct a class prototype in a pre- trained embedding space based on the given class name. To construct the class prototype, we train a conditional vari- ational autoencoder (V AE) to generate features for an ar- bitrary class conditioned on its semantic embedding. The class prototype is computed by taking the average of the generated features. We then select the patches whose em- beddings are the k-nearest neighbors of the class prototype as the class-relevant patches. After obtaining the class-relevant patches, we further se- lect among them the optimal patches to be used as counting exemplars. Here we observe that the feature maps obtained using good exemplars and badexemplars often exhibit dis- tinguishable differences. An example of the feature maps obtained with different exemplars is shown in Figure 2. The feature map from a good exemplar typically exhibits some repetitive patterns (e.g., the dots on the feature map) that center around the object areas while the patterns from a bad exemplar are more irregular and occur randomly across the image. Based on this observation, we train a model to mea- sure the goodness of an input patch based on its correspond- ing feature maps. Specifically, given an arbitrary patch and a pre-trained exemplar-based object counter, we train this model to predict the counting error of the counter when us- ing the patch as the exemplar. Here the counting error can indicate the goodness of the exemplar. After this error pre- dictor is trained, we use it to select those patches with the smallest predicted errors as the final exemplars for counting. Experiments on the FSC-147 dataset show that our method outperforms the previous exemplar-free counting method [33] by a large margin. We also provide analy- ses to show that patches selected by our method can be Pre-trainedCounter Bad ExemplarGood Exemplar Query ImageFigure 2. Feature maps obtained using different exemplars given a pre-trained exemplar-based counting model. The feature maps obtained using good exemplars typically exhibit some repetitive patterns while the patterns from bad exemplars are more irregular. used in other exemplar-based counting methods to achieve exemplar-free counting. In short, our main contributions can be summarized as follows: • We introduce the task of zero-shot object counting that counts the number of instances of a specific class in the input image, given only the class name and without relying on any human-annotated exemplars. • We propose a simple yet effective patch selec- tion method that can accurately localize the optimal patches across the query image as exemplars for zero- shot object counting. • We verify the effectiveness of our method on the FSC- 147 dataset, through extensive ablation studies and vi- sualization results.
Yang_Vector_Quantization_With_Self-Attention_for_Quality-Independent_Representation_Learning_CVPR_2023	Abstract Recently, the robustness of deep neural networks has drawn extensive attention due to the potential distribution shift between training and testing data (e.g., deep mod- els trained on high-quality images are sensitive to corrup- tion during testing). Many researchers attempt to make the model learn invariant representations from multiple corrupted data through data augmentation or image-pair- based feature distillation to improve the robustness. In- spired by sparse representation in image restoration, we opt to address this issue by learning image-quality-independent feature representation in a simple plug-and-play manner, that is, to introduce discrete vector quantization (VQ) to re- move redundancy in recognition models. Specifically, we first add a codebook module to the network to quantize deep features. Then we concatenate them and design a self-attention module to enhance the representation. During training, we enforce the quantization of features from clean and corrupted images in the same discrete embedding space so that an invariant quality-independent feature representa- tion can be learned to improve the recognition robustness of low-quality images. Qualitative and quantitative experi- mental results show that our method achieved this goal ef- fectively, leading to a new state-of-the-art result of 43.1 % mCE on ImageNet-C with ResNet50 as the backbone. On other robustness benchmark datasets, such as ImageNet-R, our method also has an accuracy improvement of almost 2%. The source code is available at https://see. xidian.edu.cn/faculty/wsdong/Projects/ VQSA.htm	1. Introduction The past few years have witnessed the remarkable devel- opment of deep convolutional neural networks (DCNNs) in many recognition tasks, such as classification [9,20,31,46], *Corresponding author. (a) Image (b) Clean (c) QualNet (d) OursFigure 1. The Grad-CAM [45] maps of different models on defo- cus blur images. (a) The clean images. (b) The maps of vanilla ResNet50 [20] model on clean images. (c) and (d) show the maps of QualNet50 [29] and our proposed method on defocus blur im- ages with severity level 3. The results show that our method still can focus on the salient object area without being seriously af- fected by corruption. Best viewed in color. detection [43, 48] and segmentation [5, 19]. Although its performance has exceeded that of humans in some datasets, its robustness still lags behind [10, 15]. Many re- searchers [11, 21, 22, 53] have shown that the performance of deep models trained in high-quality data decreases dra- matically with low-quality data encountered during deploy- ment, which usually contain common corruptions, includ- ing blur, noise, and weather influence. For example, the vanilla ResNet50 model has an accuracy of 76 %in the clean ImageNet validation set, but its average accuracy is less than 30 %in the same dataset contaminated by Gaus- sian noise. As shown in Fig. 1, the model’s attention map is seri- ously affected by image quality. That is, the model pays This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 24438 close attention to an incomplete or incorrect area of the de- focus blur image, which results in a wrong prediction. Deep degradation prior (DDP) [55] also reveals that deep repre- sentation space degradation is the reason for the decrease in model performance, and finding a mapping between de- graded and clean features will certainly benefit downstream classification or detection tasks. Generally, a simple and effective method is to consider these common corruptions and to use low-quality simulated images for augmented (fine-tuning) training. However, as described in [29], this method does not explicitly map low- quality features to high-quality features; instead, the model tends to learn the average distribution among corruptions, resulting in limited performance improvement. In addition, when the type of corruption is unknown during training, this approach is heavily based on well-designed augmentation strategies to improve generalization and robustness. To enhance the feature representation of low-quality im- ages, another feasible solution is to use pairs of clean and degraded images for training and align the degraded fea- tures with the corresponding clean ones. Both DDP [55] and QualNet [29] use high-quality features extracted from clean images as supervision and improve low-quality fea- tures through a distillation-like approach [1, 25, 26, 58]. Al- though they learn the mapping relationship between low- and high-quality feature representation, paired images are time-consuming for training, and some irrelevant features that are not useful for recognition are also forcibly aligned, such as background, color, lighting, and other features which are affected by the image quality. The vector quantization (VQ) process is essentially a special case in sparse representation, that is, the represen- tation coefficient is a one-hot vector [50]. Similarly to the application of sparse representation in image restoration [13, 59], the compression-reconstruction learning frame- work of VQ is also conducive to the removal of noisy redun- dant information and learning essential features for recog- nition. Inspired by this, we propose to use VQ to bridge the gap between low- and high-quality features and learn a quality-independent representation. Specifically, we first add a vector quantizer (codebook) to the recognition model. During the training process, due to the codebook update mechanism, both low- and high-quality features will be as- signed to the same discrete embedding space. Subsequently, since direct hard quantization (selecting and replacing) may lose some useful information, we choose to concatenate the quantized features with the original ones and use a self- attention module to further enhance the quality-independent representation and weaken irrelevant features. In summary, the main contributions of this paper are listed below. • To the best of our knowledge, it is the first time that we propose to introduce vector quantization into the recognition model for learning quality-independentfeature representation and improving the models’ ro- bustness on common corruptions. • We concatenate the quantized feature vector with the original one and use the self-attention module to en- hance the quality-independent feature representation instead of direct replacement in the standard vector quantization method. • Extensive experimental results show that our method has achieved higher accuracy on benchmark low- quality datasets than several current state-of-the-art methods.
