Split (1)

train · 2.34k rows

title stringlengths 28 135	abstract stringlengths 0 12k	introduction stringlengths 0 12k
Liu_Continual_Detection_Transformer_for_Incremental_Object_Detection_CVPR_2023	Abstract Incremental object detection (IOD) aims to train an ob- ject detector in phases, each with annotations for new ob- ject categories. As other incremental settings, IOD is sub- ject to catastrophic forgetting, which is often addressed by techniques such as knowledge distillation (KD) and exem- plar replay (ER). However, KD and ER do not work well if applied directly to state-of-the-art transformer-based ob- ject detectors such as Deformable DETR [59] and UP- DETR [9]. In this paper, we solve these issues by proposing a ContinuaL DEtection TRansformer (CL-DETR), a new method for transformer-based IOD which enables effective usage of KD and ER in this context. First, we introduce a Detector Knowledge Distillation (DKD) loss, focusing on the most informative and reliable predictions from old ver- sions of the model, ignoring redundant background predic- tions, and ensuring compatibility with the available ground- truth labels. We also improve ER by proposing a calibration strategy to preserve the label distribution of the training set, therefore better matching training and testing statistics. We conduct extensive experiments on COCO 2017 and demon- strate that CL-DETR achieves state-of-the-art results in the IOD setting.1	1. Introduction Humans inherently learn in an incremental manner, ac- quiring new concepts over time without forgetting previ- ous ones. In contrast, machine learning suffers from catas- trophic forgetting [21, 35, 36], where learning from non- i.i.d. data can override knowledge acquired previously. Un- surprisingly, forgetting also affects object detection [2, 12, 20, 37, 44, 50, 54]. In this context, the problem was formal- ized by Shmelkov et al. [44], who defined an incremental object detection (IOD) protocol, where the training samples for different object categories are observed in phases, re- 1Code: https://lyy.mpi-inf.mpg.de/CL-DETR/ Old categories All categories36384042Average Precision (%) w/ ER+KD w/ Ours Upper BoundFigure 1. The final Average Precision (AP, %) of two-phase incre- mental object detection on COCO 2017. We observe 70and10 categories in the first and second phases, respectively. The base- line is Deformable DETR [59]. “Upper bound” shows the results of joint training with all previous data accessible in each phase. stricting the ability of the trainer to access past data. Popular methods to address forgetting in tasks other than detection include Knowledge Distillation (KD) and Exem- plar Replay (ER). KD [11,16,17,26,57] uses regularization in an attempt to preserve previous knowledge when train- ing the model on new data. The key idea is to encourage the new model’s logits or feature maps to be close to those of the old model. ER methods [5, 29, 32, 33, 41, 52] work instead by memorising some of the past training data (the exemplars ), replaying them in the following phases to “re- member” the old object categories. Recent state-of-the-art results in object detection have been achieved by a family of transformer-based architec- tures that include DETR [4], Deformable DETR [59] and UP-DETR [9]. In this paper, we show that KD and ER do not work well if applied directly to these models. For instance, in Fig. 1 we show that applying KD and ER to Deformable DETR leads to much worse results compared to training with all data accessible in each phase ( i.e., the standard non-incremental setting). We identify two main issues that cause this drop in per- 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. 23799 formance. First, transformer-based detectors work by test- ing a large number of object hypotheses in parallel. Because the number of hypotheses is much larger than the typical number of objects in an image, most of them are negative, resulting in an unbalanced KD loss. Furthermore, because both old and new object categories can co-exist in any given training image, the KD loss and regular training objective can provide contradictory evidence. Second, ER methods for image classification try to sample the same number of exemplars for each category. In IOD, this is not a good strat- egy because the true object category distribution is typically highly skewed. Balanced sampling causes a mismatch be- tween the training and testing data statistics. In this paper, we solve these issues by proposing Con- tinuaL DEtection TRansformer (CL-DETR), a new method for transformer-based IOD which enables effective usage of KD and ER in this context. CL-DETR introduces the con- cept of Detector Knowledge Distillation (DKD), selecting the most confident object predictions from the old model, merging them with the ground-truth labels for the new cate- gories while resolving conflicts, and applying standard joint bipartite matching between the merged labels and the cur- rent model’s predictions for training. This approach sub- sumes the KD loss, applying it only for foreground predic- tions correctly matched to the appropriate model’s hypothe- ses. CL-DETR also improves ER by introducing a new cal- ibration strategy to preserve the distribution of object cat- egories observed in the training data. This is obtained by carefully engineering the set of exemplars remembered to match the desired distribution. Furthermore, each phase consists of a main training step followed by a smaller one focusing on better calibrating the model. We also propose a more realistic variant of the IOD benchmark protocol. In previous works [12, 44], in each phase, the incremental detector is allowed to observe all im- ages that contain a certain type of object. Because images often contain a mix of object classes, both old and new, this means that the same images can be observed in different training phases. This is incompatible with the standard def- inition of incremental learning [16, 33, 41] where, with the exception of the examples deliberately stored in the exem- plar memory, the images observed in different phases do not repeat. We redefine the IOD protocol to avoid this issue. We demonstrate CL-DETR by applying it to dif- ferent transformer-based detectors including Deformable DETR [59] and UP-DETR [9]. As shown in Fig. 1, our results on COCO 2017 show that CL-DETR leads to signif- icant improvements compared to the baseline, boosting AP by4.2percentage points compared to a direct application of KD and ER to the underlying detector model. We further study and justify our modelling choices via ablations. To summarise, we make four contributions : (1) The DKD loss that improves KD for knowledge distillation byresolving conflicts between distilled knowledge and new ev- idence and by ignoring redundant background detections; (2) A calibration strategy for ER to match the stored ex- emplars to the training set distribution; (3) A revised IOD benchmark protocol that avoids observing the same images in different training phases; (4) Extensive experiments on COCO 2017, including state-of-the-art results, an in-depth ablation study, and further visualizations.
Ling_PanoSwin_A_Pano-Style_Swin_Transformer_for_Panorama_Understanding_CVPR_2023	Abstract In panorama understanding, the widely used equirectan- gular projection (ERP) entails boundary discontinuity and spatial distortion. It severely deteriorates the conventional CNNs and vision Transformers on panoramas. In this pa- per, we propose a simple yet effective architecture named PanoSwin to learn panorama representations with ERP . To deal with the challenges brought by equirectangular projec- tion, we explore a pano-style shift windowing scheme and novel pitch attention to address the boundary discontinu- ity and the spatial distortion, respectively. Besides, based on spherical distance and Cartesian coordinates, we adapt absolute positional embeddings and relative positional bi- ases for panoramas to enhance panoramic geometry infor- mation. Realizing that planar image understanding might share some common knowledge with panorama understand- ing, we devise a novel two-stage learning framework to facilitate knowledge transfer from the planar images to panoramas. We conduct experiments against the state-of- the-art on various panoramic tasks, i.e., panoramic object detection, panoramic classification, and panoramic layout estimation. The experimental results demonstrate the effec- tiveness of PanoSwin in panorama understanding.	1. Introduction Panoramas are widely used in many real applications, such as virtual reality, autonomous driving, civil surveil- lance, etc. Panorama understanding has attracted increas- ing interest in the research community [5, 27, 34]. Among these methods, the most popular and convenient represen- tation of panorama is adopted via equirectangular projec- tion (ERP), which maps the latitude and longitude of the spherical representation to horizontal and vertical grid co- ordinates. However, the inherent omnidirectional vision re- mains the challenge of the panorama understanding. Al- though convolutional neural networks (CNNs) [11, 14, 28] have shown outstanding performances on planar image un- derstanding, most CNN-based methods are unsuitable for panoramas because of two fundamental problems entailed a. original windowsdistant in ERP but close in sphere pitch attention b. shifted windows c. rotated panoramashift Figure 1. (1). Fig. ais how a panoramic image looks, just like a planar world map, where top/bottom regions are connected to the earth’s poles; the right side is connected to the left. (2). Our PanoSwin is based on window attention [19]. Fig. aalso shows the original window partition in dotted orange, where the two windows in bold orange are separated by equirectangular projec- tion(ERP). (3). Fig. bshows pano-style shift windowing scheme, which brings the two departed regions together. (4). Fig. cshows our pitch attention module, which helps a distorted window to in- teract with an undistorted one. by ERP: (1)polar and side boundary discontinuity and(2) spatial distortion . Specifically, the north/south polar region in spherical representations are closely connected. But the converted region covers the whole top/bottom boundaries. On this account, polar boundary continuity is destroyed by ERP. Similarly, side boundary continuity is also destroyed since the left and right sides are split by ERP. Meanwhile, spatial distortion also severely deforms the image content, especially in polar regions. A common solution is to adapt convolution to the spher- ical space [4, 5, 24, 34]. However, these methods might suf- fer from high computation costs from the adaptation pro- cess. Besides, Spherical Transformer [2] and PanoFormer [22] specially devise patch sampling approaches to remove 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. 17755 panoramic distortion. However, the specially designed patch sampling approaches might not be feasible for planar images. In our experiments, we demonstrate that exploiting planar knowledge can boost the performance of panorama understanding. Inspired by Swin Transformer [19], we propose PanoSwin Transformer to reduce the distortion of panoramic images, as briefly shown in Fig. 1. To cope with boundary discontinuity , we explore a pano-style shift win- dowing scheme (PSW). In PSW, side continuity is estab- lished by horizontal shift. To establish polar continuity, we first split the panorama in half and then rotate the right half counterclockwise. To overcome spatial distortion , we first rotate the pitch of the panorama by 0.5π. So the polar re- gions of the original feature map are “swapped” with some equator regions of the rotated panorama. For each win- dow in the original panorama, we locate a corresponding window in the rotated panorama. Then we perform cross- attention between these two windows. We name the module pitch attention (PA), which is plug-and-play and can be in- serted in various backbones. Intuitively, pitch attention can help a window “know” how it looks without distortion. To leverage planar knowledge, some works [24, 25] proposed to make novel panoramic kernel mimick out- puts from planar convolution kernel layer by layer. How- ever, PanoSwin is elaborately designed to be compati- ble with planar images: PanoSwin can be switched from pano mode to vanilla swin mode . Let PanoSwin in these two modes be denoted as PanoSwin pand PanoSwin s. PanoSwin p/PanoSwin scan be adopted to process panora- mas/planar images, details about which will be introduced in Sec. 3.6. In our paper, PanoSwin is under pano mode by default. The double-mode feature of PanoSwin makes it possible to devise a simple two-stage learning paradigm based on knowledge preservation to leverage planar knowl- edge: we first pretrain PanoSwin swith planar images; then we switch it to PanoSwin pand train it with a knowledge preservation (KP) loss and downstream task losses. This paradigm is able to facilitate transferring common visual knowledge from planar images to panoramas. Our main contributions are summarized as follows: (1) We propose PanoSwin to learn panorama features, in which Pano-style Shift Windowing scheme (PSW) is proposed to resolve polar and side boundary discontinuity; (2)we pro- pose pitch attention module (PA) to overcome spatial dis- tortion introduced by ERP; (3)PanoSwin is designed to be compatible with planar images. Therefore, we proposed a KP-based two-stage learning paradigm to transfer common visual knowledge from planar images to panoramas; (4)we conduct experiments on various panoramic tasks, including panoramic object detection, panoramic classification, and panoramic layout estimation on five datasets. The results have validated the effectiveness of our proposed method.
Li_LoGoNet_Towards_Accurate_3D_Object_Detection_With_Local-to-Global_Cross-Modal_Fusion_CVPR_2023	Abstract LiDAR-camera fusion methods have shown impressive performance in 3D object detection. Recent advancedmulti-modal methods mainly perform global fusion, whereimage features and point cloud features are fused across the whole scene. Such practice lacks ﬁne-grained region-level information, yielding suboptimal fusion performance. In this paper , we present the novel Local-to-Global fusionnetwork (LoGoNet), which performs LiDAR-camera fusionat both local and global levels. Concretely, the Global Fu-sion (GoF) of LoGoNet is built upon previous literature,while we exclusively use point centroids to more preciselyrepresent the position of voxel features, thus achieving bet-ter cross-modal alignment. As to the Local Fusion (LoF),we ﬁrst divide each proposal into uniform grids and thenproject these grid centers to the images. The image featuresaround the projected grid points are sampled to be fusedwith position-decorated point cloud features, maximally uti-lizing the rich contextual information around the proposals.The Feature Dynamic Aggregation (FDA) module is furtherproposed to achieve information interaction between theselocally and globally fused features, thus producing more informative multi-modal features. Extensive experimentson both Waymo Open Dataset (WOD) and KITTI datasetsshow that LoGoNet outperforms all state-of-the-art 3D de- tection methods. Notably, LoGoNet ranks 1st on Waymo 3D object detection leaderboard and obtains 81.02 mAPH (L2) detection performance. It is noteworthy that, for theﬁrst time, the detection performance on three classes sur-passes 80 APH (L2) simultaneously. Code will be availableathttps://github.com/sankin97/LoGoNet . *Corresponding author	1. Introduction 3D object detection, which aims to localize and clas- sify the objects in the 3D space, serves as an essentialperception task and plays a key role in safety-critical au-tonomous driving [ 1,20,58]. LiDAR and cameras are two widely used sensors. Since LiDAR provides accu-rate depth and geometric information, a large number ofmethods [ 24,48,63,68,72,73] have been proposed and achieve competitive performance in various benchmarks.However, due to the inherent limitation of LiDAR sensors, point clouds are usually sparse and cannot provide sufﬁcientcontext to distinguish between distant regions, thus causingsuboptimal performance. To boost the performance of 3D object detection, a nat- ural remedy is to leverage rich semantic and texture in-formation of images to complement the point cloud. As shown in Fig. 1(a), recent advanced methods introduce the global fusion to enhance the point cloud with image fea-tures [ 2,5,7,8,22,23,25,27,34,54,55,60,69,71]. They typically fuse the point cloud features with image featuresacross the whole scene. Although certain progress has beenachieved, such practice lacks ﬁne-grained local informa- tion. For 3D detection, foreground objects only account fora small percentage of the whole scene. Merely performing global fusion brings marginal gains. To address the aforementioned problems, we propose a novel Local-to-Global fusion Network, termed LoGoNet,which performs LiDAR-camera fusion at both global andlocal levels, as shown in Fig. 1(b). Our LoGoNet is comprised of three novel components, i.e., Global Fusion (GoF), Local Fusion (LoF) and Feature Dynamic Aggrega-tion (FDA). Speciﬁcally, our GoF module is built on previ- ous literature [ 8,25,34,54,55] that fuse point cloud features and image features in the whole scene, where we use thepoint centroid to more accurately represent the position 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. 17524 Global Fusion Local FusionSpatial Location Guidance Instance Semantic Guidancefuse fuse Point Cloud Image ROI Point Cloud ROI Image (b)(a) (c)75.5475.6776.29 76.3377.1077.6478.41 78.4579.6079.94 79.9781.02 7576777879808182mAPH (L2) Local-to-Global fusion (Ours) Global fusion LiDAR-only (c)* * * * * * * *Test time augmentations or ensemble Time Figure 1. Comparison between (a) global fusion and (b) local fusion. Global fusion methods perform fusion of point cloud features and image features across the whole scene, which lacks ﬁne-grained region-level information. The proposed local fusion method fuses featuresof two modalities on each proposal, complementary to the global fusion methods. (c) Performance comparison of various methods in Waymo 3D detection leaderboard [ 51]. Our LoGoNet attains the top 3D detection performance, clearly outperforming all state-of-the-art global fusion based and LiDAR-only detectors. Please refer to Table 1for a detailed comparison with more methods. each voxel feature, achieving better cross-modal alignment. And we use the global voxel features localized by point cen-troids to adaptively fuse image features through deformable cross-attention [ 75] and adopt the ROI pooling [ 9,48]t o generate the ROI-grid features. To provide more ﬁne-grained region-level information for objects at different distances and retain the original po-sition information within a much ﬁner granularity, we pro-pose the Local Fusion (LoF) module with the Position In-formation Encoder (PIE) to encode position information of the raw point cloud in the uniformly divided grids of eachproposal and project the grid centers onto the image plane tosample image features. Then, we fuse sampled image fea-tures and the encoded local grid features through the cross-attention [ 53] module. To achieve more information inter- action between globally fused features and locally fusedROI-grid features for each proposal, we propose the FDAmodule through self-attention [ 53] to generate more infor- mative multi-modal features for second-stage reﬁnement. Our LoGoNet achieves superior performance on two 3D detection benchmarks, i.e., Waymo Open Dataset (WOD) and KITTI datasets. Notably, LoGoNet ranks 1st on Waymo 3D object detection leaderboard and obtains 81.02 mAPH (L2) detection performance. Note that, for the ﬁrst time, thedetection performance on three classes surpasses 80 APH(L2) simultaneously. The contributions of our work are summarized as fol- lows: • We propose a novel local-to-global fusion network,termed LoGoNet , which performs LiDAR-camera fu- sion at both global and local levels. • Our LoGoNet is comprised of three novel components, i.e., GoF, LoF and FDA modules. LoF provides ﬁne- grained region-level information to complement GoF.FDA achieves information interaction between glob-ally and locally fused features, producing more infor-mative multi-modal features. • LoGoNet achieves state-of-the-art performance on WOD and KITTI datasets. Notably, our Lo- GoNet ranks 1st on Waymo 3D detection leaderboard with 81.02 mAPH (L2).