Zhang_PHA_Patch-Wise_High-Frequency_Augmentation_for_Transformer-Based_Person_Re-Identification_CVPR_2023	Abstract Although recent studies empirically show that injecting Convolutional Neural Networks (CNNs) into Vision Trans- formers (ViTs) can improve the performance of person re- identification, the rationale behind it remains elusive. From a frequency perspective, we reveal that ViTs perform worse than CNNs in preserving key high-frequency components (e.g, clothes texture details) since high-frequency compo- nents are inevitably diluted by low-frequency ones due to the intrinsic Self-Attention within ViTs. To remedy such inadequacy of the ViT, we propose a Patch-wise High- frequency Augmentation (PHA) method with two core de- signs. First , to enhance the feature representation ability of high-frequency components, we split patches with high- frequency components by the Discrete Haar Wavelet Trans- form, then empower the ViT to take the split patches as aux- iliary input. Second , to prevent high-frequency components from being diluted by low-frequency ones when taking the entire sequence as input during network optimization, we propose a novel patch-wise contrastive loss. From the view of gradient optimization, it acts as an implicit augmentation to improve the representation ability of key high-frequency components. This benefits the ViT to capture key high- frequency components to extract discriminative person rep- resentations. PHA is necessary during training and can be removed during inference, without bringing extra complex- ity. Extensive experiments on widely-used ReID datasets validate the effectiveness of our method.	1. Introduction *Corresponding author https://github.com/zhangguiwei610/PHA This work was partially supported by the National Natural Science Foundation of China (No. 62072022) and the Fundamental Research Funds for the Central Universities. 2540557085100ResNet101TransReID 0612182430ResNet101TransReID Low-frequency High-frequency Person images Texture details5060708090100ResNet101TransReID mAP (Market1501)Rank-1 (Market1501)mAP (MSMT17)Rank-1 (MSMT17)88.49558.880.695.2↑0.267.4↑8.685.3↑4.788.9↑0.5 75.690.346.272.476.8↑1.291.5↑1.257.4↑11.278.8↑6.4 8.7↑2.421.8↑4.67.3↑1.924.4↑5.96.317.25.418.5DHWT(a)(b)(c)Figure 1. Performance comparisons between ResNet101 ( 44.7M #Param ) and TransReID ( 100M #Param ) on (a) original person images, (b) low-frequency components, and (c) high-frequency components of Market1501 and MSMT17 datasets, respectively. Note that “#Param” refers to the number of parameters. Person re-identification (ReID) aims to retrieve a specific person, given in a query image, from the search over a large set of images captured by various cameras [33, 36, 37, 45]. Most research to date has focused on extracting discrimi- native person representations from single images, either by Convolutional Neural Networks (CNNs) [24, 25, 39, 42], Vision Transformers (ViTs) [2, 8, 23, 27, 41, 43] or Hybrid- based approaches [10, 11, 14, 16, 35]. These studies empir- ically show that injecting CNNs into ViTs can improve the discriminative of person representations [8, 35]. Despite a couple of empirical solutions, the rationale for why ViTs can be improved by CNNs remains elusive. To this end, we explore possible reasons from a frequency per- spective, which is of great significance in digital image pro- cessing [4, 12, 32]. As shown in Fig. 1, we first employ Discrete Haar Wavelet Transform (DHWT) [19] to trans- form original person images into low-frequency compo- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14133 nents and high-frequency components, then conduct perfor- mance comparisons between the ResNet101 [6] and Tran- sReID [8] on original images, the low-frequency compo- nents and high-frequency components of Market1501 and MSMT17 datasets, respectively. Our comparisons reveal: (1) Certain texture details of person images, which are more related to the high-frequency components, are crucial for ReID tasks. Specifically, from Fig. 1 (a) and (b), with the same model, the performance on the origi- nal images is consistently better than that on low-frequency components. Taking the TransReID as an example, the Rank-1/mAP on original images of MSMT17 dataset is 6.5%/10.0% higher than that on the low-frequency compo- nents. The root reason might be that low-frequency compo- nents only reflect coarse-grained visual patterns of images, and lose texture details (e.g., bags and edges). In contrast, the lost details are more related to the high-frequency com- ponents, as shown in Fig. 1 (c). The degradation from Fig. 1 (a) to (b) indicates that certain details are key components to improve the performance of ReID. (2) The ViT performs worse than CNNs in preserv- ing key high-frequency components (e.g., texture de- tails of clothes and bags) of person images. As shown in Fig. 1 (c), although the TransReID outperforms the ResNet101 consistently on the original images and low- frequency components, the ResNet101 exceeds the Tran- sReID by 4.6%/2.4% and 5.9%/1.9% Rank-1/mAP on high-frequency components of Market1501 and MSMT17 datasets. The poor performance of the TransReID on high- frequency components shows its inadequacy in capturing key high-frequency details of person images. In view of the above, we analyze the possible reason for such inadequacy of the ViT by revisiting Self-Attention from a frequency perspective (Sec. 3.1). We reveal that high-frequency components of person images are inevitably diluted by low-frequency ones due to the Self-Attention mechanism within ViTs. To remedy such inadequacy of the ViT without modifying its architecture, we propose a Patch-wise High-frequency Augmentation (PHA) method with two core designs (Sec. 3.2). First , unlike previous works that directly take frequency subbands as network in- put [1, 4, 5, 17], we split patches with high-frequency com- ponents by the Discrete Haar Wavelet Transform (DHWT) and drop certain low-frequency components correspond- ingly, then empower the ViT to take the split patches as auxiliary input. This benefits the ViT to enhance the fea- ture representation ability of high-frequency components. Note that the dropped components are imperceptible to hu- man eyes but essential for the model, thereby preventing the model from overfitting to low-frequency components. Second , to prevent high-frequency components from be- ing diluted by low-frequency ones when taking the entire sequence as input during network optimization, we pro-pose a novel patch-wise contrastive loss. From the view of gradient optimization, it acts as an implicit augmentation to enhance the feature representation ability of key high- frequency components to extract discriminative person rep- resentations (Sec. 3.3). With it, our PHA is necessary dur- ing training and can be discarded during inference, without bringing extra complexity. Our contributions include: •We reveal that due to the intrinsic Self-Attention mech- anism, the ViT performs worse than CNNs in capturing high-frequency components of person images, which are key ingredients for ReID. Hence, we develop a Patch-wise High-frequency Augmentation (PHA) to extract discriminative person representations by en- hancing high-frequency components. •We propose a novel patch-wise contrastive loss, en- abling the ViT to preserve key high-frequency com- ponents of person images. From the view of gradi- ent optimization, it acts as an implicit augmentation to enhance the feature representation ability of key high- frequency components. By virtue of it, our PHA is necessary during training and can be removed during inference, without bringing extra complexity. •Extensive experimental results perform favorably against the mainstream methods on CUHK03-NP, Market-1501, and MSMT17 datasets.
Yao_HGNet_Learning_Hierarchical_Geometry_From_Points_Edges_and_Surfaces_CVPR_2023	Abstract Parsing an unstructured point set into constituent local geometry structures (e.g., edges or surfaces) would be help- ful for understanding and representing point clouds. This motivates us to design a deep architecture to model the hierarchical geometry from points, edges, surfaces (trian- gles), to super-surfaces (adjacent surfaces) for the thor- ough analysis of point clouds. In this paper, we present a novel Hierarchical Geometry Network (HGNet) that in- tegrates such hierarchical geometry structures from super- surfaces, surfaces, edges, to points in a top-down man- ner for learning point cloud representations. Technically, we first construct the edges between every two neighbor points. A point-level representation is learnt with edge-to- point aggregation, i.e., aggregating all connected edges into the anchor point. Next, as every two neighbor edges com- pose a surface, we obtain the edge-level representation of each anchor edge via surface-to-edge aggregation over all neighbor surfaces. Furthermore, the surface-level repre- sentation is achieved through super-surface-to-surface ag- gregation by transforming all super-surfaces into the an- chor surface. A Transformer structure is finally devised to unify all the point-level, edge-level, and surface-level fea- tures into the holistic point cloud representations. Extensive experiments on four point cloud analysis datasets demon- strate the superiority of HGNet for 3D object classification and part/semantic segmentation tasks. More remarkably, HGNet achieves the overall accuracy of 89.2% on ScanOb- jectNN, improving PointNeXt-S by 1.5%.	1. Introduction With the growing popularity of 3D acquisition technolo- gies (such as 3D scanners, RGBD cameras, and LiDARs), 3D point cloud analysis has been a fast-developing topic in the past decade. Practical point cloud analysis systems have great potential for numerous applications, e.g., autonomous driving, robotics, and virtual reality. In comparison to 2D RGB image data, 3D data of point clouds has an additional dimension of depth, leading to a three dimensional world as Point Surface Super -surface Edge Figure 1. The progressive development of a four-level hierarchi- cal geometry structure among points, edges, surfaces, and super- surfaces. Such hierarchical geometry structure further in turn strengthens surface-level, edge-level, and point-level features in a top-down manner. in our daily life. The 3D point clouds are armed with several key merits, e.g., preserving rich geometry information and being invariant to lighting conditions. Nevertheless, consid- ering that point clouds are naturally unstructured without any discretization, it is not trivial to directly applying the typical CNN operations widely adopted in 2D vision over the primary point set for analyzing 3D data. To alleviate this limitation, a series of efforts have at- tempted to interpret the geometry information rooted in unstructured point clouds, such as the regular point-based [30, 38], edge/relation-based [32, 46], or surface-based [39] representations. For instance, PointNet [30] builds up the foundation of regular point-based feature extraction via stacked Multi-Layer Perceptrons (MLPs) plus symmetric aggregation. The subsequent works (e.g., PointNet++ [32] and DGCNN [46]) further incorporate edge/relation-based features by using set abstraction or graph models for aggre- gating/propagating points and edges, pursuing the mining of finer geometric structures. More recently, RepSurf [39] learns surface-based features over triangle meshes to enable an amplified expression of local geometry. Despite showing encouraging performances, most existing point cloud anal- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21846 ysis techniques lack a thorough understanding of the com- positional local geometry structures in the point clouds. In this work, we propose to mitigate this issue from the standpoint of parsing an unstructured point set into a hierarchical structure of constituent local geometry struc- tures to comprehensively characterize the point clouds. Our launching point is to construct a bottom-up local hierarchi- cal topology from the leaf level of irregular and unordered points, to the intermediary levels of edges and surfaces, and the root level of super-surfaces (i.e., adjacent surfaces). Fig- ure 1 conceptually depicts the progressive development of a hierarchical geometry structure for a local set of points. Such hierarchical structure in turn triggers the reinforce- ment of 3D model capacity for capturing comprehensive ge- ometry information. The learning of multi-granular geom- etry features, i.e., surface-level, edge-level, and point-level features, is enhanced by aggregating geometry information from super-surfaces, surfaces, or edges derived from the hi- erarchical topology in a top-down fashion. By consolidating the idea of delving into the local hier- archical geometry in point clouds, we present a new Hier- archical Geometry Network (HGNet) to facilitate the learn- ing of point cloud representations. Specifically, we con- struct a four-level hierarchy among points, edges, surfaces, and super-surfaces. The edges are first built between two neighbor points and each is extended to a surface (trian- gle) by connecting another neighbor point. The adjacent surfaces are then grouped as a super-surface. Next, HGNet executes top-down feature aggregation from super-surfaces, surfaces, and edges, leading to three upgraded geometry features of surfaces, edges, and points, respectively. Finally, we capitalize on a Transformer structure to unify these re- fined surface-level, edge-level, and point-level geometry features with holistic contextual information, thereby im- proving point cloud analysis. We analyze and evaluate our HGNet through extensive experiments over multiple 3D vision tasks for point cloud analysis (e.g., 3D object classification and part/semantic segmentation tasks), which empirically demonstrate its su- perior advantage against state-of-the-art backbones. More remarkably, on the real-world challenging benchmark of ScanObjectNN, HGNet achieves to-date the best published overall accuracy of 89.2%, which absolutely improves PointNeXt-S (87.7%) with 1.5%. For part and semantic segmentation on PartNet and S3DIS datasets, HGNet sur- passes PosPool and PointNeXt by 1.8% and 1.8% in mIoU.