Li_Metadata-Based_RAW_Reconstruction_via_Implicit_Neural_Functions_CVPR_2023	Abstract Many low-level computer vision tasks are desirable to utilize the unprocessed RAW image as input, which remains the linear relationship between pixel values and scene ra- diance. Recent works advocate to embed the RAW im- age samples into sRGB images at capture time, and recon- struct the RAW from sRGB by these metadata when needed. However, there still exist some limitations in making full use of the metadata. In this paper, instead of following the perspective of sRGB-to-RAW mapping, we reformulate the problem as mapping the 2D coordinates of the meta- data to its RAW values conditioned on the corresponding sRGB values. With this novel formulation, we propose to reconstruct the RAW image with an implicit neural function, which achieves significant performance improvement (more than 10dB average PSNR) only with the uniform sampling. Compared with most deep learning-based approaches, our method is trained in a self-supervised way that requiring no pre-training on different camera ISPs. We perform further experiments to demonstrate the effectiveness of our method, and show that our framework is also suitable for the task of guided super-resolution.	1. Introduction Low-level computer vision tasks benefit a lot from the scene-referred RAW images [7, 39, 19, 17, 16], which is rendered to the display-referred standard RGB (sRGB) im- ages via camera image signal processors (ISPs). Compared with sRGB images, typical RAW images has the advantages of linear relationship between pixel values and scene ra- diance, as well as higher dynamic range. However, RAW images occupy obviously more memory than the sRGB im- ages in common format like JPEG, which is unfavourable for transferring and sharing. Moreover, since most dis- play and printing devices are designed for images stored and shared in sRGB format, it is inconvenient to directly re- place sRGB with RAW. Consequently, mapping sRGB im- ages back to their RAW counterparts, which is also called RAW reconstruction, is regarded as the appropriate way to PSNR: 49.15dB PSNR: 46.72dB 1% sRGB (input) RIR [25] SAM [28] PSNR: 51.12dB PSNR: 63.59dB 1% RAW (GT) CAM [24] Ours Figure 1. As RAW images are beneficial to many low-level com- puter vision tasks, we aim to reconstruct the RAW image from the corresponding sRGB image with the assistance of extra meta- data. In this figure, the reconstructed RAW images are visual- ized through error maps. As can be seen, our method remarkably outperforms other related methods with the improvement of more than 10 dB PSNR. We owe this performance boost to the effec- tiveness of our proposed implicit neural function (INF) . utilize the advantage of RAW data [23, 25, 36, 10, 28, 24]. Early RAW reconstruction methods focus on building standard models to reverse ISPs, which is parameterized by either explicit functions [4, 18, 14, 5] or neural net- works [23, 36, 10]. However, these approaches are faced with the same issue that a parameterized model is only suitable for a specific ISP. Meanwhile, a series of meth- ods [25, 26, 28, 24] propose to overcome this problem by embedding extra metadata into sRGB images at capture time. For such methods, the main challenge is to improve the accuracy with lower metadata generation cost. RIR [25] implements complex optimization algorithm to estimate the global mapping parameters as metadata, but suffers high computational cost. SAM [28] adopts a uniform sampling on RAW images to generate the metadata, which is further replaced with a sampler network by CAM [24]. For the metadata-based methods of SAM [28] and CAM [24], the embedded RAW samples stores partial infor- mation of ISPs which helps to reconstruct the RAW images better; also, by conditioning the reconstruction algorithm on 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. 18196 these metadata, the recovery of the RAW data turns into a conditional mapping function instead of a function fitted to a specific case, enabling the potential to achieve better gen- eralization. Therefore, we adopt this strategy in the paper. Despite the progress that SAM [28] and CAM [24] have made, there still exist some limitations for the metadata- based RAW reconstruction methods. SAM utilizes RBF in- terpolation [3], the main idea of which is to calculate the difference between sampling and target points by a kernel function. However, a fixed kernel function lacks the flexi- bility to model various sRGB-to-RAW mappings. CAM di- rectly uses a neural network for reconstruction but requires pre-training on pairs of sRGB and raw data from different types of ISPs. Also, we observe that the results of these methods fail to recover the saturated regions [26] (i.e., pix- els with any channel value close to the maximum), as is shown in Figure 1. To address the limitation, we propose a two-way RAW reconstruction algorithm based on an implicit neural func- tion (INF). Previously, RAW reconstruction is formulated as mapping a sRGB image and the metadata to its RAW im- age. In this paper, we reformulate the problem as mapping the 2D coordinates of the metadata to its RAW values con- ditioned on the corresponding sRGB values, i.e. an implicit function. With this novel formulation, we can also decom- pose the problem into two aspects: a mapping function from the sRGB values to the corresponding raw values; a super- resolution function to interpolate the RAW image from the sparse samples. We observe that the super-resolution part usually exhibits much higher errors, indicating the latter a more challenging task. Accordingly, two branches are de- signed for each task inside an implicit neural network and the hyper-parameters for these branches are tuned to accom- modate the difficulty of the tasks. Also, notice that with this formulation, the network can be trained in a self-supervised way, without the need of corresponding RAW images. Our contribution can be summarized as follows: • We reformulate the RAW reconstruction problem as a RAW image approximation problem that learns the 2D-to-RAW mapping of image coordinates to RAW values conditioned on its sRGB image. • We decompose the reconstruction into two aspects and design the implicit neural network accordingly. • We conduct extensive experiments on different cam- eras and demonstrate our algorithm outperforms exist- ing work significantly.
Kawahara_Teleidoscopic_Imaging_System_for_Microscale_3D_Shape_Reconstruction_CVPR_2023	Abstract This paper proposes a practical method of microscale 3D shape capturing by a teleidoscopic imaging system. The main challenge in microscale 3D shape reconstruction is to capture the target from multiple viewpoints with a large enough depth-of-ﬁeld. Our idea is to employ a teleidoscopic measurement system consisting of three planar mirrors and monocentric lens. The planar mirrors virtually deﬁne mul- tiple viewpoints by multiple reﬂections, and the monocen- tric lens realizes a high magniﬁcation with less blurry and surround view even in closeup imaging. Our contributions include, a structured ray-pixel camera model which han- dles refractive and reﬂective projection rays efﬁciently, an- alytical evaluations of depth of ﬁeld of our teleidoscopic imaging system, and a practical calibration algorithm of the teleidoscopic imaging system. Evaluations with real im- ages prove the concept of our measurement system.	1. Introduction Microscale 3D reconstruction has found profound appli- cations in a wide range of domains including medical imag- ing, life science, and aquaculture, due to its non-constrained and non-invasive measurements. The main challenges in image-based microscopic 3D shape measurement is its shal- low depth of ﬁeld and camera arrangement in the closeup scenario. Applying conventional multiple camera system designed for human-size capture [ 14,30] cannot be a feasi- ble solution due to limitations on camera placement. Con-ventional multiple mirror system [ 33] also have difﬁculties inevitably in depth-of-focus due to differences in their opti- cal path lengths with varying numbers of bounces. In this paper, we show that the fuller 3D shape of a mi- croscale object can be recovered. Our key idea is to employ a catadioptric imaging system which realizes a practical closeup multi-view imaging. The point of our design is that the system has a monocentric front lens like a teleidoscope, instead of using microscopic system in the camera side. That is, as shown in Fig. 1(a), we introduce a kaleidoscopic multi-facet mirror between the front lens and the camera, where the design realizes a deeper depth-of-ﬁeld and results in less blurring imaging. Unlike conventional microscopic imaging system such as differential phase contrast (DPC) microscopy [ 4,34] and multi-focus approaches [ 15,24,26], our method realizes a multi-view capture of the target from a single physical viewpoint which can contribute to free- viewpoint rendering, 3D shape reconstruction, and reﬂec- tion analysis. Our system also has an advantage over ex- isting imaging systems that build multiple views behind the main lens [ 7,21] in a closeup environment. The wide FoV with a monocentric main lens and virtual multiple views with mirrors allow closeup and surrounding view capturing in focus. We call our system teleidoscopic imaging system and show that the system can be compactly modeled by a struc- tured ray-pixel camera model [ 11], which handles refractive and reﬂective projection rays efﬁciently. Based on our ray- pixel camera model, we derive a practical calibration algo- 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. 20813 rithm that estimates the positions of the monocentric lens and multi-facet mirror w.r.t. the camera by using a single reference planar pattern ( i.e., a checkerboard). Given the calibration parameters, a scene point can then be linearly triangulated from its teleidoscopic projection in a direct lin- ear transform (DLT) manner [ 13]. We implement our method with an imaging system consisting of a camera, three planar mirrors, a monocen- tric lens, and a projector placed outside the mirrors for structured-light casting. We quantitatively evaluate the computation cost of numerical projections, the robustness of the calibration, and the depth of ﬁeld of our teleido- scopic imaging system. We also validate the effectiveness of our method qualitatively on a number of real-world mi- croscale objects. These results demonstrate that our method enables holistic and dense reconstruction of microscale ob- jects. We believe our method expands the avenues of three- dimensional analysis of microscale objects and scenes in real world scenarios.
Li_DynaMask_Dynamic_Mask_Selection_for_Instance_Segmentation_CVPR_2023	Abstract The representative instance segmentation methods mostly segment different object instances with a mask of the ﬁxed resolution, e.g., 2828grid. However, a low- resolution mask loses rich details, while a high-resolution mask incurs quadratic computation overhead. It is a chal- lenging task to predict the optimal binary mask for each in- stance. In this paper, we propose to dynamically select suit- able masks for different object proposals. First, a dual-level Feature Pyramid Network (FPN) with adaptive feature ag- gregation is developed to gradually increase the mask grid resolution, ensuring high-quality segmentation of objects. Speciﬁcally, an efﬁcient region-level top-down path (r-FPN) is introduced to incorporate complementary contextual and detailed information from different stages of image-level FPN (i-FPN). Then, to alleviate the increase of computa- tion and memory costs caused by using large masks, we de- velop a Mask Switch Module (MSM) with negligible compu- tational cost to select the most suitable mask resolution for each instance, achieving high efﬁciency while maintaining high segmentation accuracy. Without bells and whistles, the proposed method, namely DynaMask, brings consistent and noticeable performance improvements over other state-of- the-arts at a moderate computation overhead. The source code: https://github.com/lslrh/DynaMask .	1. Introduction Instance segmentation (IS) is an important computer vi- sion task, aiming at simultaneously predicting the class la- bel and the binary mask for each instance of interest in an image. It works as the cornerstone of many downstream vision applications, such as autonomous driving, video surveillance, and robotics, to name a few. Recent years have witnessed the signiﬁcant advances of deep convolu- tional neural networks (CNNs) based IS methods with a rich amount of training data as the rocket fuel [1, 10, 17, 25, 26]. Existing IS methods can be roughly divided into two ma- *denotes the equal contribution, †denotes the corresponding author. This work is supported by the Hong Kong RGC RIF grant (R5001-18). Figure 1. Dynamic mask selection results. Some “hard” sam- ples with irregular shapes like “person” are assigned larger masks, while the “easy” ones like “frisbee” are assigned smaller ones. jor categories: two-stage [6, 10, 17] and single-stage meth- ods [1, 2, 26]. The former ﬁrst detect a sparse set of pro- posals and then performs mask predictions based on them, while the latter directly predict classiﬁcation scores and masks based on the pre-deﬁned anchors. Generally speak- ing, two-stage methods could produce higher segmenta- tion accuracy but cost more computational resources than single-stage methods. Among the many recently developed IS methods, the proposal-based two-stage methods [6, 10, 17], which fol- low a detection-and-segmentation paradigm, are still the top performers. These methods need to predict a binary grid mask of uniform resolution for all proposals, e.g.,2828, and then upsample it to the original image size. For in- stance, Mask R-CNN [10] ﬁrst generates a group of propos- als with an object detector and then performs per pixel fore- ground/background segmentation on the Regions of Inter- est (RoIs) [24]. Despite achieving promising performance, the low-resolution mask of Mask R-CNN is insufﬁcient to capture more detailed information, resulting in unsatisfac- tory predictions, especially over object boundaries. An in- tuitive solution to improve the segmentation quality is to adopt a larger mask. Nevertheless, high-resolution masks often generate excessive predictions on the smooth regions, resulting in high computational complexity. It is difﬁcult to segment different objects in an image 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. 11279 with masks of the same resolution. Objects with irregu- lar shapes and complicated boundaries demand more ﬁne- grained masks to predict, referred to as “hard” samples, such as the “person” in Fig. 1. In comparison, the “easy” samples with regular shapes and fewer details can be efﬁ- ciently segmented using coarser masks, like the “frisbee” in Fig. 1. Inspired by the above observations, we propose to adaptively adjust the mask size for each instance for bet- ter IS performance. Speciﬁcally, instead of using a uniform resolution for all instances, we assign high-resolution masks to “hard” objects and low-resolution masks to “easy” ob- jects. In this way, the redundant computation for “easy” samples can be reduced while the accuracy of “hard” sam- ples can be improved, achieving a balance between accu- racy and speed. As shown in Tab. 1, however, directly pre- dicting a high-resolution mask by Mask R-CNN [10] unex- pectedly degrades the mask average precision (AP). This at- tributes to two main reasons. First, the RoI features of larger objects are extracted from higher pyramid levels, which are very coarse due to the downsampling operations [20]. Thus simply increasing the mask size of these RoIs will not bring extra useful information. Second, the mask head of Mask R-CNN is oversimpliﬁed, so it cannot make more precise predictions as the mask grid size increases. To overcome the above mentioned problems, we propose a dual-level FPN framework to gradually enlarge the mask grid. Speciﬁcally, in addition to traditional image-level FPN (i-FPN), a region-level FPN (r-FPN) is designed to achieve coarse-to-ﬁne mask prediction. Wherein we construct infor- mation ﬂows between i-FPN and r-FPN at different pyra- mid levels, aiming to incorporate complementary contex- tual and detailed information from multiple feature levels for high-quality segmentation. With the dual-level FPN, we present a data-dependent Mask Switch Module (MSM) with negligible computational cost to adaptively select masks for each instance. The overall approach, namely DynaMask, is evaluated on benchmark instance segmentation datasets to demonstrate its superiority over state-of-the-arts. The major contributions of this work are summarized as follows: A dynamic mask selection method (DynaMask) is pro- posed to adaptively assign appropriate masks to dif- ferent instances. Speciﬁcally, it assigns low-resolution masks to “easy” samples for efﬁciency while assigning high-resolution masks to “hard” samples for accuracy. A dual-level FPN framework is developed for IS. We construct direct information ﬂows from i-FPN to r- FPN at multiple levels, facilitating complementary in- formation aggregation from multiple pyramid levels. Extensive experiments demonstrate that DynaMask achieves a good trade-off between IS accuracy and efﬁ- ciency, outperforming many state-of-the-art two-stage IS methods at a moderate computation overhead.Method Resolution AP FLOPs Mask R-CNN [10]1414 32.9 0.2G 2828 34.7 0.5G 5656 33.8 2.0G 112112 32.5 8.0G DynaMask1414 32.9 0.2G 2828 36.1 0.6G 5656 37.1 1.0G 112112 37.6 1.4G Table 1. Mask AP and FLOPs with different mask resolutions. For Mask R-CNN, directly increasing the mask resolution will de- crease the mask AP. While for our DynaMask, higher mask reso- lution results in better performance.
Li_DSFNet_Dual_Space_Fusion_Network_for_Occlusion-Robust_3D_Dense_Face_CVPR_2023	Abstract Sensitivity to severe occlusion and large view angles lim- its the usage scenarios of the existing monocular 3D dense face alignment methods. The state-of-the-art 3DMM-based method, directly regresses the model’s coefficients, under- utilizing the low-level 2D spatial and semantic information, which can actually offer cues for face shape and orienta- tion. In this work, we demonstrate how modeling 3D facial geometry in image and model space jointly can solve the oc- clusion and view angle problems. Instead of predicting the whole face directly, we regress image space features in the visible facial region by dense prediction first. Subsequently, we predict our model’s coefficients based on the regressed feature of the visible regions, leveraging the prior knowl- edge of whole face geometry from the morphable models to complete the invisible regions. We further propose a fusion network that combines the advantages of both the image and model space predictions to achieve high robustness and accuracy in unconstrained scenarios. Thanks to the pro- posed fusion module, our method is robust not only to occlu- sion and large pitch and roll view angles, which is the bene-fit of our image space approach, but also to noise and large yaw angles, which is the benefit of our model space method. Comprehensive evaluations demonstrate the superior per- formance of our method compared with the state-of-the-art methods. On the 3D dense face alignment task, we achieve 3.80% NME on the AFLW2000-3D dataset, which outper- forms the state-of-the-art method by 5.5% . Code is avail- able at https://github.com/lhyfst/DSFNet .	1. Introduction 3D dense face alignment is an important prob- lem with many applications, e.g. video conferencing, AR/VR/metaverse, games, facial analysis, etc. Many meth- ods have been proposed [8–12, 16, 21, 26, 27, 29, 33, 39, 41, 47, 49]. However, these methods are sensitive to severe oc- clusion and large view angles [19, 26, 31, 32], limiting their applicability of 3D dense face alignment on wild images where occlusion and view angles often occur. 3D dense face alignment from a single image is an ill- posed problem, mainly because of the depth ambiguity. 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. 4531 (a) (i)(c)(b) 3DMM coefficients (ii) (iii)Figure 2. In this case, only one eye is visible. (a) The 3DMM- based method fails. (b) Face parsing algorithm [43] still works. (c) Our method first (i) predicts reliable geometry in visible region by dense prediction, then (ii) completes the whole face by facial geometry prior, producing a reasonable result. (iii) Viewed in im- age view. existing methods [8,10,27,29,33] use a contractive CNN to predict the coefficients of 3DMM [4] directly. However, contractive CNNs are essentially ill-suited for this task [18] due to some reasons, including: mixed depth and 2D spatial information, the loss of low-level 2D spatial information as a result of invariance attribute of CNNs, and mixed facial region and occluder in the process of the contraction. Severe occlusion and large view angles pose problems due to the complexity of the many-to-one mapping from 2D image to 3D shape. In contrast, low-level vision features are less variant according to geometry transform. There- fore, dense prediction is essentially more robust to the above problem in the visible region, because dense prediction re- lies more on local information, where an example is shown in Fig. 2 (b). Even if most of the face is masked out and only the left eye is visible, the face parsing algorithm is still able to deduce a reasonable parsing result. Based on this observation, we decentralize the instance- level 3DMM coefficients regression (i.e., whole-face level) to pixel-level dense prediction in image space to improve the robustness against occlusion and large view angles, by proposing a 3D facial geometry’s 2D image space repre- sentation. To complete the invisible region due to extra- or self-occlusion, a novel post-process algorithm is proposed to convert the dense prediction for the visible face region into 3D facial geometry that includes the whole face area. Fig. 2 (c) shows that our image space prediction recovers reasonable results only seeing one eye, while the SOTA method fails to produce a reasonable result. We further compare the robustness and accuracy be- tween the image space prediction with the model space prediction that directly regresses 3DMM’s coefficients, and discover that there is a complementary relationship between these two spaces. Thus, we propose a dual space fusion network (DSFNet) that predicts using the image and model spaces using a two-branch architecture. With the fusionmodule, our DSFNet effectively combines the advantages of both spaces. In summary, the main contributions of this paper are: • We propose a novel 3D facial geometry’s 2D im- age space representation, followed by a novel post- processing algorithm. It achieves robust 3D dense face alignment to occlusion and large view angles. • We introduce a fusion network, which combines the advantages of both the image and model space predic- tions to achieve high robustness and accuracy in un- constrained scenarios. • On the 3D dense face alignment task, we achieve 3.80% NME on AFLW2000-3D dataset, which out- performs the state-of-the-art method by 5.5% .