Yang_BEVHeight_A_Robust_Framework_for_Vision-Based_Roadside_3D_Object_Detection_CVPR_2023	Abstract While most recent autonomous driving system focuses on developing perception methods on ego-vehicle sensors, people tend to overlook an alternative approach to lever- age intelligent roadside cameras to extend the perception ability beyond the visual range. We discover that the state- of-the-art vision-centric bird’s eye view detection methods have inferior performances on roadside cameras. This is because these methods mainly focus on recovering the depth regarding the camera center, where the depth difference be- *Work done during an internship at DAMO Academy, Alibaba Group. †Corresponding Author.tween the car and the ground quickly shrinks while the dis- tance increases. In this paper, we propose a simple yet effective approach, dubbed BEVHeight, to address this is- sue. In essence, instead of predicting the pixel-wise depth, we regress the height to the ground to achieve a distance- agnostic formulation to ease the optimization process of camera-only perception methods. On popular 3D detection benchmarks of roadside cameras, our method surpasses all previous vision-centric methods by a significant mar- gin. The code is available at https://github.com/ ADLab-AutoDrive/BEVHeight . This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21611	1. Introduction The rising tide of autonomous driving vehicles draws vast research attention to many 3D perception tasks, of which 3D object detection plays a critical role. While most recent works tend to only rely on ego-vehicle sensors, there are certain downsides of this line of work that hinders the perception capability under given scenarios. For example, as the mounting position of cameras is relatively close to the ground, obstacles can be easily occluded by other vehi- cles to cause severe crash damage. To this end, people have started to develop perception systems that leverage intelli- gent units on the roadside, such as cameras, to address such occlusion issue and enlarge perception range so to increase the response time in case of danger [5,11,28,34,36,37]. To facilitate future research, there are two large-scale bench- mark datasets [36,37] of various roadside cameras and pro- vide an evaluation of certain baseline methods. Recently, people discover that, in contrast to directly pro- jecting the 2D images into a 3D space, leveraging a bird’s eye view (BEV) feature space can significantly improve the perception performance of vision centric system. One line of the recent approach, which constitutes the state-of-the-art camera-only method, is to generate implicitly or explicitly the depth for each pixel to ease the optimization process of bounding box regression. However, as shown in Fig. 1, we visualize the per-pixel depth of a roadside image and notice a phenomenon. Consider two points, one on the roof of a car and another on the nearest ground. If we measure the depth of these points to the camera center, namely dandd′, the difference between these depth d−d′would drastically decrease when the car moves away from the camera. We conjecture this leads to two potential downsides: i) unlike the autonomous vehicle that has a consistent camera pose, roadside ones usually have different camera poses across the datasets, which makes regressing depth hard; ii) depth prediction is very sensitive to the change of extrinsic param- eter, where it happens quite often in the real world. On the contrary, we notice that the height to the ground is consistent regardless of the distance between car and cam- era center. To this end, we propose a novel framework to predict the per-pixel height instead of depth, dubbed BEVHeight. Specifically, our method firstly predicts cat- egorical height distribution for each pixel to project rich contextual feature information to the appropriate height in- terval in wedgy voxel space. Followed by a voxel pooling operation and a detection head to get the final output de- tections. Besides, we propose a hyperparameter-adjustable height sampling strategy. Note that our framework does not depend on explicit supervision like point clouds. We conduct extensive experiments on two popular roadside perception benchmarks, DAIR-V2X [37] and Rope3D [36]. On traditional settings where there is no disruption to the cameras, our BEVHeight achieves state-of-the-art performance and surpasses all previous methods, regardless of monocular 3D detectors or recent bird’s eye view methods by a margin of 5%. In realistic scenarios, the extrinsic parameters of these roadside units can be subject to changes due to various reasons, such as maintenance and wind blows. We simulate these scenarios following [38] and observe a severe performance drop of the BEVDepth, from 41.97% to 5.49%. Compared to these methods, we show- case the benefit of predicting the height instead of depth and achieve 26.88% improvement over the BEVDepth [15], which further evidences the robustness of our method.
Zhang_Diverse_Embedding_Expansion_Network_and_Low-Light_Cross-Modality_Benchmark_for_Visible-Infrared_CVPR_2023	Abstract For the visible-infrared person re-identification (VIReID) task, one of the major challenges is the modality gaps between visible (VIS) and infrared (IR) images. However, the training samples are usually lim- ited, while the modality gaps are too large, which leads that the existing methods cannot effectively mine diverse cross-modality clues. To handle this limitation, we propose a novel augmentation network in the embedding space, called diverse embedding expansion network (DEEN). The proposed DEEN can effectively generate diverse em- beddings to learn the informative feature representations and reduce the modality discrepancy between the VIS and IR images. Moreover, the VIReID model may be seriously affected by drastic illumination changes, while all the existing VIReID datasets are captured under suffi- cient illumination without significant light changes. Thus, we provide a low-light cross-modality (LLCM) dataset, which contains 46,767 bounding boxes of 1,064 identities captured by 9 RGB/IR cameras. Extensive experiments on the SYSU-MM01, RegDB and LLCM datasets show the superiority of the proposed DEEN over several other state-of-the-art methods. The code and dataset are released at:https://github.com/ZYK100/LLCM	1. Introduction Person re-identification (ReID) aims to match a given person with gallery images captured by different cameras [3, 9, 52]. Most existing ReID methods [22, 24, 30, 38, 50] only focus on matching RGB images captured by visible cameras at daytime. However, these methods may fail *Corresponding author. CNN dimensionamplitude 0VIS imageIR image CNN with DEEN dimensionamplitude 0VIS imageIR image(a) Feature Learning without the proposed DEEN (b) Feature Learning with the proposed DEEN VIS image IR image VIS image IR image Embeddings EmbeddingsFigure 1. Motivation of the proposed DEEN, which aims to gen- erate diverse embeddings to make the network focus on learning with the informative feature representations to reduce the modality gaps between the VIS and IR images. to achieve encouraging results when visible cameras can- not effectively capture person’s information under complex conditions, such as at night or low-light environments. To solve this problem, some visible (VIS)-infrared (IR) per- son re-identification (VIReID) methods [15,39,41,48] have been proposed to retrieve the VIS (IR) images according to the corresponding IR (VIS) images. Compared with the widely studied person ReID task, the VIReID task is much more challenging due to the ad- ditional cross-modality discrepancy between the VIS and IR images [33, 45, 49, 51]. Typically, there are two popu- lar types of methods to reduce this modality discrepancy. One type is the feature-level methods [5, 11, 16, 35, 40, 42], which try to project the VIS and IR features into a common embedding space, where the modality discrepancy can be minimized. However, the large modality discrepancy makes these methods difficult to project the cross-modality images into a common feature space directly. The other type is the image-level methods [4,28,29,32], which aim to reduce the modality discrepancy by translating an IR (or VIS) image 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2153 IR images VIS images IR images VIS images SYSU-MM01 RegDB LLCMFigure 2. Comparison of person images on the SYSU-MM01 (1st row), RegDB (2nd row), and LLCM (3rd-5th rows) datasets. Each row shows four VIS images and four IR images of two identities. It is obvious that our LLCM contains a more challenging and real- istic VIReID environment. into its VIS (or IR) counterpart by using the GANs [8]. De- spite their success in reducing the modality gaps, the gen- erated cross-modality images are usually accompanied by some noises due to the lack of the VIS-IR image pairs. In this paper, we propose a novel augmentation network in the embedding space for the VIReID task, called diverse embedding expansion network (DEEN), which consists of a diverse embedding expansion (DEE) module and a multi- stage feature aggregation (MFA) block. The proposed DEE module can generate more embeddings followed by a novel center-guided pair mining (CPM) loss to drive the DEE module to focus on learning with the diverse feature rep- resentations. As illustrated in Fig. 1, by exploiting the gen- erated embeddings with diverse information, the proposed DEE module can achieve the performance improvement by using more diverse embeddings. The proposed MFA block can aggregate the features from different stages for mining potential channel-wise and spatial feature representations, which increases the network’s capacity for mining different- level diverse embeddings. Moreover, we observe that the existing VIReID datasets are captured under the environments with sufficient illumi- nation. However, the performance of the VIReID methods may be seriously affected by drastic illumination changes or low illuminations. Therefore, we collect a challenging low- light cross-modality dataset, called LLCM dataset, which is shown in Fig. 2. Compared with the other VIReID datasets, the LLCM dataset contains a larger number of identities and images captured under low-light scenes, which introduces more challenges to the real-world VIReID task. In summary, the main contributions are as follows: •We propose a novel diverse embedding expansion (DEE) module with a center-guided pair mining (CPM) loss to generate more embeddings for learning the diverse fea-ture representations. We are the first to augment the embed- dings in the embedding space in VIReID. Besides, we also propose an effective multistage feature aggregation (MFA) block to mine potential channel-wise and spatial feature representations. •With the incorporation of DEE, CPM loss and MFA into an end-to-end learning framework, we propose an effective diverse embedding expansion network (DEEN), which can effectively reduce the modality discrepancy be- tween the VIS and IR images. •We collect a low-light cross-modality (LLCM) dataset, which contains 46,767 images of 1,064 identities captured under the environments with illumination changes and low illuminations. The LLCM dataset has more new and impor- tant features, which can facilitate the research of VIReID towards practical applications. •Extensive experiments show that the proposed DEEN outperforms the other state-of-the-art methods for the VIReID task on three challenging datasets.