Kim_Sampling_Is_Matter_Point-Guided_3D_Human_Mesh_Reconstruction_CVPR_2023	Abstract This paper presents a simple yet powerful method for 3D human mesh reconstruction from a single RGB image. Most recently, the non-local interactions of the whole mesh vertices have been effectively estimated in the transformer while the relationship between body parts also has begun to be handled via the graph model. Even though those ap- proaches have shown the remarkable progress in 3D hu- man mesh reconstruction, it is still difficult to directly infer the relationship between features, which are encoded from the 2D input image, and 3D coordinates of each vertex. To resolve this problem, we propose to design a simple fea- ture sampling scheme. The key idea is to sample features in the embedded space by following the guide of points, which are estimated as projection results of 3D mesh ver- tices (i.e., ground truth). This helps the model to concen- trate more on vertex-relevant features in the 2D space, thus leading to the reconstruction of the natural human pose. Furthermore, we apply progressive attention masking to precisely estimate local interactions between vertices even under severe occlusions. Experimental results on bench- mark datasets show that the proposed method efficiently im- proves the performance of 3D human mesh reconstruction. The code and model are publicly available at: https: //github.com/DCVL-3D/PointHMR_release .	1. Introduction The goal of 3D human mesh reconstruction is to esti- mate 3D coordinates of points, which make up the human body surface. Since the high-quality 3D human model has been consistently required for various immersive applica- tions, many studies have devoted considerable efforts to accurately reconstruct the 3D human mesh. In the early stage of this field, complex optimization techniques were *equal contribution †corresponding author Figure 1. (a) Traditional process of feature extraction for estimat- ing 3D coordinates. (b) Vertex-relevant feature extraction process based on the proposed point-guided sampling method for estimat- ing 3D coordinates. adopted to generate the 3D human model based on the re- lationship between multiple scenes, which are acquired by using stereo or multiple-view camera systems. Recently, owing to the great success of deep learning, the problem of 3D human mesh reconstruction now can be resolved only with a single RGB image, thus the majority has begun to develop compact network architectures and efficient train- ing strategies. Even though such deep learning-based ap- proaches have shown the significant progress in 3D human mesh reconstruction, this task is still challenging due to se- vere occlusions by diverse human poses and depth ambigu- ities by the monocular setting. Deep learning-based approaches can be divided into two main groups: model-based and model-free methods. In the former, most methods aim to estimate shape and pose parameters of the skinned multi-person linear (SMPL) model [24], which is capable of yielding the whole vertices via these two simple factors, thus most widely employed in this field. Traditional encoder-decoder architectures, which are mostly composed of stacked convolutional layers, are sufficient to conduct the regression for estimating those pa- rameters. Despite their great performance, model-based 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. 12880 methods have the obvious shortcoming, i.e., reconstruction results are limited to the pre-defined types of human body models. On the other hand, model-free methods have at- tempted to directly infer 3D coordinates of mesh vertices from input features without using any specific human body model. Compared to the model-based approach, which ob- tains the well-defined full mesh by adjusting shape and pose parameters, the model-free approach needs to estimate 3D coordinates of whole vertices directly from the network. Most methods in this category are based on the transformer to grasp non-local interactions between mesh vertices. The graph model (e.g., graph convolution) also has been utilized together to allow for body part relations in a local manner. One important advantage of the model-free approach is the flexibility to adapt to other applications, e.g., hand pose es- timation, without significant changes of the data format and the training strategy. However, inferring the 3D coordinate from a single monocular image is still challenging due to lack of learning the correspondence between encoded fea- tures and spatial positions. In this paper, we propose a simple yet powerful method for 3D human mesh reconstruction. To this end, we con- duct feature sampling at vertex-relevant points of the input image as shown in Fig. 1, which are estimated through the heatmap decoder trained by projection results of 3D mesh vertices (i.e., ground truth). These sampled features are sub- sequently fed into the transformer encoder as the form of the vertex token (see Fig. 2). In a similar way of [6], we apply attention masking to the transformer encoder, how- ever, the difference is that the local connection is defined with the range of multiple levels through the sequence of transformer encoders. This progressive attention masking helps the model understand local relations between vertices precisely even in occlusions. The main contribution of the proposed method can be summarized as follows: • We propose to utilize the correspondence between en- coded features and vertex positions, which are pro- jected into the 2D space, via our point-guided fea- ture sampling scheme. By explicitly indicating such vertex-relevant features to the transformer encoder, co- ordinates of the 3D human mesh are accurately esti- mated. • Our progressive attention masking scheme helps the model efficiently deal with local vertex-to-vertex rela- tions even under complicated poses and occlusions.
Liu_Few-Shot_Non-Line-of-Sight_Imaging_With_Signal-Surface_Collaborative_Regularization_CVPR_2023	Abstract The non-line-of-sight imaging technique aims to recon- struct targets from multiply reﬂected light. For most exist- ing methods, dense points on the relay surface are raster scanned to obtain high-quality reconstructions, which re- quires a long acquisition time. In this work, we propose a signal-surface collaborative regularization (SSCR) frame- work that provides noise-robust reconstructions with a min- imal number of measurements. Using Bayesian inference, we design joint regularizations of the estimated signal, the 3D voxel-based representation of the objects, and the 2D surface-based description of the targets. To our best knowl- edge, this is the ﬁrst work that combines regularizations in mixed dimensions for hidden targets. Experiments on syn- thetic and experimental datasets illustrated the efﬁciency of the proposed method under both confocal and non-confocal settings. We report the reconstruction of the hidden targets with complex geometric structures with only 55confocal measurements from public datasets, indicating an acceler- ation of the conventional measurement process by a factor of 10,000. Besides, the proposed method enjoys low time and memory complexity with sparse measurements. Our approach has great potential in real-time non-line-of-sight imaging applications such as rescue operations and au- tonomous driving.	1. Introduction The non-line-of-sight (NLOS) imaging technique en- ables reconstructions of targets out of the direct line of sight, which is attractive in various applications such as au- tonomous driving, remote sensing, rescue operations and medical imaging [1,5,6,10,15,16,19,21,26,33–35,38–40]. A typical scenario of NLOS imaging is shown in Figure 1. Several points on the visible surface are illuminated by a laser and the back-scattered light from the target is de- tected to reconstruct the target. The NLOS detection sys- tem is confocal if each illumination point is the same with Figure 1. A typical non-line-of-sight imaging scenario. a) The time resolved signals are measured at only 33focal points. b) The three views of the reconstructed target obtained with the pro- posed SSCR method. the detection point, and non-confocal otherwise. The time- correlated single-photon counting (TCSPC) technique is ap- plied in the detection process due to the extremely low pho- ton intensity after multiple diffuse reﬂections. In practice, a single-photon avalanche diode (SPAD) in the Geiger-mode can be used to record the photon events with time-of-ﬂight (TOF) information [3]. The ﬁrst experimental demonstra- tion of NLOS imaging dates back to 2012, where the targets are reconstructed with the back-projection (BP) method [37]. Extensions of this approach include its fast implemen- tation [2], the ﬁltering technique for reconstruction quality enhancement [17], and weighting factors for noise reduc- tion [11]. A number of efﬁcient methods have been designed 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. 13303 for fast reconstructions. The light cone transform (LCT) method [30] formulates the physical model as a convolu- tion operator, so that the reconstructions can be obtained us- ing the Wiener deconvolution method with the fast Fourier transform. The directional light cone transform (D-LCT) [42] generalizes the LCT and reconstructs the albedo and surface normal simultaneously. The method of frequency wavenumber migration (F-K) [20] formulates the propaga- tion of light using the wave equation, and also provides a fast inversion algorithm with the frequency-domain interpo- lation technique. Whereas the LCT, D-LCT and F-K meth- ods only work directly in confocal measurement scenarios, the phasor ﬁeld (PF) method [23,24,32] converts the NLOS imaging scenarios to LOS cases and works for the gen- eral non-confocal setting with low computation complex- ity. For high-quality and noise-robust reconstructions, the signal-object collaborative regularization (SOCR) method can be applied, but brings additional computational cost. In recent years, deep learning-based methods are also intro- duced to the ﬁeld of NLOS imaging [7, 8, 27, 43]. Besides, advances in hardware enhance the distance of NLOS detec- tion to kilometers [39], or make it possible to reconstruct targets on the scale of millimeters [38]. Despite these breakthroughs, the trade off between the acquisition time and the imaging quality is inevitable. In the raster scanning mode, the acquisition time is propor- tional to the number of measurement points with ﬁxed scan- ning speed. Due to the intrinsic ill-posedness of the NLOS reconstruction problem [22] and heavy measurement noise [11], dense measurements are necessary for high quality re- constructions [20, 23, 30]. The measurement process may take from seconds to hours, which poses a great challenge for applications such as autonomous driving, where real- time reconstruction of the video stream is needed. The ac- quisition process can be accelerated by reducing the number of pulses used for each illumination point. In the work [18], the pulse number that record the ﬁrst returning photon is used to reconstruct the target. Another way to reduce the ac- quisition time is to design array detectors for non-confocal measurements. For example, the implementation of the phasor ﬁeld method with SPAD arrays realizes low-latency real-time video imaging of the hidden scenes [28]. A third way to accelerate the NLOS detection process is to reduce the number of measurement points. It is shown that 1616 confocal measurements are enough to reconstruct the hid- den target by incorporating the compressed sensing tech- nique [41]. In this paper, we study the randomness in the photon de- tection process of NLOS scenarios and propose an imag- ing method that deals with a very limited number of spatial measurements, which we term the few-shot NLOS detec- tion scenarios. We design joint regularizations of the esti- mated signal, the 3D voxel-based representation of the ob- Figure 2. The least-squares solution of the statue with 33con- focal measurements [20]. The target cannot be identiﬁed even though its simulated signal matches the measurements well (see Fig. 3). Strong regularizations are needed to reconstruct the target. See also Fig. 6 for a comparison. Figure 3. Comparisons of the measured data and the simulated data of the least-squares solution for the instance of the statue [20]. The measured signals are shown in black. The simulated data of the least-squares solution are shown in red. The shapes of the signals are very close to each other. jects, and the 2D surface-based description of the targets, which leads to faithful reconstruction results. The main contributions of this work are as follows. We propose a signal-surface collaborative regulariza- tion (SSCR) framework for few-shot non-line-of-sight reconstructions, which works under both confocal and non-confocal settings. We report the reconstruction of the hidden targets with complex geometric structures with only 55confo- cal measurements from public datasets, indicating an acceleration of the conventional measurement process by a factor of 10,000. 13304
Ko_MELTR_Meta_Loss_Transformer_for_Learning_To_Fine-Tune_Video_Foundation_CVPR_2023	Abstract Foundation models have shown outstanding perfor- mance and generalization capabilities across domains. Since most studies on foundation models mainly focus on the pretraining phase, a naive strategy to minimize a sin- gle task-specific loss is adopted for fine-tuning. However, such fine-tuning methods do not fully leverage other losses that are potentially beneficial for the target task. There- fore, we propose MEtaLossTRansformer ( MELTR ), a plug-in module that automatically and non-linearly com- bines various loss functions to aid learning the target task via auxiliary learning. We formulate the auxiliary learn- ing as a bi-level optimization problem and present an ef- ficient optimization algorithm based on Approximate Im- plicit Differentiation (AID). For evaluation, we apply our framework to various video foundation models (UniVL, Violet and All-in-one), and show significant performance gain on all four downstream tasks: text-to-video retrieval, video question answering, video captioning, and multi- modal sentiment analysis. Our qualitative analyses demon- strate that MELTR adequately ‘transforms’ individual loss functions and ‘melts’ them into an effective unified loss. Code is available at https://github.com/mlvlab/ MELTR .	1. Introduction Large-scale models trained on a huge amount of data have gained attention due to their adaptability to a wide range of downstream tasks. As introduced in [1], deep learning models with the generalizability are referred to as foundation models. In recent years, several foundation models for various domains have been proposed ( e.g., [2,3] *Equal contribution. †Corresponding author.for natural language processing, [4, 5] for images and lan- guage, and [6–8] for videos) and they mainly focus on pre- train the model often with various multiple pretext tasks. On the other hand, strategies for fine-tuning on downstream tasks are less explored. For instance, a recently proposed video foundation model UniVL [7] is pretrained with a lin- earcombination of several pretext tasks such as text-video alignment, masked language/frame modeling, and caption generation. However, like other domains, fine-tuning is simply performed by minimizing a single target loss. Other potentially beneficial pretext tasks have remained largely unexplored for fine-tuning. Auxiliary learning is a natural way to utilize multiple pretext task losses for learning. Contrary to multi-task learning that aims for generalization across tasks, auxiliary learning focuses only on the primary task by taking ad- vantage of several auxiliary tasks. Most auxiliary learning frameworks [9,10] manually selected auxiliary tasks, which require domain knowledge and may not always be benefi- cial for the primary task. To automate task selection, meta learning was integrated into auxiliary learning [11–13]. Here, the model learns to adaptively leverage multiple aux- iliary tasks to assist learning of the primary task. Likewise, the pretext task losses can be unified into a single auxiliary loss to be optimized in a way that helps the target down- stream task. To this end, we propose Meta Loss Transformer (MELTR), a plug-in module that automatically andnon- linearly transforms various auxiliary losses into a unified loss. MELTR built on Transformers [14] takes the tar- get task loss as well as pretext task losses as input and learns their relationship via self-attention. In other words, MELTR learns to fine-tune a foundation model by combin- ing the primary task with multiple auxiliary tasks, and this can be viewed as a meta-learning (or ‘learning-to-learn’) problem. Similar to meta-learning-based auxiliary learning frameworks [13,15], this can be formulated as a bi-level op- 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. 20105 timization problem, which generally involves a heavy com- putational cost due to the second-order derivative and its in- verse, e.g., the inverse Hessian matrix. To circumvent this, we present an efficient training scheme that approximates the inverse Hessian matrix. We further provide empirical analyses on the time-performance trade-off of various opti- mization algorithms. To verify the generality of our proposed method, we ap- ply it to three video foundation models: UniVL [7], Vi- olet [16], and All-in-one [17]. These foundation models are originally pretrained with a linear combination of sev- eral pretext tasks such as text-video alignment, masked lan- guage/frame modeling, and caption generation. We exper- iment by fine-tuning on the text-to-video retrieval, video question answering, video captioning, and multi-modal sentiment analysis task with five datasets: YouCook2, MSRVTT, TGIF, MSVD, and CMU-MOSI. For each task and dataset, our MELTR improves both previous foun- dation models and task-specific models by large margins. Furthermore, our extensive qualitative analyses and abla- tion studies demonstrate that MELTR effectively learns to non-linearly combine pretext task losses, and adaptively re- weights them for the target downstream task. To sum up, our contributions are threefold: • We propose MEtaLossTRansformer ( MELTR ), a novel fine-tuning framework for video foundation models. We also present an efficient optimization al- gorithm to alleviate the heavy computational cost of bi-level optimization. • We apply our framework to three video foundation models in four downstream tasks on five benchmark video datasets, where MELTR significantly outper- forms the baselines fine-tuned with single-task and multi-task learning schemes. • We provide in-depth qualitative analyses on how MELTR non-linearly transforms individual loss func- tions and combines them into an effective unified loss for the target downstream task.
Li_ACSeg_Adaptive_Conceptualization_for_Unsupervised_Semantic_Segmentation_CVPR_2023	Abstract Recently, self-supervised large-scale visual pre-training models have shown great promise in representing pixel- level semantic relationships, significantly promoting the de- velopment of unsupervised dense prediction tasks, e.g., un- supervised semantic segmentation (USS). The extracted re- lationship among pixel-level representations typically con- tains rich class-aware information that semantically iden- tical pixel embeddings in the representation space gather together to form sophisticated concepts. However, lever- aging the learned models to ascertain semantically con- sistent pixel groups or regions in the image is non-trivial since over/ under-clustering overwhelms the conceptualiza- tion procedure under various semantic distributions of dif- ferent images. In this work, we investigate the pixel-level semantic aggregation in self-supervised ViT pre-trained models as image Segmentation and propose the Adaptive Conceptualization approach for USS, termed ACSeg . Con- cretely, we explicitly encode concepts into learnable proto- types and design the Adaptive Concept Generator (ACG), which adaptively maps these prototypes to informative con- cepts for each image. Meanwhile, considering the scene complexity of different images, we propose the modularity loss to optimize ACG independent of the concept number based on estimating the intensity of pixel pairs belonging to the same concept. Finally, we turn the USS task into clas- sifying the discovered concepts in an unsupervised manner. Extensive experiments with state-of-the-art results demon- strate the effectiveness of the proposed ACSeg.	1. Introduction Semantic segmentation is one of the primary tasks in computer vision, which has been widely used in many do- mains, such as autonomous driving [7, 14] and medical *Corresponding author. Project page: https://lkhl.github.io/ACSeg. (a) Under-clustering (b) Over-clustering (c) Our Adaptive ConceptualizationPixel-level representation Prototype Inactive prototype Pixels in a concept Concept Adaptive Concept Generator ConceptFigure 1. Comparison between existing methods and our adaptive conceptualization on finding underlying “concepts” in the pixel- level representations produced by a pre-trained model. While under-clustering just focuses on a single object and over-clustering splits objects, our adaptive conceptualization processes different images adaptively through updating the initialized prototypes with the representations for each image. imaging [12, 24, 41]. With the development of deep learn- ing and the increasing amount of data [7, 11, 28, 57], uplift- ing performance has been achieved on this task by optimiz- ing deep neural networks with pixel-level annotations [29]. However, large-scale pixel-level annotations are expensive and laborious to obtain. Different kinds of weak supervision have been explored to achieve label efficiency [37], e.g., image-level [1,49], scribble-level [27], and box-level super- vision [35]. More than this, some methods also achieve se- mantic segmentation without relying on any labels [19, 20], namely unsupervised semantic segmentation (USS). Early approaches for USS are based on pixel-level self- supervised representation learning by introducing cross- 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. 7162 view consistency [6,20], edge detection [19,56], or saliency prior [43]. Recently, the self-supervised ViT [4] provides a new paradigm for USS due to its property of containing se- mantic information in pixel-level representations. We make it more intuitive through Figure 1, which shows that in the representation space of an image, the pixel-level representa- tions produced by the self-supervised ViT contain underly- ing clusters. When projecting these clusters into the image, they become semantically consistent groups of pixels or re- gions representing “concepts”. In this work, we aim to achieve USS by accurately ex- tracting and classifying these “concepts” in the pixel rep- resentation space of each image. Unlike the previous at- tempts which only consider foreground-background parti- tion [40, 44, 48] or divide each image into a fixed number of clusters [18, 32], we argue that it is crucial to consider different images distinguishably due to the complexity of various scenarios (Figure 1). We thus propose the Adaptive Conceptualization for unsupervised semantic Segmentation (ACSeg), a framework that finds these underlying concepts adaptively for each image and achieves USS by classifying the discovered concepts in an unsupervised manner. To achieve conceptualization, we explicitly encode con- cepts to learnable prototypes and adaptively update them for different images by a network, as shown in Figure 2. This network, named as Adaptive Concept Generator (ACG), is implemented by iteratively applying scaled dot-product at- tention [45] on the prototypes and pixel-level representa- tions in the image to be processed. Through such a struc- ture, the ACG learns to project the initial prototypes to the concept in the representation space depending on the in- put pixel-level representations. Then the concepts are ex- plicitly presented in the image as different regions by as- signing each pixel to the nearest concept in the representa- tion space. The ACG is end-to-end optimized without any annotations by the proposed modularity loss. Specifically, we construct an affinity graph on the pixel-level representa- tions and use the connection relationship of two pixels in the affinity graph to adjust the strength of assigning two pixels to the same concept, motivated by the modularity [34]. As the main part of ACSeg, the ACG achieves precise conceptualization for different images due to its adaptive- ness, which is reflected in two aspect: Firstly, it can adap- tively operate on pixel-level representations of different im- ages thanks to the dynamic update structure. Secondly, the training objective does not enforce the number of concepts, resulting in adaptive number of concepts for different im- ages. With these properties, we get accurate partition for images with different scene complexity via the concepts produced by the ACG, as shown in Figure 1(c). Therefore, in ACSeg, the semantic segmentation of an image can fi- nally be achieved by matting the corresponding regions in the image and classifying them with the help of powerful Update Updated prototypes Initialized prototypesPrototypes Pixel-level representation Pixels in a concept Update by pixel-level representations Update by each otherFigure 2. Intuitive explanation for the basic idea of the ACG. The concepts are explicitly encoded to learnable prototypes and dynamically updated according to the input pixel-level representa- tions. After update, the pixels are assigned to the nearest concept in the representation space. image-level pre-trained models. For evaluation, we apply ACSeg on commonly used semantic segmentation datasets, including PASCAL VOC 2012 [11] and COCO-Stuff [20, 28]. The experimental results show that the proposed ACSeg surpasses previous methods on different settings of unsupervised semantic seg- mentation tasks and achieves state-of-the-art performance on the PASCAL VOC 2012 unsupervised semantic segmen- tation benchmark without post-processing and re-training. Moreover, the visualization of the pixel-level representa- tions and the concepts shows that the ACG is applicable for decomposing images with various scene complexity. Since the ACG is fast to converge without learning new represen- tations and the concept classifier is employed in a zero-shot manner, we draw the proposed ACSeg as a generalizable method which is easy to modify and adapt to a wide range of unsupervised image understanding.
Lin_Adaptive_Human_Matting_for_Dynamic_Videos_CVPR_2023	Abstract The most recent efforts in video matting have focused on eliminating trimap dependency since trimap annotations are expensive and trimap-based methods are less adapt- able for real-time applications. Despite the latest tripmap- free methods showing promising results, their performance often degrades when dealing with highly diverse and un- structured videos. We address this limitation by introduc- ingAdaptive Matting for Dynamic Videos, termed AdaM , which is a framework designed for simultaneously differen- tiating foregrounds from backgrounds and capturing alpha matte details of human subjects in the foreground. Two in- terconnected network designs are employed to achieve this goal: (1) an encoder-decoder network that produces alpha mattes and intermediate masks which are used to guide the transformer in adaptively decoding foregrounds and back- grounds, and (2) a transformer network in which long- and short-term attention combine to retain spatial and tempo- ral contexts, facilitating the decoding of foreground de- tails. We benchmark and study our methods on recently introduced datasets, showing that our model notably im- proves matting realism and temporal coherence in complex real-world videos and achieves new best-in-class general- izability. Further details and examples are available at https://github.com/microsoft/AdaM .	1. Introduction Video human matting aims to estimate a precise alpha matte to extract the human foreground from each frame ofan input video. In comparison with image matting [6,10,14, 21, 30, 39, 44, 47], video matting [2, 5, 11, 18, 19, 33, 36, 38] presents additional challenges, such as preserving spatial and temporal coherence. Many different solutions have been put forward for the video matting problem. A straightforward approach is to build on top of image matting models [44], which is to im- plement an image matting approach frame by frame. It may, however, result in inconsistencies in alpha matte predictions across frames, which will inevitably lead to flickering ar- tifacts [42]. On the other hand, top performers leverage dense trimaps to predict alpha mattes, which is expensive and difficult to generalize across large video datasets. To alleviate the substantial trimap limitation, OTVM [35] pro- posed a one-trimap solution recently. BGM [22, 33] pro- poses a trimap-free solution, which needs to take an addi- tional background picture without the subject at the time of capture. While the setup is less time-consuming than cre- ating trimaps, it may not work well if used in a dynamic background environment. The manual prior required by these methods limits their use in some real-time applica- tions, such as video conferencing. Lately, more general so- lutions, e.g., MODNet [16] and RVM [23], have been pro- posed which involve manual-free matting without auxiliary inputs. However, in challenging real-world videos, back- grounds are inherently non-differentiable at some points, causing these solutions to produce blurry alpha mattes. It is quite challenging to bring together the benefits of both worlds, i.e., a manual-free model that produces accu- rate alpha mattes in realistic videos. In our observation, 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. 10229 Figure 2. Qualitative sample results of MODNet [16], RVM [23] and the proposed AdaM on real video scenes. significant challenges can mostly be explained by the inher- ent unconstrained and diverse nature of real-world videos. As a camera moves in unstructured and complex scenes, foreground colors can resemble passing objects in the back- ground, which makes it hard to separate foreground subjects from cluttered backgrounds. This might result in blurred foreground boundaries or revealing backgrounds behind the subjects, as shown in Fig. 2. MODNet [16] and RVM [23] are both nicely designed models with auxiliary-free ar- chitectures to implicitly model backgrounds. In complex scenes, however, models without guidance may experience foreground-background confusion, thereby degrading the alpha matte accuracy. In this paper, we aim to expand the applicability of the matting architecture such that it can serve as a reli- able framework for human matting in real-world videos. Our method does not require manual efforts (e.g., manu- ally annotated trimaps or pre-captured backgrounds). The main idea is straightforward: understanding and structur- ing background appearance can make the underlying mat- ting network easier to render high-fidelity foreground alpha mattes in dynamic video scenes. Toward this goal, an in- terconnected two-network framework is employed: (1) an encoder-decoder network with skip connections produces alpha mattes and intermediate masks that guide the trans- former network in adaptively enhancing foregrounds and backgrounds, and (2) a transformer network with long- and short-term attention that retain both spatial and temporal contexts, enabling foreground details to be decoded. Based on a minimal-auxiliary strategy, the transformer network obtains an initial mask from an off-the-shelf segmenter for coarse foreground/background (Fg/Bg) guidance, but the decoder network predicts subsequent masks automatically in a data-driven manner. The proposed method seeks to pro- duce accurate alpha mattes in challenging real-world envi- ronments while eliminating the sensitivities associated with handling an ill-initialized mask. Compared to the recently published successes in video matting study, our main con- tributions are as follows: • We propose a framework for human matting with uni-fied handling of complex unconstrained videos without requiring manual efforts. The proposed method pro- vides a data-driven estimation of the foreground masks to guide the network to distinguish foregrounds and backgrounds adaptively. • Our network architecture and training scheme have been carefully designed to take advantage of both long- and short-range spatial and motion cues. It reaches top-tier performance on the VM [23] and CRGNN [42] benchmarks.
Liu_SCOTCH_and_SODA_A_Transformer_Video_Shadow_Detection_Framework_CVPR_2023	Abstract Shadows in videos are difficult to detect because of the large shadow deformation between frames. In this work, we argue that accounting for shadow deformation is essen- tial when designing a video shadow detection method. To this end, we introduce the shadow deformation attention trajectory ( SODA ), a new type of video self-attention mod- ule, specially designed to handle the large shadow defor- mations in videos. Moreover, we present a new shadow contrastive learning mechanism ( SCOTCH ) which aims at guiding the network to learn a unified shadow represen- tation from massive positive shadow pairs across differ- ent videos. We demonstrate empirically the effectiveness of our two contributions in an ablation study. Furthermore, we show that SCOTCH andSODA significantly outperforms existing techniques for video shadow detection. Code is available at the project page: https://lihaoliu- cambridge.github.io/scotch_and_soda/	1. Introduction Shadow is an inherent part of videos, and they have an adverse effect on a wide variety of video vision tasks. Therefore, the development of robust video shadow detec- tion techniques, to alleviate those negative effects, is of great interest for the community. Video shadow detection is usually formulated as a segmentation problem for videos, however and due to the nature of the problem, shadow de- tection greatly differs from other segmentation tasks such as object segmentation. For inferring the presence of shad- ows in an image, one has to account for the global content information such as light source orientation, and the pres- ence of objects casting shadows. Importantly, in a given video, shadows considerably change appearance (deforma- tion) from frame to frame due to light variation and object motion. Finally, shadows can span over different back- grounds over different frames, making approaches relying EncoderBlock1t×!"×#"×c1t×!$%×#$%×c2t×!&'×#&'×c3 Block3Block2Stage1 DecoderMLPLayert×!&'×#&'×c3 t×!$%×#$%×c2t×!"×#"×c1Stage3 UpsampleLayerInputVideo SegmentationMasks t = 3t = 2t = 1t = 4Stage 2Deformation Trajectory AttentionShadowContrast Shadow FeatureNon-shadow FeatureGet SimilarGet DifferentA is the input of BDeformation Trajectoryz1z2z3 z!"z#"z$"Figure 1. Overview of our SCOTCH andSODA framework. A MiT encoder extracts multi-scale features for each frame of the video (stage 1). Then, our deformation attention trajectory is applied to features individually to incorporate temporal information (stage 2). Finally, an MLP layer combines the multi-scale information to generate the segmentation masks (stage 3). The model is trained to contrast shadow and non-shadow features, by minimising our shadow contrastive loss with massive positive shadow pairs. on texture information unreliable. Particularly, video shadow detection methods can be broadly divided into two main categories. The first category refers to image shadow detection (ISD) [9,15,35,36,43,46]. This family of techniques computes the shadow detec- tion frame by frame. Although computationally saving, these methods are incapable of handling temporal informa- tion. The second category refers to video shadow detection (VSD) [6, 9, 14, 16, 25]. These methods offer higher per- formance as the analysis involves spatial-temporal informa- tion. Hence, our main focus is video shadow detection. 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. 10449 State-of-the-art video shadow detection methods rely on deep neural networks, which are trained on large annotated datasets. Specifically, those methods are composed of three parts: (i) a feature extraction network that extracts spatial features for each frame of the video: (ii) a temporal ag- gregation mechanism [6, 14] enriching spatial features with information from different frames; and (iii) a decoder, that maps video features to segmentation masks. Additionally, some works enforce consistency between frames prediction by using additional training criterion [9,25]. We retain from these studies that the design of the temporal aggregation mechanism and the temporal consistency loss is crucial to the performance of a video shadow detection network, and we will investigate both of those aspects in this work. The current temporal aggregation mechanisms available in the literature were typically designed for video tasks such as video action recognition, or video object segmenta- tion. Currently, the most widely used temporal aggregation mechanism is based on a variant of the self-attention mech- anism [1, 29, 32, 40, 41]. Recently, trajectory attention [29] has been shown to provide state-of-the-art results on video processing. Intuitively, trajectory attention aggregates in- formation along the object’s moving trajectory, while ignor- ing the context information, deemed as irrelevant. However, shadows in videos are subject to strong deformations, mak- ing them difficult to track, and thus they might cause the trajectory attention to fail. In this work, we first introduce the ShadOw Deformation Attention trajectory ( SODA ), a spatial-temporal aggregation mechanism designed to better handle the large shadow de- formations that occur in videos. SODA operates in two steps. First, for each spatial location, an associated token is com- puted between the given spatial location and the video, which contains information in every time-step for the given spatial location. Second, by aggregating every associated spatial token, a new token is yielded with enriched spatial deformation information. Aggregating spatial-location-to- video information along the spatial dimension helps the net- work to detect shape changes in videos. Besides, we introduce the Shadow COnTrastive meCH- anism ( SCOTCH ), a supervised contrastive loss with massive positive shadow pairs aiming to drive our network to learn more discriminative features for the shadow regions in dif- ferent videos. Specifically, in training, we add a contrastive loss at the coarsest layer of the encoder, driving the fea- tures from shadow regions close together, and far from the features from the non-shadow region. Intuitively, this con- trastive mechanism drives the encoder to learn high-level representations of shadow, invariant to all the various fac- tors of shadow variations, such as shape and illumination. In summary, our contributions are as follows: • We introduce a new video shadow detection frame- work, in which we highlight:–SODA , a new type of trajectory attention that har- monise the features of the different video frames at each resolution. –SCOTCH , a contrastive loss that highlights a mas- sive positive shadow pairs strategy in order to make our encoder learn more robust high-level representations of shadows. • We evaluate our proposed framework on the video shadow benchmark dataset ViSha [6], and compare with the state-of-the-art methods. Numerical and vi- sual experimental results demonstrate that our ap- proach outperforms, by a large margin, existing ones on video shadow detection. Furthermore, we provide an ablation study to further support the effectiveness of the technical contributions.