Ying_Mapping_Degeneration_Meets_Label_Evolution_Learning_Infrared_Small_Target_Detection_CVPR_2023	Abstract Training a convolutional neural network (CNN) to detect infrared small targets in a fully supervised manner has gained remarkable research interests in recent years, but is highly labor expensive since a large number of per-pixel annotations are required. To handle this problem, in this paper, we make the ﬁrst attempt to achieve infrared small target detection with point-level supervision. Interestingly, during the training phase supervised by point labels, we discover that CNNs ﬁrst learn to segment a cluster of pixels near the targets, and then gradually converge to predict groundtruth point labels. Motivated by this “mapping degeneration” phenomenon, we propose a label evolution framework named label evolution with single point supervision (LESPS) to progressively expand the point label by leveraging the intermediate predictions of CNNs. In this way, the network predictions can ﬁnally approximate the updated pseudo labels, and a pixel-level target mask can be obtained to train CNNs in an end-to-end manner. We conduct extensive experiments with insightful visualizations to validate the effectiveness of our method. Experimental results show that CNNs equipped with LESPS can well recover the target masks from corresponding point labels, and can achieve over 70% and 95% of their fully supervised performance in terms of pixel-level intersection over union (IoU ) and object-level probability of detection (P d), respectively. Code is available at https://github.com/XinyiYing/LESPS.	1. Introduction Infrared small target detection has been a longstanding, fundamental yet challenging task in infrared search and tracking systems, and has various important applications in civil and military fields [49,57], including traffic monitoring This work was supported by National Key Research and Development Program of China No. 2021YFB3100800. 0.001.00 0.50 0.000.26 0.12 0.000.20 0.10 0.000.16 0.08 loss Image Prediction Label Training Epochs Figure 1. An illustration of mapping degeneration under point supervision. CNNs always tend to segment a cluster of pixels near the targets with low conﬁdence at the early stage, and then gradually learn to predict GT point labels with high conﬁdence. [24, 54], maritime rescue [52, 53] and military surveillance [7,47]. Due to the rapid response and robustness to fast-moving scenes, single-frame infrared small target (SIRST) detection methods have always attracted much more attention, and numerous methods have been proposed. Early methods, including ﬁltering-based [9, 40], local contrast-based [3, 16] and low rank-based [11, 43] methods, require complex handcrafted features with carefully tuned hyper-parameters. Recently, compact deep learning has been introduced in solving the problem of SIRST detection [24,45,54]. However, there are only a few attempts, and its potential remains locked, unlike the extensive explorations of deep learning for natural images. This is mainly due to potential reasons, including lack of large-scale, accurately annotated datasets and high stake application scenarios. Infrared small targets are usually of very small size, weak, shapeless and textureless, and are easily submerged in diverse complex background clutters. As a result, directly adopting existing popular generic object detectors like RCNN series [13, 14,19,39], YOLO series [25, 37, 38] and SSD [29] to SIRST detection cannot produce satisfactory performance. Realizing this, researchers have been focusing on developing deep networks tailored for This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15528 infrared small targets by adequately utilizing the domain knowledge. However, most existing deep methods for SIRST detection [8, 24, 54] are fully supervised, which usually requires a large dataset with accurate target mask annotations for training. Clearly, this is costly [5, 26]. Therefore, a natural question arises: Can we develop a new framework for SIRST detection with single point supervision ? In fact, to substantially reduce the annotation cost for object detection tasks, weakly supervised object detection methods with point supervision [4, 5, 26, 56] have been studied in the ﬁeld of computer vision. Although these weakly supervised methods achieve promising results, they are not designed for the problem of SIRST detection, and the class-agnostic labels ( i.e., only foreground and background) of infrared small targets hinder their applications [42, 58]. Therefore, in this work, we intend to conduct the ﬁrst study of weakly supervised SIRST detection with single-point supervision. A key motivation of this work comes from an interesting observation during the training of SIRST detection networks. That is, with single point labels serving as supervision, CNNs always tend to segment a cluster of pixels near the targets with low conﬁdence at the early stage, and then gradually learn to predict groundtruth (GT) point labels with high conﬁdence, as shown in Fig. 1. It reveals the fact that region-to-region mapping is the intermediate result of the ﬁnal region-to-point mapping1. We attribute this “mapping degeneration” phenomenon to the special imaging mechanism of infrared system [24, 54], the local contrast prior of infrared small targets [3, 8], and the easy-to-hard learning property of CNNs [44], in which the ﬁrst two factors result in extended mapping regions beyond the point labels, and the last factor contributes to the degeneration process. Based on the aforementioned discussion, in this work, we propose a novel framework for the problem of weakly supervised SIRST detection, dubbed label evolution with single point supervision (LESPS). Speciﬁcally, LESPS leverages the intermediate network predictions in the training phase to update the current labels, which serve as supervision until the next label update. Through iterative label update and network training, the network predictions can ﬁnally approximate the updated pseudo mask labels, and the network can be simultaneously trained to achieve pixel-level SIRST detection in an end-to-end2manner. Our main contributions are summarized as: (1) We present the ﬁrst study of weakly supervised SIRST 1“region-to-region mapping” represents the mapping learned by CNNs from target regions in images to a cluster of pixels near the targets, while “region-to-point mapping” represents the mapping from target regions in images to the GT point labels. 2Different from generic object detection [32, 59], “end-to-end” here represents achieving point-to-mask label regression and direct pixel-level inference in once training.detection, and introduce LESPS that can signiﬁcantly reduce the annotation cost. (2) We discover the mapping degeneration phenomenon, and leverage this phenomenon to automatically regress pixel-level pseudo labels from the given point labels via LESPS. (3) Experimental results show that our framework can be applied to different existing SIRST detection networks, and enable them to achieve over 70% and 95% of its fully supervised performance in terms of pixel-level intersection over union ( IoU) and object-level probability of detection ( Pd), respectively.