Karunratanakul_HARP_Personalized_Hand_Reconstruction_From_a_Monocular_RGB_Video_CVPR_2023	Abstract We present HARP (HAnd Reconstruction and Personal- ization), a personalized hand avatar creation approach that takes a short monocular RGB video of a human hand as input and reconstructs a faithful hand avatar exhibiting a high-ﬁdelity appearance and geometry. In contrast to the major trend of neural implicit representations, HARP mod- els a hand with a mesh-based parametric hand model, a vertex displacement map, a normal map, and an albedo without any neural components. The explicit nature of our representation enables a truly scalable, robust, and efﬁcient approach to hand avatar creation as validated by our ex- periments. HARP is optimized via gradient descent from a short sequence captured by a hand-held mobile phone and can be directly used in AR/VR applications with real- time rendering capability. To enable this, we carefully de- sign and implement a shadow-aware differentiable render- ing scheme that is robust to high degree articulations and self-shadowing regularly present in hand motions, as wellas challenging lighting conditions. It also generalizes to un- seen poses and novel viewpoints, producing photo-realistic renderings of hand animations. Furthermore, the learned HARP representation can be used for improving 3D hand pose estimation quality in challenging viewpoints. The key advantages of HARP are validated by the in-depth analyses on appearance reconstruction, novel view and novel pose synthesis, and 3D hand pose reﬁnement. It is an AR/VR- ready personalized hand representation that shows superior ﬁdelity and scalability.	1. Introduction Advancements in AR/VR devices are introducing a new reality in which the physical and digital worlds merge. The human hand is a crucial element for an intimate and in- teractive experience in these environments, serving as the primary interface between humans and the digital world. Therefore, it is essential to capture, reconstruct, and animate life-like digital hands for AR and VR applications. Without 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. 12802 this capability, the authenticity and practicality of AR/VR consumer products will always be limited. Despite its importance, the research into hand avatar cre- ation has so far been limited. Most works [8, 35, 59] focus on creating an appearance space on top of a parametric hand model such as MANO [62]. Such an appearance space pro- vides a compact way to represent hand texture but is rather limited in expressivity to handle non-standard textures. The recent LISA [12] model has emerged as an alternative, us- ing an implicit function to represent hand geometry and tex- ture color ﬁelds. Training a new identity in LISA, how- ever, requires a multi-view capturing setup as well as a large amount of data and computing power. In the nearby ﬁelds of face and body avatar creation, many works that leverage an implicit function [19,40,41,74] or NeRF-based [42] volume rendering [39,53,77] have also been recently explored. The NeRF-based method such as HumanNeRF [77] produces a convincing novel view synthesis but still shows blurry ar- tifacts around highly articulated parts and cannot be easily exported to other applications. We argue that democratizing hand avatar creation for AR/VR users requires a method that is (1) accurate : so that personalized hand appearance and geometry can be faith- fully reconstructed; (2) scalable : allowing hand avatars to be obtained using a commodity camera; (3) robust : ca- pable of handling out-of-distribution appearance and self- shadows between ﬁngers and palm; and (4) efﬁcient : with real-time rendering capability. To this end, we propose HARP, a personalized hand re- construction method that can create a faithful hand avatar from a short RGB video captured by a hand-held mobile phone. HARP leverages a parametric hand model, an ex- plicit appearance, and a differentiable rasterizer and shader to reconstruct a hand avatar and environment lighting in an analysis-by-synthesis manner, without any neural net- work component . Our observation is that human hands are highly articulated. The appearance changes of observed hands in a captured sequence can be dramatic and largely attributed to articulations and light interaction. Learning neural representations, such as implicit texture ﬁelds [12] or volume-based representations like NeRF [70], is vulnerable to the over-ﬁtting to a short monocular training sequence and can hardly generalize well to sophisticated and dexter- ous hand movements. By properly disentangling geometry, appearance, and self-shadow with explicit representations, HARP can signiﬁcantly improve the reconstruction quality and generate life-like renderings on novel views and novel animations performing highly articulated motions. Further- more, the nature of the explicit representation allows the results from HARP to be conveniently exported to standard graphics applications. In summary, the key advantages of HARP are: (1) HARP is a simple personalized hand avatar creation methodthat reconstructs high-ﬁdelity appearance and geometry us- ing only a short monocular video. HARP demonstrates that an explicit representation with a differentiable raster- izer and shader is enough to obtain life-like hand avatars. (2) The hand avatar from HARP is controllable and com- patible with standard rasterization graphics pipelines allow- ing for photo-realistic rendering in AR/VR applications. (3) Moreover, HARP can be used to improve 3D hand pose estimation in challenging viewpoints. We perform exten- sive experiments on the tasks of appearance reconstruction, novel-view-and-pose synthesis, and 3D hand poses reﬁne- ment. Compared to existing approaches, HARP is more ac- curate, robust, and generalizable with superior scalability.
Li_KERM_Knowledge_Enhanced_Reasoning_for_Vision-and-Language_Navigation_CVPR_2023	Abstract Vision-and-language navigation (VLN) is the task to en- able an embodied agent to navigate to a remote location following the natural language instruction in real scenes. Most of the previous approaches utilize the entire fea- tures or object-centric features to represent navigable can- didates. However, these representations are not efficient enough for an agent to perform actions to arrive the tar- get location. As knowledge provides crucial information which is complementary to visible content, in this paper, we propose a Knowledge Enhanced Reasoning Model (KERM) to leverage knowledge to improve agent navigation ability. Specifically, we first retrieve facts (i.e., knowledge described by language descriptions) for the navigation views based on local regions from the constructed knowledge base. The re- trieved facts range from properties of a single object (e.g., color, shape) to relationships between objects (e.g., action, spatial position), providing crucial information for VLN. We further present the KERM which contains the purification, fact-aware interaction, and instruction-guided aggregation modules to integrate visual, history, instruction, and fact features. The proposed KERM can automatically select and gather crucial and relevant cues, obtaining more accurate action prediction. Experimental results on the REVERIE, R2R, and SOON datasets demonstrate the effectiveness of the proposed method. The source code is available at https://github.com/XiangyangLi20/KERM .	1. Introduction Vision-and-language navigation (VLN) [3,12,23,24,36, 38] is one of the most attractive embodied AI tasks, where agents should be able to understand natural language in- structions, perceive visual content in dynamic 3D environ- ments, and perform actions to navigate to the target loca- tion. Most previous methods [9,15,22,31,34] depend on se- Panoramic ViewInstructionGoupthestairs through theoutdoor living area and … thekitchen …between thekitchen island and pantry . Retrieved Facts <sofa next to wall> <lamp to right of sofa> <a living room area> …<light hanging over kitchen island> <laminate countertop> …<ceiling recessed light> <wood countertop> <contemporary table> …Navigable Candidate 1 <lamp to right of sofa> <white wall partition> <sofa next to wall>Knowledge database<workspace kitchen island> <contemporary table> <switch above sofa> … ✔ ✔Navigable Candidate 2Navigable Candidate 3Figure 1. Illustration of knowledge related navigable candidates, which provides crucial information such as attributes and relation- ships between objects for VLN. Best viewed in color. quential models ( e.g., LSTMs and Transformers) to contin- uously receive visual observations and align them with the instructions to predict actions at each step. More recently, transformer-based architectures [5, 7, 25] which make use of language instructions, current observations, and histori- cal information have been widely used. Most of the previous approaches utilize the entire fea- tures [5,12,13,25] or object-centric features [1,7,10,20] to represent navigable candidates. For example, Qi et al. [22] and Gao et al. [10] encode discrete images within each panorama with detected objects. Moudgil et al. [20] utilize both object-level and scene-level features to represent visual observations. However, these representations are not effi- cient enough for an agent to navigate to the target location. For example, as shown in Figure 1, there are three candi- 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. 2583 dates. According to the instruction and the current location, candidate2 is the correct navigation. Based on the entire features of a candidate view, it is hard to select the correct one, as candidate2 and candidate3 belong to the same cat- egory ( i.e.,“dining room”). Meanwhile, it is also hard to differentiate them from individual objects, as “lamp” and “light” are the common components for them. As humans make inferences under their knowledge [11], it is important to incorporate knowledge related to navigable candidates for VLN tasks. First, knowledge provides cru- cial information which is complementary to visible content. In addition to visual information, high-level abstraction of the objects and relationships contained by knowledge pro- vides essential information. Such information is indispens- able to align the visual objects in the view image with the concepts mentioned in the instruction. As shown in Fig- ure 1, with the knowledge related to candidate2 ( i.e.,<light hanging over kitchen island >), the agent is able to navigate to the target location. Second, the knowledge improves the generalization ability of the agent. As the alignment be- tween the instruction and the navigable candidate is learned in limited-seen environments, leveraging knowledge ben- efits the alignment in the unseen environment, as there is no specific regularity for target object arrangement. Third, knowledge increases the capability of VLN models. As rich conceptual information is injected into VLN models, the correlations among numerous concepts are learned. The learned correlations are able to benefit visual and language alignment, especially for tasks with high-level instructions. In this work, we incorporate knowledge into the VLN task. To obtain knowledge for view images, facts ( i.e., knowledge described by language descriptions) are re- trieved from the knowledge base constructed on the Vi- sual Genome dataset [16]. The retrieved facts by CLIP [26] provide rich and complementary information for vi- sual view images. And then, a knowledge enhanced rea- soning model (KERM) which leverages knowledge for suf- ficient interaction and better alignment between vision and language information is proposed. Especially, the proposed KERM consists of a purification module, a fact-aware inter- action module, and an instruction-guided aggregation mod- ule. The purification model aims to extract key informa- tion in the fact representations, the visual region representa- tions, and the historical representations respectively guided by the instruction. The fact-aware interaction module al- lows visual and historical representations to obtain the in- teraction of the facts with cross-attention encoders. And the instruction-guided aggregation module extracts the most relevant components of the visual and historical representa- tions according to the instruction for fusion. We conduct the experiments on three VLN datasets, i.e., the REVERIE [24], SOON [36], and R2R [3]. Our ap- proach outperforms state-of-the-art methods on all splits ofthese datasets under most metrics. The further experimental analysis demonstrates the effectiveness of our method. In summary, we make the following contributions: • We incorporate region-centric knowledge to compre- hensively depict navigation views in VLN tasks. For each navigable candidate, the retrieved facts ( i.e., knowledge described by language descriptions) are complementary to visible content. • We propose the knowledge enhanced reasoning model (KERM) to inject fact features into the visual represen- tations of navigation views for better action prediction. • We conduct extensive experiments to validate the ef- fectiveness of our method and show that it outperforms existing methods with a better generalization ability.
Klingner_X3KD_Knowledge_Distillation_Across_Modalities_Tasks_and_Stages_for_Multi-Camera_CVPR_2023	Abstract Recent advances in 3D object detection (3DOD) have obtained remarkably strong results for LiDAR-based mod- els. In contrast, surround-view 3DOD models based on multiple camera images underperform due to the neces- sary view transformation of features from perspective view (PV) to a 3D world representation which is ambiguous due to missing depth information. This paper introduces X3KD, a comprehensive knowledge distillation framework across different modalities, tasks, and stages for multi- camera 3DOD. Specifically, we propose cross-task distil- lation from an instance segmentation teacher (X-IS) in the PV feature extraction stage providing supervision without ambiguous error backpropagation through the view trans- formation. After the transformation, we apply cross-modal feature distillation (X-FD) and adversarial training (X-AT) to improve the 3D world representation of multi-camera features through the information contained in a LiDAR- based 3DOD teacher. Finally, we also employ this teacher for cross-modal output distillation (X-OD), providing dense supervision at the prediction stage. We perform extensive ablations of knowledge distillation at different stages of multi-camera 3DOD. Our final X3KD model outperforms previous state-of-the-art approaches on the nuScenes and Waymo datasets and generalizes to RADAR-based 3DOD. Qualitative results video at https://youtu.be/1do9DPFmr38.	1. Introduction 3D object detection (3DOD) is an essential task in vari- ous real-world computer vision applications, especially au- tonomous driving. Current 3DOD approaches can be cate- gorized by their utilized input modalities, e.g., camera im- ages [28, 40, 46] or LiDAR point clouds [25, 55, 60], which dictates the necessary sensor suite during inference. Re- cently, there has been significant interest in surround-view *These authors contributed equally to this work. †Automated Driving, Qualcomm Technologies, Inc. ‡Qualcomm AI Research, an initiative of Qualcomm Technologies, Inc. §Automated Driving, QT Technologies Ireland Limited Figure 1. While previous approaches considered multi-camera 3DOD in a standalone fashion or with depth supervision, we pro- pose X3KD, a knowledge distillation framework using cross- modal and cross-task information by distilling information from LiDAR-based 3DOD and instance segmentation teachers into dif- ferent stages (marked by red arrows) of the multi-camera 3DOD. multi-camera 3DOD, aiming to leverage multiple low-cost monocular cameras, which are conveniently embedded in current vehicle designs in contrast to expensive LiDAR scanners. Existing solutions to 3DOD are mainly based on extracting a unified representation from multiple cameras [28,30,37,41] such as the bird’s-eye view (BEV) grid. How- ever, predicting 3D bounding boxes from 2D perspective- view (PV) images involves an ambiguous 2D to 3D transfor- mation without depth information, which leads to lower per- formance compared to LiDAR-based 3DOD [1, 28, 30, 55]. While LiDAR scanners may not be available in commer- cially deployed vehicle fleets, they are typically available in training data collection vehicles to facilitate 3D annotation. Therefore, LiDAR data is privileged; it is often available 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. 13343 Model LSS++ DS GFLOPS mAP↑ NDS↑ BEVDepth†✗ ✗ 298 32.4 44.9 ✗ ✓ 298 33.1 44.9 ✓ ✗ 316 34.9 47.0 ✓ ✓ 316 35.9 47.2 X3KD(Ours) ✓ ✓ 316 39.0 50.5 Table 1. Analysis of BEVDepth†(re-implementation of [28]): We compare the architectural improvement of a larger Lift-Splat- Shoot (LSS++) transform to using depth supervision (DS). during training but not during inference. The recently intro- duced BEVDepth [28] approach pioneers using accurate 3D information from LiDAR data at training time to improve multi-camera 3DOD, see Fig. 1 (top part). Specifically, it proposed an improved Lift-Splat-Shoot PV-to-BEV trans- form (LSS++) and depth supervision (DS) by projected Li- DAR points, which we analyze in Table 1. We observe that the LSS++ architecture yields significant improvements, though depth supervision seems to have less effect. This motivates us to find additional types of supervision to trans- fer accurate 3D information from LiDAR point clouds to multi-camera 3DOD. To this end, we propose cross-modal knowledge distillation (KD) to not only use LiDAR data but a high-performing LiDAR-based 3DOD model , as in Fig. 1 (middle part). To provide an overview of the effectiveness of cross-modal KD at various multi-camera 3DOD network stages, we present three distillation techniques: feature dis- tillation (X-FD) and adversarial training (X-AT) to improve the feature representation by the intermediate information contained in the LiDAR 3DOD model as well as output dis- tillation (X-OD) to enhance output-stage supervision. For optimal camera-based 3DOD, extracting useful PV features before the view transformation to BEV is equally essential. However, gradient-based optimization through an ambiguous view transformation can induce non-optimal su- pervision signals. Recent work proposes pre-training the PV feature extractor on instance segmentation to improve the extracted features [49]. Nevertheless, neural networks are subject to catastrophic forgetting [23] such that knowl- edge from pre-training will continuously degrade if not re- tained by supervision. Therefore, we propose cross-task in- stance segmentation distillation (X-IS) from a pre-trained instance segmentation teacher into a multi-camera 3DOD model, see Fig. 1 (bottom part). As shown in Table 1, our X3KD framework significantly improves upon BEVDepth without additional complexity during inference. To summarize, our main contributions are as follows: • We propose X3KD, a KD framework across modali- ties, tasks, and stages for multi-camera 3DOD. • Specifically, we introduce cross-modal KD from a strong LiDAR-based 3DOD teacher to the multi- camera 3DOD student, which is applied at multiple network stages in bird’s eye view, i.e., feature-stage(X-FD and X-AT) and output-stage (X-OD). • Further, we present cross-task instance segmentation distillation (X-IS) at the PV feature extraction stage. • X3KD outperforms previous approaches for multi- camera 3DOD on the nuScenes and Waymo datasets. • We transfer X3KD to RADAR-based 3DOD and train X3KD only through KD without using ground truth. • Our extensive ablation studies on nuScenes and Waymo provide a comprehensive evaluation of KD at different network stages for multi-camera 3DOD.
Liu_Multiple_Instance_Learning_via_Iterative_Self-Paced_Supervised_Contrastive_Learning_CVPR_2023	Abstract Learning representations for individual instances when only bag-level labels are available is a fundamental challenge in multiple instance learning (MIL). Recent works have shown promising results using contrastive self- supervised learning (CSSL), which learns to push apart representations corresponding to two different randomly- selected instances. Unfortunately, in real-world applica- tions such as medical image classiﬁcation, there is often class imbalance, so randomly-selected instances mostly be- long to the same majority class, which precludes CSSL from learning inter-class differences. To address this issue, we propose a novel framework, Iterative Self-paced Supervised Contrastive Learning for MIL Representations (ItS2CLR), which improves the learned representation by exploiting instance-level pseudo labels derived from the bag-level la- bels. The framework employs a novel self-paced sampling strategy to ensure the accuracy of pseudo labels. We evalu- ate ItS2CLR on three medical datasets, showing that it im- proves the quality of instance-level pseudo labels and repre- sentations, and outperforms existing MIL methods in terms of both bag and instance level accuracy.1	1. Introduction The goal of multiple instance learning (MIL) is to per- form classiﬁcation on data that is arranged in bags of in- stances. Each instance is either positive or negative, but these instance-level labels are not available during training; only bag-level labels are available. A bag is labeled as pos- itive if anyof the instances in it are positive, and negative otherwise. An important application of MIL is cancer diag- nosis from histopathology slides. Each slide is divided into *Equal Contribution †Joint Last Author 1Code is available at https://github.com/Kangningthu/ ItS2CLRhundreds or thousands of tiles but typically only slide-level labels are available [ 6,9,17,35,39,53]. Histopathology slides are typically very large, in the or- der of gigapixels (the resolution of a typical slide can be as high as 105⇥105), so end-to-end training of deep neu- ral networks is typically infeasible due to memory limi- tations of GPU hardware. Consequently, state-of-the-art approaches [ 6,35,39,44,53] utilize a two-stage learning pipeline: (1) a feature-extraction stage where each instance is mapped to a representation which summarizes its content, and (2) an aggregation stage where the representations ex- tracted from all instances in a bag are combined to produce a bag-level prediction (Figure 1). Notably, our results indi- cate that even in the rare settings where end-to-end training is possible, this pipeline still tends to be superior (see Sec- tion4.3). In this work, we focus on a fundamental challenge in MIL: how to train the feature extractor. Currently, there are three main strategies to perform feature-extraction, which have signiﬁcant shortcomings. (1) Pretraining on a large natural image dataset such as ImageNet [ 39,44] is problem- atic for medical applications because features learned from natural images may generalize poorly to other domains [ 38]. (2) Supervised training using bag-level labels as instance- level labels is effective if positive bags contain mostly posi- tive instances [ 11,34,50], but in many medical datasets this is not the case [ 5,35]. (3) Contrastive self-supervised learn- ing (CSSL) outperforms prior methods [ 14,35], but is not as effective in settings with heavy class imbalance, which are of crucial importance in medicine. CSSL operates by push- ing apart the representations of different randomly selected instances. When positive bags contain mostly negative in- stances, CSSL training ends up pushing apart negative in- stances from each other, which precludes it from learning features that distinguish positive samples from the negative ones (Figure 2). We discuss this ﬁnding in Section 2. Our goal is to address the shortcomings of current feature-extraction methods. We build upon several key in- sights. First, it is possible to extract instance-level pseudo This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted versio
Kolmogorov_Solving_Relaxations_of_MAP-MRF_Problems_Combinatorial_In-Face_Frank-Wolfe_Directions_CVPR_2023	Abstract We consider the problem of solving LP relaxations of MAP-MRF inference problems, and in particular the method proposed recently in [16, 35]. As a key computa- tional subroutine, it uses a variant of the Frank-Wolfe (FW) method to minimize a smooth convex function over a combi- natorial polytope. We propose an efﬁcient implementation of this subroutine based on in-face Frank-Wolfe directions , introduced in [4] in a different context. More generally, we deﬁne an abstract data structure for a combinatorial sub- problem that enables in-face FW directions, and describe its specialization for tree-structured MAP-MRF inference subproblems. Experimental results indicate that the result- ing method is the current state-of-art LP solver for some classes of problems. Our code is available at pub.ist. ac.at/ ˜vnk/papers/IN-FACE-FW.html .	1. Introduction The main focus of this paper is on the problem of min- imizing a function of discrete variables z= (z1;:::;zn) with unary and pairwise terms: min z2D1:::DnX v2[n]fv(zv) +X uv2Efuv(zu;zv)(1) HereG= ([n];E)is an undirected graph and D1;:::;Dn are ﬁnite sets. This problem is often referred to as MAP-MRF inference (maximum a posteriori inference in a Markov Random Field ). A prominent approach to tackle this NP-hard problem in practice is to solve its natural LP relaxation (see e.g. [39]), also called Basic LP relaxation [17]: min 0X v2[n] a2Dvfv(a)va+X uv2E (a;b)2DuDvfuv(a;b)ua;vb (2a) X b02Dvua;vb0=ua;X a02Duua0;vb=vb8uv;a;b (2b) X a2Dvua= 18v (2c)Designing algorithms to (approximately) solve this relax- ation for large-scale problems has been a very active area of research. A popular approach is to use message passing techniques, which perform a block-coordinate ascent on the dual objective [5, 11, 13, 36, 38, 40] This strategy is very ef- fective for some problems, but for other problems it may get stuck in a suboptimal point. Many techniques have been developed that are guaranteed to converge to the optimal so- lution of the LP relaxation [8,9,16,18,20,21,25,26,28–32, 34, 35]. In this paper we revisit the approach in [16, 35]. Its key computational subroutine is to minimize a quadratic con- vex function over combinatorial polytope, which is done by invoking a variant of the Frank-Wolfe (FW) algorithm [3]. We study efﬁcient implementations of the latter in the con- text of MAP-MRF inference. Our main contribution is in- corporating in-face FW directions introduced in [4]. The idea is to speed-up computations by running FW algorithm on a smaller “contracted” subproblem obtained by taking a face of the polytope containing the current point. It has been used for applications such as low-rank matrix com- pletion [4], cluster detection in networks [1], and training sparse neural networks with `1regularization [6]. We inves- tigate the use of in-face FW directions for general combina- torial polytopes, and describe an abstract data structure that enables such directions. We then specialize it to subprob- lems corresponding to tree-structured MAP-MRF inference problems. Our approach has the following features: It may happen that the contracted subproblem splits into independent subproblems. These subproblems are han- dled by a block-coordinate version of FW. We store a cache of “atoms” for each contracted sub- problem. We describe how to efﬁciently transform these atoms when the current face is recomputed. For an edgeuv2Eand ﬁxed fractional unary vectors for u;vwe can compute an optimal fractional pairwise vec- tor for edge uvby solving a small-scale optimal trans- portation (OT) problems. Such computations were used in [27] for computing primal feasible solutions of relax- 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. 11980 ation (2). We show how to use them for improving the performance of in-face FW directions. To our knowledge, the issues above have not been discussed in the literature so far. We remark that in-face FW directions effectively imple- ment the following rather natural idea: the optimization should be performed only over “active” pairs (v;a)that are likely to be present in the support of an optimal solution; the other pairs should be ﬁxed. A related idea appeared in the context of message passing algorithms in [37], where mes- sages are updated only in a subgraph in which the current best labels keep changing. The method in [37] works only with dual variables, and uses heuristic criteria for choosing the subgraph. We believe that in-face FW directions allow a more principled criterion for deciding which variables to ﬁx and for how long. Note that Freund et al. [4] proved that their criterion retains the convergence rate of the basic FW algorithm. We use a different criterion that also takes into account the ratio of runtimes on the original and on con- tracted subproblems. In Section 4 we test the algorithms on benchmark prob- lems in the evaluation [10], and compare them with LP solvers used in [10]. Results suggest that the method in [16, 35] with in-face FW directions is the current state- of-the-art LP solver for certain classes of problems.
Liu_InstMove_Instance_Motion_for_Object-Centric_Video_Segmentation_CVPR_2023	Abstract Despite significant efforts, cutting-edge video segmenta- tion methods still remain sensitive to occlusion and rapid movement, due to their reliance on the appearance of ob- jects in the form of object embeddings, which are vul- nerable to these disturbances. A common solution is to use optical flow to provide motion information, but essen- tially it only considers pixel-level motion, which still re- lies on appearance similarity and hence is often inaccu- rate under occlusion and fast movement. In this work, we study the instance-level motion and present InstMove, which stands for Instance Motion for Object-centric Video Segmentation. In comparison to pixel-wise motion, Inst- Move mainly relies on instance-level motion information that is free from image feature embeddings, and features physical interpretations, making it more accurate and ro- bust toward occlusion and fast-moving objects. To better fit in with the video segmentation tasks, InstMove uses in- stance masks to model the physical presence of an object and learns the dynamic model through a memory network to predict its position and shape in the next frame. With only a few lines of code, InstMove can be integrated into current SOTA methods for three different video segmentation tasks and boost their performance. Specifically, we improve the previous arts by 1.5 AP on OVIS dataset, which features heavy occlusions, and 4.9 AP on YouTubeVIS-Long dataset, which mainly contains fast moving objects. These results suggest that instance-level motion is robust and accurate, and hence serving as a powerful solution in complex sce- narios for object-centric video segmentation.	1. Introduction Segmenting and tracking object instances in a given video is a critical topic in computer vision, with vari- ous applications in video understanding, video editing, au- tonomous driving, augmented reality, etc. Three represen- tative tasks include video object segmentation (VOS), video *First two authors contributed equally. Work done during an internship at ByteDance. The code and models are available for research purposes at https://github.com/wjf5203/VNext. Framet-1Framet GTmaskFramet-1MaskOpticalFlowPropagatedMaskMotionMaskInstMove PreviousMasks+InstanceMotionPixelMotionFigure 1. Different from optical flow that estimates pixel-level motion, InstMove learns instance-level motion and deformation directly from previous instance masks and predicts more accurate and robust position and shape estimates for the current frame, even in scenarios with occlusions and rapid motion. instance segmentation (VIS), and multi-object tracking and segmentation (MOTS). These tasks differ significantly from video semantic segmentation [16, 19, 31, 52], which aims to classify every pixel in a video frame, hence we refer to them as object-centric video segmentation in this paper. De- spite significant progress, state-of-the-art (SOTA) methods are still struggle with occlusion, rapid motion, and signifi- cant changes in objects, resulting in a marked drop in han- dling longer or more complex videos. One reason we observe is that most methods rely solely on appearance to localize objects and track them across frames. Specifically, a majority of VOS methods [10, 30, 35, 40, 51, 67] use the previous frames as target templates and construct a feature memory bank of embeddings for all target objects. This is then used to match the pixel-level fea- ture in the new frame. Online VIS [6, 15, 21, 28, 65, 72, 73] and MOTS [25, 71] methods directly perform per-frame in- stance segmentation based on image features and use the object embeddings to track them through the video. While these paradigms work well on simple videos, they are sen- sitive to intense appearance changes and struggle with han- dling multiple object instances with similar appearances, re- sulting in large errors when dealing with complex scenarios 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. 6344 with complex motion patterns, occlusion, or deformation. Apart from appearance cues, object motion, which is an- other crucial piece of information provided by videos, has also been extensively studied for video segmentation. The majority of motion models in related fields fall into two cat- egories: One line of work uses optical flow to learn pixel- level motion. However, this approach does not help solve the problem of occlusion or fast motion since flow itself is often inaccurate in these scenarios [39,57]. The main reason causing the failure we argue is that optical flow still heavily relies on appearance cues to compute the pixel-level mo- tion across the frame. The other line of work uses a linear speed model, which helps alleviate these tracking problems caused by occlusion and fast motion in MOT [5,44,63,77]. However, it oversimplifies the problem and thus provides limited benefits in other tasks such as VOS and VIS. In this work, we aim at narrowing the gap between the two aforementioned lines of work by reformulating the mo- tion module and providing InstMove, a simple yet efficient motion prediction plugin that enjoys the advantages of both solutions. First, it is portable and is compatible with and beneficial to approaches of video segmentation tasks. More importantly, similar to optical flow, it also provides high- dimensional information of position and shape, which can be beneficial for a range of downstream tasks in a variety of ways, and, similar to the dynamic motion model, it learns physical interpretation to model motion information, im- proving robustness toward occlusion and fast motion. To achieve our objective, we utilize an instance mask to indicate the position and shape of a target object, and provide an RNN-based module with a memory network to extract motion features from previous masks, store and re- trieve dynamic information, and predict the position and shape information of the next frame based on motion cues. However, while being robust towards appearance changes, predicting shape without the object appearance or image features results in an even less accurate boundary in sim- ple cases. To solve this, we incorporate the low-level image features at the end of InstMove. Finally, to prove the effec- tiveness of InstMove on object-centric video segmentation tasks, we present two simple ways to integrate InstMove into recent SOTA methods in VIS, VOS, and MOTS, which improve their robustness with minimal modifications. In the experiments section, we first validate that our mo- tion module is more accurate and compatible with existing methods compared with learning motion and deformation with optical flow methods such as RAFT [55]. We also show that it is more robust to occlusion and rapid move- ments. Then we demonstrate the improvement of integrat- ing our motion plugin into all recent SOTA methods in VOS, VIS, and MOTS tasks, particularly in complex sce- narios with heavy occlusion and rapid motion. Remarkably, with only a few lines of code, we significantly boost the cur-rent art by 1.5 AP on OVIS [47], 4.9 AP on YouTubeVIS- Long [69], and reduce IDSw on BDD100K [75] by 28.6%. In summary, we have revisited the motion models used in video segmentation tasks and propose InstMove, which contains both pixel-level information and instance-level dy- namic information to predict shape and position. It pro- vides additional information that is robust to occlusion and rapid motion. The improvements in SOTA methods of all three tasks demonstrate the effectiveness of incorporating instance-level motion in tackling complex scenarios.
Li_Exploring_the_Effect_of_Primitives_for_Compositional_Generalization_in_Vision-and-Language_CVPR_2023	Abstract Compositionality is one of the fundamental properties of human cognition (Fodor & Pylyshyn, 1988). Compositional generalization is critical to simulate the compositional ca- pability of humans, and has received much attention in the vision-and-language (V&L) community. It is essential to understand the effect of the primitives, including words, im- age regions, and video frames, to improve the compositional generalization capability. In this paper, we explore the ef- fect of primitives for compositional generalization in V&L. Specifically, we present a self-supervised learning based framework that equips existing V&L methods with two char- acteristics: semantic equivariance andsemantic invari- ance . With the two characteristics, the methods understand primitives by perceiving the effect of primitive changes on sample semantics and ground-truth. Experimental results on two tasks: temporal video grounding and visual question answering, demonstrate the effectiveness of our framework.	1. Introduction Compositionality is one of the fundamental properties of human cognition argued by Fodor and Pylyshyn [11]. Com- positional generalization in vision-and-language (V&L) has received increasing attention and significant progress in re- cent years, but has not been fully explored. Compositional generalization requires V&L methods to generalize well to sentences with novel combinations of seen words, which is critical to simulate the compositional properties of human cognition. *Corresponding author: Chenchen Jing and Yuwei Wu Query: Aperson opens the door. \| \|Video: Query: The person opens adoor. \| \|Query: Aperson closes the door. \| \|Figure 1. An example in the context of temporal video grounding, showing that primitives are the determinants of sample semantics and ground-truth. An indispensable premise for improving compositional generalization is to understand the effect of the primitives, including words, image regions, and video frames. Primi- tives are compositional building blocks mainly involved in V&L tasks and the determinants of sample semantics. For example, for a sample with the query “A person opens the door” in the context of temporal video grounding (TVG), its semantics are changed completely when the primitive “opens” is changed to “closed”, but are unchanged when the primitives “A” and “the” are modified to “The” and “a”, respectively, as shown in Fig. 1. We investigate if existing V&L methods are sensitive to the sample semantic changes brought by primitive changes. Our observations show that the methods erroneously keep almost 90% of the predic- tions unchanged when the sample semantics are corrupted by replacing 50% critical words ( e.g., nouns, verbs) in sen- tences. This suggests that existing methods cannot correctly establish the relationship between the primitives and the sample semantics and thus the ground-truth, so they cannot achieve compositional generalization. In this paper, we explore the effect of primitives for com- positional generalization from two aspects: semantic equiv- 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. 19092 (a) An original example in the context of temporal video grounding.Prediction r \| \| Query: A person is smiling next to a refrigerator. (b) Equivariant samples generated by masking critical primitives. (c) Invariant samples generated by masking irrelevant primitives.Query: A [MASK] is [MASK] next to a refrigerator. Query: A person is smiling next to a refrigerator. Prediction v \| \| Prediction v \| \| Query: A person is smiling [MASK] to a refrigerator. Query: A person is smiling next to a refrigerator. Prediction r \| \| Prediction r \| \|Figure 2. The samples generated in our framework by masking different primitives. ariance andsemantic invariance . Semantic equivariance means that the predictions of the methods should be equiv- ariant with the sample semantics, which are determined by the primitives. Once the semantics of the sample are changed, the predictions of the methods should faithfully change. To ensure semantic equivariance, methods are en- couraged to learn which primitives have a high effect on the sample semantics and thus the ground-truth. Seman- tic invariance means that the methods should maintain the same predictions when irrelevant primitives ( e.g., function words and background in visual content) are changed. This helps the methods to learn the primitives with low richness semantics and a low effect on ground-truth, and is comple- mented with the semantic equivariance. With the two char- acteristics, the methods understand the effect of primitive changes on sample semantics and ground-truth. We propose a self-supervised learning based framework to equip existing methods with semantic equivariance and semantic invariance . By masking critical and irrelevant primitives, we generate numerous labeled training samples, including equivariant samples and invariant samples, re- spectively, as shown in Fig. 2. To assign labels to the gen- erated samples, we estimate the effect of masked primitives on ground-truth. The larger the effect of the masked prim- itives, we assign the generated sample with a label more different from the ground-truth of the original sample. By training with the generated samples, the methods learn to make equivariant and invariant predictions when the sam- ple semantics change and do not, respectively. Extensiveexperiments on two V&L tasks: temporal video grounding [2] and visual question answering [3], demonstrate that our framework improves the compositional generalization ca- pability of existing methods. In summary, our contributions are as follows: • We explore the effect of primitives on improving the compositional generalization capability of exist- ing V&L methods by perceiving the effect of primitive changes on sample semantics and ground-truth. • We propose a self-supervised learning based frame- work for compositional generalization, in which nu- merous labeled samples are generated to equip existing V&L methods with semantic equivariance and seman- tic invariance.