Zhang_Learning_3D_Representations_From_2D_Pre-Trained_Models_via_Image-to-Point_Masked_CVPR_2023	Abstract Pre-training by numerous image data has become de- facto for robust 2D representations. In contrast, due to the expensive data processing, a paucity of 3D datasets severely hinders the learning for high-quality 3D features. In this paper, we propose an alternative to obtain superior 3D representations from 2D pre-trained models via Image-to- Point Masked Autoencoders, named as I2P-MAE . By self- supervised pre-training, we leverage the well learned 2D knowledge to guide 3D masked autoencoding, which recon- structs the masked point tokens with an encoder-decoder ar- chitecture. Specifically, we first utilize off-the-shelf 2D mod- els to extract the multi-view visual features of the input point cloud, and then conduct two types of image-to-point learn- ing schemes. For one, we introduce a 2D-guided masking strategy that maintains semantically important point tokens to be visible. Compared to random masking, the network can better concentrate on significant 3D structures with key spatial cues. For another, we enforce these visible to- kens to reconstruct multi-view 2D features after the decoder. This enables the network to effectively inherit high-level 2D semantics for discriminative 3D modeling. Aided by our image-to-point pre-training, the frozen I2P-MAE, without any fine-tuning, achieves 93.4% accuracy for linear SVM on ModelNet40, competitive to existing fully trained meth- ods. By further fine-tuning on on ScanObjectNN’s hard- est split, I2P-MAE attains the state-of-the-art 90.11% ac- curacy, +3.68% to the second-best, demonstrating supe- rior transferable capacity. Code is available at https: //github.com/ZrrSkywalker/I2P-MAE .	1. Introduction Driven by huge volumes of image data [11, 40, 56, 61], pre-training for better visual representations has gained †Corresponding author 2D Pre-trained Models Image-to-Point Learning Large-scale 2D DataLimited 3D DataPre-train Pre-train I2P-MAEMAEEncoderMAEDecoderRich2DKnowledgeResNet,ViT,Swin,…Figure 1. Image-to-Point Masked Autoencoders. We leverage the 2D pre-trained models to guide the MAE pre-training in 3D, which alleviates the need of large-scale 3D datasets and learns from 2D knowledge for superior 3D representations. much attention in computer vision, which benefits a variety of downstream tasks [8, 29, 39, 55, 62]. Besides supervised pre-training with labels, many researches develop advanced self-supervised approaches to fully utilize raw image data via pre-text tasks, e.g., image-image contrast [5, 9, 10, 21, 28], language-image contrast [12, 53], and masked image modeling [3, 4, 18, 19, 27, 31, 42, 70]. Given the popularity of 2D pre-trained models, it is still absent for large-scale 3D datasets in the community, attributed to the expensive data acquisition and labor-intensive annotation. The widely adopted ShapeNet [6] only contains 50k point clouds of 55 object categories, far less than the 14 million ImageNet [11] and 400 million image-text pairs [53] in 2D vision. Though there have been attempts to extract self-supervisory signals for 3D pre-training [33, 41, 47, 64, 69, 76, 78, 83], raw point clouds with sparse structural patterns cannot provide suf- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21769 RandomMasking 3DCoord.2D-guided Masking 2D Semantics… ……3DCoord.… …2D Attention Maps2D VisualFeaturesVisibleTokens MaskedTokensPoint-MAE&-M2AEI2P-MAEFigure 2. Comparison of (Left) Existing Methods [47, 78] and (Right) our I2P-MAE. On top of the general 3D MAE architec- ture, I2P-MAE introduces two schemes of image-to-point learn- ing: 2D-guided masking and 2D-semantic reconstruction. ficient and diversified semantics compared to colorful im- ages, which constrain the generalization capacity of pre- training. Considering the homology of images and point clouds, both of which depict certain visual characteristics of objects and are related by 2D-3D geometric mapping, we ask the question: can off-the-shelf 2D pre-trained mod- els help 3D representation learning by transferring robust 2D knowledge into 3D domains? To tackle this challenge, we propose I2P-MAE , a Masked Autoencoding framework that conducts Image-to- Point knowledge transfer for self-supervised 3D point cloud pre-training. As shown in Figure 1, aided by 2D semantics learned from abundant image data, our I2P-MAE produces high-quality 3D representations and exerts strong transfer- able capacity to downstream 3D tasks. Specifically, refer- ring to 3D MAE models [47, 78] in Figure 2 (Left), we first adopt an asymmetric encoder-decoder transformer [14] as our fundamental architecture for 3D pre-training, which takes as input a randomly masked point cloud and recon- structs the masked points from the visible ones. Then, to acquire 2D semantics for the 3D shape, we bridge the modal gap by efficiently projecting the point cloud into multi-view depth maps. This requires no time-consuming offline ren- dering and largely preserves 3D geometries from different perspectives. On top of that, we utilize off-the-shelf 2D models to obtain the multi-view 2D features along with 2D attention maps of the point cloud, and respectively guide the pre-training from two aspects, as shown in Figure 2 (Right). Firstly , different from existing methods [47, 78] to ran- domly sample visible tokens, we introduce a 2D-guided masking strategy that reserves point tokens with more spa- tial semantics to be visible for the MAE encoder. In de- tail, we back-project the multi-view attention maps into 3D space as a spatial attention cloud. Each element in the 9493929190898887 93.4%Linear SVM (%)on ModelNet4092.9% 11623691.0%Pre-trainingEpochsFigure 3. Pre-training Epochs vs. Linear SVM Accuracy on ModelNet40 [67]. With the image-to-point learning schemes, I2P- MAE exerts superior transferable capability with much faster con- vergence speed than Point-MAE [47] and Point-M2AE [78]. cloud indicates the semantic significance of the correspond- ing point token. Guided by this, the 3D network can bet- ter focus on the visible critical structures to understand the global 3D shape, and also reconstruct the masked tokens from important spatial cues. Secondly , in addition to the recovering of masked point tokens, we propose to concurrently reconstruct 2D seman- tics from the visible point tokens after the MAE decoder. For each visible token, we respectively fetch its projected 2D representations from different views, and integrate them as the 2D-semantic learning target. By simultaneously re- constructing the masked 3D coordinates and visible 2D con- cepts, I2P-MAE is able to learn both low-level spatial pat- terns and high-level semantics pre-trained in 2D domains, contributing to superior 3D representations. With the aforementioned image-to-point guidance, our I2P-MAE significantly accelerates the convergence speed of pre-training and exhibits state-of-the-art performance on 3D downstream tasks, as shown in Figure 3. Learning from 2D ViT [14] pre-trained by CLIP [53], I2P-MAE, without any fine-tuning, achieves 93.4% classification ac- curacy by linear SVM on ModelNet40 [67], which has surpassed the fully fine-tuned results of Point-BERT [76] and Point-MAE [33]. After fine-tuning, I2P-MAE further achieves 90.11% classification accuracy on the hardest split of ScanObjectNN [63], significantly exceeding the second- best Point-M2AE [78] by +3.68%. The experiments fully demonstrate the effectiveness of learning from pre-trained 2D models for superior 3D representations. Our contributions are summarized as follows: • We propose Image-to-Point Masked Autoencoders (I2P-MAE), a pre-training framework to leverage 2D pre-trained models for learning 3D representations. 21770 • We introduce two strategies, 2D-guided masking and 2D-semantic reconstruction, to effectively transfer the well learned 2D knowledge into 3D domains. • Extensive experiments have been conducted to indicate the significance of our image-to-point pre-training.