Liu_Marching-Primitives_Shape_Abstraction_From_Signed_Distance_Function_CVPR_2023	Abstraction from Signed Distance Function Weixiao Liu1,2Yuwei Wu1Sipu Ruan1Gregory S. Chirikjian1* 1National University of Singapore2Johns Hopkins University {mpewxl, yw.wu, ruansp, mpegre }@nus.edu.sg Abstract Representing complex objects with basic geometric primitives has long been a topic in computer vision. Primitive-based representations have the merits of com- pactness and computational efficiency in higher-level tasks such as physics simulation, collision checking, and robotic manipulation. Unlike previous works which extract polyg- onal meshes from a signed distance function (SDF), in this paper, we present a novel method, named Marching- Primitives, to obtain a primitive-based abstraction directly from an SDF . Our method grows geometric primitives (such as superquadrics) iteratively by analyzing the connectivity of voxels while marching at different levels of signed dis- tance. For each valid connected volume of interest, we march on the scope of voxels from which a primitive is able to be extracted in a probabilistic sense and simulta- neously solve for the parameters of the primitive to cap- ture the underlying local geometry. We evaluate the per- formance of our method on both synthetic and real-world datasets. The results show that the proposed method out- performs the state-of-the-art in terms of accuracy, and is di- rectly generalizable among different categories and scales. The code is open-sourced at https://github.com/ ChirikjianLab/Marching-Primitives.git .	1. Introduction Recent years have witnessed great progress in the areas of 3D shape representation and environmental perception. Low-level representations such as surface meshes, point clouds, and occupancy grids are widely used as inputs to high-level computer vision algorithms and artificial intel- ligence tasks. They have the advantage of being able to represent and visualize objects with high accuracy and rich local geometric features. However, the low-level represen- tations are ineffective in delivering a general and intuitive sense of structural geometry as well as part-level scene un- derstanding. Studies [3, 20] show that human vision, unlike *Corresponding author Figure 1. Primitive-base representation versus mesh. For each pair of objects, the left one is the superquadric abstraction obtained by our algorithm, and the right one is the original mesh. The mesh of the chair is 6MB in size, while our representation only needs 4KB. An SDF representation discretized on a 1283voxel grid oc- cupies 19MB. Our abstraction is equivalent to an implicit continu- ous SDF, which is an approximation to the discrete SDF. computer vision, tends to perceive and understand scenes as combinations of simple primitive shapes. Human beings perform well and robustly in complex tasks, providing a basic geometric description of the scene is available [32]. Therefore, researchers turn to exploring the possibility of interpreting complex objects and scenes with basic geo- metric primitives. Taking advantage of the primitive-based representation, many higher-level tasks, such as segmenta- tion [14, 16, 21, 30], scene understanding [29, 31, 41, 47], grasping [33, 44, 45] and motion planning [35, 36], are able to be solved efficiently. However, it still remains challenging to extract primitive- based abstractions from low-level representations. Start- ing from the 1990s, Solina et al. [1, 17, 39] aim to extract a single superquadric representation from a simple object by minimizing the least-square error between the primitive and the measured points. Later in [7, 22], their method is extended to represent more complex objects with multiple primitives. More recently, the authors of [24, 47] reformu- late the task as a probabilistic inference problem with en- hanced accuracy and robustness to noise and outliers. At the same time, with the surge of data-driven techniques, re- searchers attempt to train neural networks to infer cuboids [27,38,41,48,50] and superquadrics [29,31] representations in an end-to-end fashion. However, both the computational and learning-based approaches have their own limitations. The computational methods are vulnerable to the inherent ambiguity of the point-to-surface relationship. For exam- 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. 8771 ple, the algorithms tend to fill empty spaces of a non-convex object with primitives by mistake, due to the inside/outside ambiguity of a surface depicted by a set of points [24, 48]. The main drawback of the learning approaches lies in the lack of generalizability beyond the object category on which the model is trained [24,31,47,48]. Also, the shape abstrac- tion accuracy is inferior to the computational methods. The signed distance function (SDF) has been a success- ful 3D volumetric representation in varieties of computer vi- sion and graphics tasks. It is the basic framework for many classic 3D reconstruction algorithms such as TSDF volume reconstruction [10, 15], KinectFusion [19], and Dynamic- Fusion [26]. Recently, the SDF representation is adapted to the deep learning frameworks, and exhibits boosted poten- tials in shape encoding [8, 18, 28, 43], surface reconstruc- tion [23, 46], and shape completion [11, 12, 34]. Usually, triangular mesh surfaces are extracted from the SDF rep- resentation with the marching cubes algorithm [25]. Point cloud and occupancy grid representations are also obtained by keeping the vertices of the meshes and the sign of each voxel point, respectively. The SDF is among the most in- formative 3D representations since it encodes not only the surface geometry but also the distance and side of a point relative to the shape. Meanwhile, it is easily achievable via range images from 3D sensors [10], or learnable from other input modalities [8, 18, 28]. Since we are able to extract meshes from an SDF, it is natural to think about the pos- sibility of extracting primitives as well. Furthermore, the primitive-based abstraction is a continuous interpretation of the complete geometric information encoded in the original discrete SDF, but requires much less storage size (Fig.1). Motivated by the aforementioned facts and the bottle- neck of the current shape abstraction algorithms, we pro- posed a general shape abstraction method by reasoning di- rectly on the informative SDF representation. The goal of our method is to find a combination of geometric primi- tives whose underlying SDF values match the target values evaluated on the evenly spaced discrete grid points (Sec. 3.1 and Sec. 3.2). To solve this problem, we propose a two-step iterative algorithm called the Marching-Primitives. Our algorithm ‘marches’ on two domains: the signed dis- tance domain and the voxelized space domain, alternately. Firstly, the connectivity of volumes are analyzed by gen- erating isosurfaces on a sequence of decreasing levels of negative signed distances (Sec.3.3). By doing so, volumes of interest (VOIs) where primitives are likely to be encoded can be identified sequentially. In the second step, for each of the VOIs, our algorithm marches on the neighbouring vox- els to infer their probabilistic correspondences to the primi- tive and simultaneously optimizes the shape and pose of the primitive (Sec.3.4). After the primitive representation of a VOI is achieved, the fitted volumes are deactivated from the voxel grid. Our algorithm continues marching on the signeddistance domain until it approaches zero, i.e., all the inte- rior volumes of the SDF have been captured by the recov- ered primitives. We compare our algorithm with the state- of-the-art of both the computational and learning-based ap- proaches on the ShapeNet object dataset [6] and D-FAUST human shape dataset [4] (Sec. 4.1). We also study the per- formance of our algorithm on different conditions(Sec. 4.2). Finally, we demonstrate the scene abstraction result of the Stanford Reading Room [49], which contains several pieces of furniture of various categories(Sec. 4.3).
Liu_SAP-DETR_Bridging_the_Gap_Between_Salient_Points_and_Queries-Based_Transformer_CVPR_2023	Abstract Recently, the dominant DETR-based approaches apply central-concept spatial prior to accelerating Transformer detector convergency. These methods gradually refine the reference points to the center of target objects and imbue ob- ject queries with the updated central reference information for spatially conditional attention. However, centralizing ref- erence points may severely deteriorate queries’ saliency and confuse detectors due to the indiscriminative spatial prior. To bridge the gap between the reference points of salient queries and Transformer detectors, we propose SAlient Point-based DETR (SAP-DETR ) by treating object detec- tion as a transformation from salient points to instance ob- jects. Concretely, we explicitly initialize a query-specific reference point for each object query, gradually aggregate them into an instance object, and then predict the distance from each side of the bounding box to these points. By rapidly attending to query-specific reference regions and the conditional box edges, SAP-DETR can effectively bridge the gap between the salient point and the query-based Trans- former detector with a significant convergency speed. Exper- imentally, SAP-DETR achieves 1.4 ×convergency speed with competitive performance and stably promotes the SoTA ap- proaches by ∼1.0 AP . Based on ResNet-DC-101, SAP-DETR achieves 46.9 AP . The code will be released at https: //github.com/liuyang-ict/SAP-DETR.	1. Introduction Object detection is a fundamental task in computer vi- sion, whose target is to recognize and localize each ob- ject from input images. In the last decade, various detec- *This work was done when working as an intern at AI Lab, Lenovo Research, Beijing, China. †Corresponding author. Figure 1. Comparison of SAP-DETR and DAB-DETR under 3- layer decoder model and 36-epoch training scheme. (a) Statistics of the query count in different classification score intervals. (b) and (c) Visualize the distribution of all reference points (pink) and highlight the top-20 classification score queries with their bounding boxes (blue) and reference points (red) in different decoder layers. (d) Visualize the outputs of positive queries and ground truth (red) during the training process, each query has a representative color for its reference point and bounding box. Best viewed in color. tors [6,11,14,18,20,22] based on Convolutional Neural Net- works (CNNs), have received widespread attention and made significant progress. Recently, Carion et al. [2] proposed a new end-to-end paradigm for object detection based on the Transformer [24], called DEtection TRansformer (DETR), which treats object detection as a problem of set prediction. In DETR, a set of learnable positional encodings, namely object queries, are employed to aggregate instance features from the context image in Transformer Decoder. The predic- tions of queries are finally assigned to the ground truth via bipartite matching to achieve end-to-end detection. Despite the promising results of DETR, its application is largely limited by considerably longer training time com- pared to conventional CNNs. To address this problem, many 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. 15539 variants attempted to take a close look at query paradigm and introduced various spatial priors for model convergency and efficacy. According to the type of spatial prior, they can be categorized into implicit and explicit methods. The implicit ones [5, 16, 31] attempt to decouple a reference point from the object query and make use of this spatial prior to attend to the image context features efficiently. The current state-of- the-arts (SoTAs) are dominated by the explicit ones [13, 25], which suggest to instantiate a position with spatial prior for each query, i.e., explicit reference coordinates with a center point or an anchor box. These reference coordinates serve as helpful priors and enable the queries to focus on their expected regions easily. For instance, Anchor DETR [25] in- troduced an anchor concept (center point with different box size patterns) to formulate the query position and directly regressed the central offsets of the bounding boxes. DAB- DETR [13] further stretched the center point to a 4D anchor box concept [cx, cy, w, h ]to refine proposal bonding boxes in a cascaded manner. However, instantiating the query loca- tion as a target center may severely degrade the classification accuracy and convergency speed. As illustrated in Fig. 1, there exist many plausible queries [19] with high-quality classification scores (Fig. 1(a) within red box) and box In- tersection over Union (IoU, see the redundant blue boxes in Fig. 1(b) and (c)), which only brings a slight improvement on precision rate but inevitably confuses the detector on the positive query assignments when training with bipartite matching strategy. This is because the plausible predictions are considered in negative classification loss, which severely decelerates the model convergency. As shown in Fig. 1(b) and (c), the predefined reference point of the positive query may not be the nearest one to the center of the ground truth bounding box, and the reference points tend to be centralized or marginalized (cyan arrows in Fig. 1(b)), hence losing the spatial specificity. With further insight into the one-to-one label assignment during the training process, we find that the query, whose reference point is closest to the center point, also has a high-quality IoU, but it still exists a disparity with the positive query in the classification confidence. Therefore, we argue that such a centralized spatial prior may cause de- generation of target consistency in both classification and localization tasks, which leads to inconsistent predictions. Furthermore, the mentioned central point-based variants also have difficulties in detecting occluded objects. For ex- ample, Fig. 1(d) shows that DAB-DETR detects the left baseman twice, while the query point in SAP-DETR is not necessarily the center point, so the query point of the bound- ing box for the occluded baseman is from a pixel on the occluded baseman on the top right area instead of from the left baseman. One solution for center-based method [25] is to predefine different receptive fields (similar to the scaling anchor box in YOLO [17]) for the position of each query. However, increasing the diversity of the receptive fields foreach position query is unsuitable for non-overlapped targets, as it still generates massive indistinguishable predictions for one position as same as other center-based models. To bridge these gaps, in this paper, we present a novel framework for Transformer detector, called SAlient Point- based DETR (SAP-DETR), which treats object detection as a transformation from salient points to instance objects. Instead of regressing the reference point to the target center, we define the reference point belonging to one positive query as a salient point , keep this query-specific spatial prior with a scaling amplitude, and then gradually update them to an in- stance object by predicting the distance from each side of the bounding box. Specifically, we tile the mesh-grid referenced points and initialize their center/corner as the query-specific reference point. To disentangle the reference sparsity as well as stabilize the training process, a movable strategy with scaling amplitude is applied for reference point adjustment, which prompts queries to consider their reference grid as the salient region to perform image context attention. By localizing each side of the bounding box layer by layer, such query-specific spatial prior enables compensation for the over-smooth/inadequacy problem during center-based de- tection, thereby vastly promoting model convergency speed. Inspired by [5, 13, 16], we also take advantage of both Gaus- sian spatial prior and conditional cross-attention mechanism, and then a salient point enhanced cross-attention mecha- nism is developed to distinguish the salient region and other conditional extreme regions from the context image features. We bridge the gap between salient points and query-based Transformer detector by speedily attending to the query- specific region and other conditional regions. The extensive experiments have shown that SAP-DETR achieves superior convergency speed and performance. To the best of our knowledge, this is the first work to introduce the salient point based regression into end-to-end query-based Transformer detectors. Our contributions can be summarized as follows. 1)We introduce the salient point concept into query-based Transformer detectors by assigning query-specific reference points to object queries. Unlike center-based methods, we restrict the reference location and define the point of the posi- tive query as the salient one , hence enlarging the discrepancy of query as well as reducing the redundant predictions (see Fig. 1). Thanks to the efficacy of the query-specific prior, our SAP-DETR accelerates the convergency speed greatly, achieving competitive performance with 30% fewer train- ing epochs. The proposed movable strategy further boosts SAP-DETR to a new SoTA performance. 2)We devise a point-enhanced cross-attention mechanism to imbue query with spatial prior based on both reference point and box sides for final specific region attention. 3)Evaluation over COCO dataset has demonstrated that SAP-DETR achieves superior convergency speed and detec- tion accuracy. Under the same training settings, SAP-DETR 15540 outperforms the SoTA approaches with a large margin.
Li_DropKey_for_Vision_Transformer_CVPR_2023	Abstract In this paper, we focus on analyzing and improving the dropout technique for self-attention layers of Vision Trans- former, which is important while surprisingly ignored by prior works. In particular, we conduct researches on three core questions: First, what to drop in self-attention layers? Different from dropping attention weights in literature, we propose to move dropout operations forward ahead of atten- tion matrix calculation and set the Key as the dropout unit, yielding a novel dropout-before-softmax scheme. We theo- retically verify that this scheme helps keep both regulariza- tion and probability features of attention weights, alleviat- ing the overfittings problem to specific patterns and enhanc- ing the model to globally capture vital information; Second, how to schedule the drop ratio in consecutive layers? In contrast to exploit a constant drop ratio for all layers, we present a new decreasing schedule that gradually decreases the drop ratio along the stack of self-attention layers. We experimentally validate the proposed schedule can avoid overfittings in low-level features and missing in high-level semantics, thus improving the robustness and stableness of model training; Third, whether need to perform struc- tured dropout operation as CNN? We attempt patch-based block-version of dropout operation and find that this use- ful trick for CNN is not essential for ViT. Given exploration on the above three questions, we present the novel Drop- Key method that regards Key as the drop unit and exploits decreasing schedule for drop ratio, improving ViTs in a gen- eral way. Comprehensive experiments demonstrate the ef- fectiveness of DropKey for various ViT architectures, e.g. T2T, VOLO, CeiT and DeiT, as well as for various vision *Equal contribution †Corresponding authortasks, e.g., image classification, object detection, human- object interaction detection and human body shape recov- ery.	1. Introduction Vision Transformer (ViT) [6] has achieved great suc- cess for various vision tasks, e.g., image recognition [7, 12, 20, 34, 35], object detection [1], human body shape esti- mation [18], etc. Prior works mainly focus on researches of patch division, architecture design and task extension. However, the dropout technique for self-attention layer, which plays the essential role to achieve good generaliz- ability, is surprisingly ignored by the community. Different from the counterpart for Convolutional Neural Networks (CNNs), the dropout in ViT directly utilizes the one in original Transformer designed for Natural Language Processing, which sets attention weights as the manipula- tion unit with a constant dropout ratio for all layers. Despite of its simplicity, this vanilla design faces three major prob- lems. First, it breaks the probability distribution of atten- tion weights due to the averaging operation on non-dropout units after softmax normalization. Although this regularizes the attention weights, it still overfits specific patterns locally due to the failure on penalizing score peaks, as shown in Fig. 1 (a) and (b); Second, the vanilla design is sensitive to the constant dropout ratio, since high ratio occurs missing of semantic information in high-level representations while low ratios overfitting in low-level features, resulting in the unstable training process; Third, it ignores the structured characteristic of input patch grid to ViT, which plays an ef- fective role to improve performance with blockwise dropout in CNNs. These three problems degrade the performance and limit the generalizability of ViTs. Motivated by this, we propose to analyze and improve 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. 22700 the dropout technique in self-attention layer, further push- ing forward the frontier of ViTs for vision tasks in a general way. Specifically, we focus on three core aspects: What to drop in self-attention layer Different from drop- ping attention weights as in the vanilla design, we propose to set the Key as the dropout unit, which is essential input of self-attention layer and significantly affects the output. This moves the dropout operation forward before calculat- ing the attention matrix as shown in Fig. 1 (c) and yields a novel dropout-before-softmax scheme. This scheme reg- ularizes attention weights and keeps their probability dis- tribution at the same time, which intuitively helps penalize weight peaks and lift weight foots. We theoretically verify this property via implicitly introducing an adaptive smooth- ing coefficient for the attention operator from the perspec- tive of gradient optimization by formulating a Lagrange function. With the dropout-before-softmax scheme, self- attention layers can capture vital information in a global manner, thus overcoming the overfittings problem to spe- cific patterns occurred in the vanilla dropout and enhancing the model generalizability as visualization of feature map in Fig. 1 (c). For the training phase, this scheme can be simply implemented by swapping the operation order of softmax and dropout in vanilla design, which provides a general way to effectively enhance ViTs. For inference phase, we con- duct an additional finetune phase to align the expectations to training phase, further improving the performance. How to schedule the drop ratio In contrast to exploiting a constant drop ratio for all layers, we present a new linear decreasing schedule that gradually decreases the drop ratio along the stack of self-attention layers. This schedule leads to a high drop ratio in shallow layers while the low one in deep layers, thus avoiding overfittings to low-level features and preserving sufficient high-level semantics. We experi- mentally verify the effectiveness of the proposed decreasing schedule for drop ratio to stable the training phase and im- prove the robustness. Whether need to perform structured drop Inspired by the DropBlock [10] method for CNNs, we implement two structured versions of the dropout operation for ViTs: the block-version dropout that drops keys corresponding to contiguous patches in images or feature maps; the cross- version dropout that drops keys corresponding to patches in horizontal and vertical stripes. We conduct thorough ex- periments to validate their efficacy and find that the struc- ture trick useful for CNN is not essential for ViT, due to the powerful capability of ViT to grasp contextual information in full image range. Given exploration on the above three aspects, we present a novel DropKey method that utilizes Key as the drop unit and decreasing schedule for drop ratio. In particular, Drop- Key overcomes drawbacks of the vanilla dropout technique for ViTs, improving performance in a general and effectiveway. Comprehensive experiments on different ViT archi- tectures and vision tasks demonstrate the efficacy of Drop- Key. Our contributions are in three folds: First, to our best knowledge, we are the first to theoretically and experimen- tally analyze dropout technique for self-attention layers in ViT from three core aspects: drop unit, drop schedule and structured necessity; Second, according to our analysis, we present a novel DropKey method to effectively improve the dropout technique in ViT. Third, with DropKey, we improve multiple ViT architectures to achieve new SOTAs on vari- ous vision tasks.