Zhang_Towards_Unbiased_Volume_Rendering_of_Neural_Implicit_Surfaces_With_Geometry_CVPR_2023	Abstract Learning surface by neural implicit rendering has been a promising way for multi-view reconstruction in recent years. Existing neural surface reconstruction methods, such as NeuS [24] and VolSDF [32], can produce reliable meshes from multi-view posed images. Although they build a bridge between volume rendering and Signed Distance Function (SDF), the accuracy is still limited. In this pa- per, we argue that this limited accuracy is due to the bias of their volume rendering strategies, especially when the view- ing direction is close to be tangent to the surface. We revise and provide an additional condition for the unbiased vol- ume rendering. Following this analysis, we propose a new rendering method by scaling the SDF field with the angle between the viewing direction and the surface normal vec- tor. Experiments on simulated data indicate that our render- ing method reduces the bias of SDF-based volume render- ing. Moreover, there still exists non-negligible bias when the learnable standard deviation of SDF is large at early stage, which means that it is hard to supervise the rendered depth with depth priors. Alternatively we supervise zero- level set with surface points obtained from a pre-trained Multi-View Stereo network. We evaluate our method on the DTU dataset and show that it outperforms the state-of-the- arts neural implicit surface methods without mask supervi- sion.	1. Introduction 3D reconstruction is an important task in 3D games and AR/VR applications. As a key technique in computer vi- sion and graphics, recovering surfaces and textures from Multi-View calibrated RGB images has been widely studied in recent decades. Early unsupervised Multi-View Stereo (MVS) approaches [14, 20] provide solutions through a *Corresponding authorcertain multistage pipeline, including grouping the related views, depth prediction, filtering with photometric con- sistency and geometry consistency, fusion of points from different views, meshing the dense points by off-the-shelf methods such as screened Poisson Surface Reconstruction [8], and texture mapping finally. Later MVS networks [5,20,27,36] are developed rapidly benefiting from the available large-scale 3D datasets. This kind of MVS networks use Convolutional Neural Network (CNN) to predict depth maps effectively, then follow the traditional pipeline to fuse a global dense point cloud and mesh it. However, MVS networks suffer from texture-less regions and sudden depth changes, so there usually exist many holes in the recovered meshes. Recently, neural implicit surface and differentiable ren- dering methods present a promising way to improve and simplify the progress of the Multi-View 3D reconstruction. The surfaces are represented as Signed Distance Functions (SDF) [18, 24, 32, 33] or occupancy field [16, 17]. At the same time, neural radiance field [13, 35] are proposed with different volume rendering. The neural surface-based ren- dering method can recover reliable and smooth surfaces, but it is hard to train without mask supervision. On the contrary, the different volume rendering can achieve good 2D views without mask supervision, but the quality of 3D geometry is rather coarse. Is there some connections between the SDF field and occupancy field? NeuS [24] and V olSDF [32] point that the connection can be conducted with a certain Cumula- tive Distribution Function (CDF). Thanks to this significant progress, it is able to learn 3D surfaces effectively from neu- ral implicit surface with the self-supervised volume render- ing. The necessary input can only be well-posed 2D images. Masks could be removed, because it is hard to obtain accu- rate masks for many complex objects in the real world. Although these great methods have made big progress on 3D reconstruction from calibrated multi-view images, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4359 the accuracy of meshes is still limited compared with those methods [5, 14, 20, 20, 27, 36] in a classical pipeline. We find that there exist unexpected convex or concave surfaces in regions with poor texture or strong highlight. In addition, the learned surface and rendered color tend to be smooth, and the high frequency details can not be captured well. Based on NeuS [24] and V olSDF [32], we further ana- lyze the precision of the bridge between the SDF field and density field, and we found that there exist a bias between real depth and rendered depth by SDF-based volume ren- dering. We argue that there are two factors leading to the bias: (1) The angle between the view direction and the nor- mal vector; (2) The learnable standard deviation of the SDF field. The bias increases with the growth of the angle. It decreases with the descent of the learnable standard devia- tion, but it is still relatively large when the deviation is not small enough at early training stage. More details of the analysis are described in Section 3. In order to reduce the bias caused by the various angles, we modify the transfor- mation between the SDF field and the density field. Further- more, we adopt dense point clouds predicted by an acces- sible MVS network to reduce the bias further. Finally we evaluate our method and compare it with other SDF-based volume rendering methods on the public benchmark. To summarise, we provides the following three contribu- tions: a) We amend the conditions of unbiased SDF-based volume rendering, and analyze the bias of V olSDF [32] and NeuS [24]. b) We propose a new transformation between the SDF field and the density field, which does not re- quire a plane assumption and outperforms V olSDF [32] and NeuS [24] even without geometry priors. In particular, we scale the SDF field by the inverse of the angle between the view direction and the normal vector. c) Geometry priors from a pre-trained MVS network are used with the anneal- ing sampling to further reduce the bias effectively at early training stage and boost the reconstruction quality.
Yan_SMAE_Few-Shot_Learning_for_HDR_Deghosting_With_Saturation-Aware_Masked_Autoencoders_CVPR_2023	Abstract Generating a high-quality High Dynamic Range (HDR) image from dynamic scenes has recently been extensively studied by exploiting Deep Neural Networks (DNNs). Most DNNs-based methods require a large amount of train- ing data with ground truth, requiring tedious and time- consuming work. Few-shot HDR imaging aims to generate satisfactory images with limited data. However, it is dif- ficult for modern DNNs to avoid overfitting when trained on only a few images. In this work, we propose a novel semi-supervised approach to realize few-shot HDR imag- ing via two stages of training, called SSHDR. Unlikely pre- vious methods, directly recovering content and removing ghosts simultaneously, which is hard to achieve optimum, we first generate content of saturated regions with a self- supervised mechanism and then address ghosts via an itera- tive semi-supervised learning framework. Concretely, con- sidering that saturated regions can be regarded as mask- ing Low Dynamic Range (LDR) input regions, we design a Saturated Mask AutoEncoder (SMAE) to learn a robust fea- ture representation and reconstruct a non-saturated HDR image. We also propose an adaptive pseudo-label selection strategy to pick high-quality HDR pseudo-labels in the sec- ond stage to avoid the effect of mislabeled samples. Ex- periments demonstrate that SSHDR outperforms state-of- the-art methods quantitatively and qualitatively within and across different datasets, achieving appealing HDR visual- ization with few labeled samples.	1. Introduction Standard digital photography sensors are unable to cap- ture the wide range of illumination present in natural scenes, resulting in Low Dynamic Range (LDR) images that often *†The first three authors contributed equally to this work. This work was partially supported by NSFC (U19B2037, 61901384), Natural Science Basic Research Program of Shaanxi (2021JCW-03, 2023-JC-QN-0685), and National Engineering Laboratory for Integrated Aero-Space-Ground- Ocean Big Data Application Technology. Corresponding author: Yu Zhu. Figure 1. The proposed method generates high-quality images with few labeled samples when compared with several methods. suffer from over or underexposed regions, which can dam- age the details of the scene. High Dynamic Range (HDR) imaging has been developed to address these limitations. This technique combines several LDR images with different exposures to generate an HDR image. While HDR imaging can effectively recover details in static scenes, it may pro- duce ghosting artifacts when used with dynamic scenes or hand-held camera scenarios. Historically, various techniques have been suggested to address such issues, such as alignment-based methods [3,10,27,37], patch-based methods [8,15,24], and rejection- based methods [5, 11, 19, 20, 35, 40]. Two categories of alignment-based approaches exist: rigid alignment ( e.g., homographies) that fail to address foreground motions, and non-rigid alignment ( e.g., optical flow) that is error-prone. Patch-based techniques merge similar regions using patch- level alignment and produce superior results, but suffer from high complexity. Rejection-based methods aim to eliminate misaligned areas before fusing images, but may result in a loss of information in motion regions. As Deep Neural Networks (DNNs) become increas- ingly prevalent, the DNN-based HDR deghosting methods This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5775 [9, 33, 36] achieve better visual results compared to tradi- tional methods. However, these alignment approaches are error-prone and inevitably cause ghosting artifacts (see Fig- ure 1 Kalantari’s results). AHDR [31, 32] proposes spatial attention to suppress motion and saturation, which effec- tively alleviate misalignment problems. Based on AHDR, ADNET [14] proposes a dual branch architecture using spatial attention and PCD-alignment [29] to remove ghost- ing artifacts. All these above methods directly learn the complicated HDR mapping function with abundant HDR ground truth data. However, it’s challenging to collect a large amount of HDR-labeled data. The reasons can be at- tributed to that 1) generating a ghost-free HDR ground truth sample requires an absolute static background, and 2) it is time-consuming and requires considerable manpower to do manual post-examination. This generates a new setting that only uses a few labeled data for HDR imaging. Recently, FSHDR [22] attempts to generate a ghost-free HDR image with only few labeled data. They use a prelim- inary model trained with a large amount of unlabeled dy- namic samples, and a few dynamic and static labeled sam- ples to generate HDR pseudo-labels and synthesize artificial dynamic LDR inputs to improve the model performance of dynamic scenes further. This approach expects the model to handle both the saturation and the ghosting problems simul- taneously, but it is hard to achieve under the condition of few labeled data, especially misaligned regions caused by saturation and motion (see Figure 1 FSHDR). In addition, FSHDR uses optical flow to forcibly synthesize dynamic LDR inputs from poorly generated HDR pseudo-labels, the errors in optical flow further intensify the degraded qual- ity of artificial dynamic LDR images, resulting in an appar- ent distribution shift between LDR training and testing data, which hampers the performance of the network. The above analysis makes it very challenging to di- rectly generate a high-quality and ghost-free HDR image with few labeled samples. A reasonable way is to address the saturation problems first and then cope with the ghost- ing problems with a few labeled samples. In this paper, we propose a semi-supervised approach for HDR deghost- ing, named SSHDR, which consists of two stages: self- supervised learning network for content completion and sample-quality-based iterative semi-supervised learning for deghosting. In the first stage, we pretrain a Saturated Mask AutoEncoder (SMAE), which learns the representation of HDR features to generate content of saturated regions by self-supervised learning. Specifically, considering that the saturated regions can be regarded as masking the short LDR input patches, inspired by [6], we randomly mask a high proportion of the short LDR input and expect the model to reconstruct a no-saturated HDR image from the remaining LDR patches in the first stage. This self-supervised ap- proach allows the model to recover the saturated regionswith the capability to effectively learn a robust represen- tation for the HDR domain and map an LDR image to an HDR image. In the second stage, to prevent the overfit- ting problem with a few labeled training samples and make full use of the unlabeled samples, we iteratively train the model with a few labeled samples and a large amount of HDR pseudo-labels from unlabeled data. Based on the pretrained SMAE, a sample-quality-based iterative semi- supervised learning framework is proposed to address the ghosting artifacts. Considering the quality of pseudo-labels is uneven, we develop an adaptive pseudo-labels selection strategy to pick high-quality HDR pseudo-labels ( i.e., well- exposed, ghost-free) to avoid awful pseudo-labels hamper- ing the optimization process. This selection strategy is guided by a few labeled samples and enhances the diver- sity of training samples in each epoch. The experiments demonstrate that our proposed approach can generate high- quality HDR images with few labeled samples and achieves state-of-the-art performance on individual and cross-public datasets. Our contributions can be summarized as follows: • We propose a novel and generalized HDR self-supervised pretraining model, which uses mask strategy to recon- struct an HDR image and addresses saturated problems from one LDR image. • We propose a sample-quality-based semi-supervised training approach to select well-exposed and ghost-free HDR pseudo-labels, which improves ghost removal. • We perform both qualitative and quantitative experi- ments, which show that our method achieves state-of-the- art results on individual and cross-public datasets.

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.