Liang_Visual_Exemplar_Driven_Task-Prompting_for_Unified_Perception_in_Autonomous_Driving_CVPR_2023	Abstract Multi-task learning has emerged as a powerful paradigm to solve a range of tasks simultaneously with good efficiency in both computation resources and inference time. However, these algorithms are designed for different tasks mostly not within the scope of autonomous driving, thus making it hard to compare multi-task methods in autonomous driving. Aiming to enable the comprehensive evaluation of present multi-task learning methods in autonomous driving, we ex- tensively investigate the performance of popular multi-task methods on the large-scale driving dataset, which covers four common perception tasks, i.e., object detection, seman- tic segmentation, drivable area segmentation, and lane de- tection. We provide an in-depth analysis of current multi- task learning methods under different common settings and find out that the existing methods make progress but there is still a large performance gap compared with single-task baselines. To alleviate this dilemma in autonomous driving, we present an effective multi-task framework, VE-Prompt, which introduces visual exemplars via task-specific prompt- ing to guide the model toward learning high-quality task- specific representations. Specifically, we generate visual exemplars based on bounding boxes and color-based mark- ers, which provide accurate visual appearances of target categories and further mitigate the performance gap. Fur- thermore, we bridge transformer-based encoders and con- volutional layers for efficient and accurate unified percep- tion in autonomous driving. Comprehensive experimental results on the diverse self-driving dataset BDD100K show that the VE-Prompt improves the multi-task baseline and further surpasses single-task models.	1. Introduction Multi-task learning (MTL) has been the source of a num- ber of breakthroughs in autonomous driving over the last †Corresponding author. [CLASS]A photo of a [class]. ...Text Encoder ImageImage EncoderAlignment (a) Textual PromptImage Patch...Embed... ... Learnable Prompts Transformer (b) Visual Prompt ... ... ... ...Prompt Generator ImageImage EncoderTask Prompts Task PromptingHead A Head B Head C Head D (c) Visual Exemplar Driven Prompt (Ours)Figure 1. Comparison of different prompts in computer vi- sion. (a) Extracting textual prompts from a text encoder to per- form image-text alignment [62]. (b) Prepend learnable prompts to the embeddings of image patches [20]. (c) Visual exemplar driven prompts for multi-task learning (ours). The generated task prompts encode high-quality task-specific knowledge for down- stream tasks. few years [23, 50, 56] and general vision tasks recently [2, 12, 26, 34, 55]. As the foundation of autonomous driv- ing, a robust vision perception system is required to pro- vide critical information, including the position of traffic participants, traffic signals like lights, signs, lanes, and ob- stacles that influence the drivable space, to ensure driving safety and comfort. These tasks gain knowledge from the same data source and present prominent relationships be- tween each other, like traffic participants, are more likely to appear within drivable spaces and traffic signs may appear near traffic lights, etc. Training these tasks independently is time costing and fails to mine the latent relationship among them. Therefore, it is crucial to solve these multiple tasks simultaneously, which can improve data efficiency and re- duce training and inference time. Some recent works have attempted to apply unified train- ing on multiple tasks in autonomous training. Uncertainty [21] trains per-pixel depth prediction, semantic segmenta- 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. 9611 tion, and instance segmentation in a single model. CIL [19] introduces an extra traffic light classifier to learn different traffic patterns following traffic light changes. CP-MTL [4] learns object detection and depth prediction together to identify dangerous traffic scenes. However, these works dif- fer in task types, evaluation matrix, and dataset, making it hard to compare their performances. For example, most of them are developed upon dense prediction [2, 55] and natu- ral language understanding [7,47], rather than being tailored for more common perception tasks for autonomous driving, thus these methods may produce poor results when applied to a self-driving system. As a result, there is an emerg- ing demand for a thorough evaluation of existing multi-task learning methods covering common tasks in autonomous driving. In this paper, we focus on heterogeneous multi-task learning in common scenarios of autonomous driving and cover popular self-driving tasks, i.e., object detection, se- mantic segmentation, drivable area segmentation, and lane detection. We provide a systematic study of present MTL methods on large-scale driving dataset BDD100K [58]. Specifically, we find that task scheduling [26] is better than zeroing loss [51], but worse than pseudo labeling [15] on most tasks. Interestingly, in task-balancing methods, Un- certainty [21] produces satisfactory results on most tasks, while MGDA [41] only performs well on lane detection. This indicates that negative transfer [8], which is a phe- nomenon that increasing the performance of a model on one task will hurt the performance on another task with different needs, is common among these approaches. To mitigate the negative transfer problem, we introduce the visual exemplar-driven task-prompting (shorten as VE- Prompt ) based on the following motivations: (1) Given the visual clues of each task, the model can extract task-related information from the pre-trained model. Different from cur- rent prompting methods which introduce textual prompts [6, 38, 62, 63] or learnable context [20], we leverage exem- plars containing information of target objects to generate task-specific prompts by considering that the visual clues should represent the specific task to some extent, and give hints for learning task-specific information; (2) Transformer has achieved competitive performance on many vision tasks but usually requires long training time, thus tackling four tasks simultaneously on a pure transformer is resource- intensive. To overcome this challenge, we efficiently bridge transformer encoders and convolutional layers to build the hybrid multi-task architecture. Extensive experiments show that VE-Prompt surpasses multi-task baselines by a large margin. We summarize the main contributions of our work be- low: • We provide an in-depth analysis of current multi-task learning approaches under multiple settings that complywith real-world scenarios, consisting of three common multi-task data split settings, two partial-label learning approaches, three task scheduling techniques, and three task balancing strategies. • We propose an effective framework VE-Prompt, which utilizes visual exemplars to provide task-specific visual clues and guide the model toward learning high-quality task-specific representations. • The VE-Prompt framework is constructed in a computa- tionally efficient way and outperforms competitive multi- task methods on all tasks.
Li_DynIBaR_Neural_Dynamic_Image-Based_Rendering_CVPR_2023	Abstract We address the problem of synthesizing novel views from a monocular video depicting a complex dynamic scene. State- of-the-art methods based on temporally varying Neural Ra- diance Fields (aka dynamic NeRFs ) have shown impressive results on this task. However, for long videos with com- plex object motions and uncontrolled camera trajectories, these methods can produce blurry or inaccurate renderings, hampering their use in real-world applications. Instead of encoding the entire dynamic scene within the weights of MLPs, we present a new approach that addresses these lim- itations by adopting a volumetric image-based rendering framework that synthesizes new viewpoints by aggregating features from nearby views in a scene motion–aware manner. Our system retains the advantages of prior methods in its ability to model complex scenes and view-dependent effects, but also enables synthesizing photo-realistic novel views from long videos featuring complex scene dynamics with unconstrained camera trajectories. We demonstrate signiﬁ-cant improvements over state-of-the-art methods on dynamic scene datasets, and also apply our approach to in-the-wild videos with challenging camera and object motion, where prior methods fail to produce high-quality renderings.	1. Introduction Computer vision methods can now produce free- viewpoint renderings of static 3D scenes with spectacular quality. What about moving scenes, like those featuring peo- ple or pets? Novel view synthesis from a monocular video of a dynamic scene is a much more challenging dynamic scene reconstruction problem. Recent work has made progress towards synthesizing novel views in both space and time, thanks to new time-varying neural volumetric representa- tions like HyperNeRF [ 50] and Neural Scene Flow Fields (NSFF) [ 35], which encode spatiotemporally varying scene content volumetrically within a coordinate-based multi-layer perceptron (MLP). However, these dynamic NeRF methods have limitations 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. 4273 that prevent their application to casual, in-the-wild videos. Local scene ﬂow–based methods like NSFF struggle to scale to longer input videos captured with unconstrained camera motions: the NSFF paper only claims good perfor- mance for 1-second, forward-facing videos [ 35]. Methods like HyperNeRF that construct a canonical model are mostly constrained to object-centric scenes with controlled camera paths, and can fail on scenes with complex object motion. In this work, we present a new approach that is scalable to dynamic videos captured with 1) long time duration, 2) un- bounded scenes, 3) uncontrolled camera trajectories, and 4) fast and complex object motion. Our approach retains the ad- vantages of volumetric scene representations that can model intricate scene geometry with view-dependent effects, while signiﬁcantly improving rendering ﬁdelity for both static and dynamic scene content compared to recent methods [ 35,50], as illustrated in Fig. 1. We take inspiration from recent methods for rendering static scenes that synthesize novel images by aggregating local image features from nearby views along epipolar lines [ 39,64,70]. However, scenes that are in motion vi- olate the epipolar constraints assumed by those methods. We instead propose to aggregate multi-view image features in scene motion–adjusted ray space, which allows us to cor- rectly reason about spatio-temporally varying geometry and appearance. We also encountered many efﬁciency and robustness chal- lenges in scaling up aggregation-based methods to dynamic scenes. To efﬁciently model scene motion across multiple views, we model this motion using motion trajectory ﬁelds that span multiple frames, represented with learned basis functions. Furthermore, to achieve temporal coherence in our dynamic scene reconstruction, we introduce a new tem- poral photometric loss that operates in motion-adjusted ray space. Finally, to improve the quality of novel views, we pro- pose to factor the scene into static and dynamic components through a new IBR-based motion segmentation technique within a Bayesian learning framework. On two dynamic scene benchmarks, we show that our approach can render highly detailed scene content and sig- niﬁcantly improves upon the state-of-the-art, leading to an average reduction in LPIPS errors by over 50% both across entire scenes, as well as on regions corresponding to dynamic objects. We also show that our method can be applied to in- the-wild videos with long duration, complex scene motion, and uncontrolled camera trajectories, where prior state-of- the-art methods fail to produce high quality renderings. We hope that our work advances the applicability of dynamic view synthesis methods to real-world videos.
Liu_AdaptiveMix_Improving_GAN_Training_via_Feature_Space_Shrinkage_CVPR_2023	Abstract Due to the outstanding capability for data generation, Gen- erative Adversarial Networks (GANs) have attracted considerable attention in unsupervised learning. However, training GANs is difficult, since the training distribution is dynamic for the discriminator, leading to unstable image representation. In this paper, we address the problem of training GANs from a novel perspective, i.e., robust image classification. Motivated by studies on robust image representation, we propose a simple yet effective module, namely AdaptiveMix, for GANs, which shrinks the regions of training data in the image representation space of the discriminator. Considering it is intractable to directly bound feature space, we propose to construct hard samples and narrow down the feature distance between hard and easy samples. The hard samples are constructed by mixing a pair of training images. We evaluate the effectiveness of our AdaptiveMix with widely-used and state-of-the-art GAN architectures. The evaluation results demonstrate that our AdaptiveMix can facilitate the training of GANs and effectively improve the image quality of generated samples. We also show that our AdaptiveMix can be further applied to image classification and Out-Of-Distribution (OOD) detection tasks, by equipping it with state-of-the- art methods. Extensive experiments on seven publicly available datasets show that our method effectively boosts the performance of baselines. The code is publicly avail- able at https://github.com/WentianZhang- ML/AdaptiveMix .	1. Introduction Artificial Curiosity [40, 41] and Generative Adversar- ial Networks (GANs) have attracted extensive attention due to their remarkable performance in image generation †Equal Contribution Corresponding Authors: Bing Li and Yuexiang Li. StyleGAN2StyleGAN2+AdaptiveMixAFHQ-Cat-5k(5,153 img)FFHQ-5k(5,000 img) Figure 1. Results generated by StyleGAN-V2 [20] and our method (StyleGAN-V2 + AdaptiveMix) on AFHQ-Cat and FFHQ-5k. We propose a simple yet effective module AdaptiveMix, which can be used for helping the training of unsupervised GANs. When trained on a small amount of data, StyleGAN-V2 generates images with artifacts, due to unstable training. However, our AdaptiveMix effectively boosts the performance of StyleGAN-V2 in terms of image quality. [18,45,55,57]. A standard GAN consists of a generator and a discriminator network, where the discriminator is trained to discriminate real/generated samples, and the generator aims to generate samples that can fool the discriminator. Nevertheless, the training of GANs is difficult and unstable, leading to low-quality generated samples [23, 34]. Many efforts have been devoted to improving the train- ing of GANs ( e.g. [1,5,12,25,27,34,39]). Previous studies [37] attempted to co-design the network architecture of the generator and discriminator to balance the iterative training. Following this research line, PG-GAN [16] gradually trains 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. 16219 the GANs with progressive growing architecture according to the paradigm of curriculum learning. More recently, data augmentation-based methods, such as APA [15], ADA [17], and adding noises into the generator [20], were further pro- posed to stabilize the training of GANs. A few works ad- dress this problem on the discriminator side. For example, WGAN [1] proposes to enforce a Lipschitz constraint by us- ing weight clipping. Instead, WGAN-GP [6] directly penal- izes the norm of the discriminator’s gradient. These meth- ods have shown that revisions of discriminators can achieve promising performance. However, improving the training of GANs remains an unsolved and challenging problem. In this paper, considering that the discriminator is critical to the training of GANs, we address the problem of training GANs from a novel perspective, i.e.,robust image classifi- cation. In particular, the discriminator can be regarded as performing a classification task that discriminates real/fake samples. Our insight is that controlling the image repre- sentation ( i.e., feature extractor) of the discriminator can improve the training of GANs, motivated by studies on ro- bust image classification [28, 44]. More specifically, recent work [44] on robust image representation presents inspir- ing observations that training data is scattered in the learn- ing space of vanilla classification networks; hence, the net- works would improperly assign high confidences to samples that are off the underlying manifold of training data. This phenomenon also leads to the vulnerability of GANs, i.e., the discriminator cannot focus on learning the distribution of real data. Therefore, we propose to shrink the regions of training data in the image representation space of the dis- criminator. Different from existing works [15, 17], we explore a question for GANs: Would the training stability of GANs be improved if we explicitly shrink the regions of training data in the image representation space supported by the discrim- inator? To this end, we propose a module named Adap- tiveMix to shrink the regions of training data in the latent space constructed by a feature extractor. However, it is non- trivial and challenging to directly capture the boundaries of feature space. Instead, our insight is that we can shrink the feature space by reducing the distance between hard and easy samples in the latent space, where hard samples are regarded as the samples that are difficult for classification networks to discriminate/classify. To this end, AdaptiveMix constructs hard samples by mixing a pair of training images and then narrows down the distance between mixed images and easy training samples represented by the feature extrac- tor for feature space shrinking. We evaluate the effective- ness of our AdaptivelyMix with state-of-the-art GAN archi- tectures, including DCGAN [37] and StyleGAN-V2 [20], which demonstrates that the proposed AdaptivelyMix facil- itates the training of GANs and effectively improves the im- age quality of generated samples. Besides image generation, our AdaptiveMix can be ap- plied to image classification [9,53] and Out-Of-Distribution(OOD) detection [11, 14, 50] tasks, by equipping it with suitable classifiers. To show the way of applying Adap- tiveMix, we integrate it with the Orthogonal classifier in recent start-of-the-art work [52] in OOD. Extensive experi- mental results show that our AdaptiveMix is simple yet ef- fective, which consistently boosts the performance of [52] on both robust image classification and Out-Of-Distribution tasks on multiple datasets. In a nutshell, the contribution of this paper can be sum- marized as: • We propose a novel module, namely AdaptiveMix, to improve the training of GANs. Our AdaptiveMix is simple yet effective and plug-and-play, which is help- ful for GANs to generate high-quality images. • We show that GANs can be stably and efficiently trained by shrinking regions of training data in image representation supported by the discriminator. • We show our AdaptiveMix can be applied to not only image generation, but also OOD and robust image classification tasks. Extensive experiments show that our AdaptiveMix consistently boosts the performance of baselines for four different tasks ( e.g., OOD) on seven widely-used datasets.
Liu_StyleRF_Zero-Shot_3D_Style_Transfer_of_Neural_Radiance_Fields_CVPR_2023	Abstract 3D style transfer aims to render stylized novel views of a 3D scene with multi-view consistency. However, most ex- isting work suffers from a three-way dilemma over accu- rate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. We propose StyleRF (Style Radiance Fields), an innovative 3D style transfer technique that resolves the three-way dilemma by performing style transformation within the feature space of a radiance field. StyleRF employs an explicit grid of high-level features to represent 3D scenes, with which high- fidelity geometry can be reliably restored via volume ren- dering. In addition, it transforms the grid features ac- cording to the reference style which directly leads to high- quality zero-shot style transfer. StyleRF consists of two innovative designs. The first is sampling-invariant con- tent transformation that makes the transformation invari- *Shijian Lu is the corresponding author.ant to the holistic statistics of the sampled 3D points and accordingly ensures multi-view consistency. The second is deferred style transformation of 2D feature maps which is equivalent to the transformation of 3D points but greatly re- duces memory footprint without degrading multi-view con- sistency. Extensive experiments show that StyleRF achieves superior 3D stylization quality with precise geometry recon- struction and it can generalize to various new styles in a zero-shot manner. Project website: https://kunhao- liu.github.io/StyleRF/	1. Introduction Given a set of multi-view images of a 3D scene and an image capturing a target style, 3D style transfer aims to gen- erate novel views of the 3D scene that have the target style consistently across the generated views (Fig. 1). Neural style transfer has been investigated extensively, and state- of-the-art methods allow transferring arbitrary styles in a 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. 8338 zero-shot manner. However, most existing work focuses on style transfer across 2D images [15, 21, 24] but cannot extend to a 3D scene that has arbitrary new views. Prior studies [19, 22, 37, 39] have shown that naively combining 3D novel view synthesis and 2D style transfer often leads to multi-view inconsistency or poor stylization quality, and 3D style transfer should optimize novel view synthesis and style transfer jointly. However, the current 3D style transfer is facing a three- way dilemma over accurate geometry reconstruction, high- quality stylization, and being generalizable to new styles. Different approaches have been investigated to resolve the three-way dilemma. For example, multiple style trans- fer [11, 22] requires a set of pre-defined styles but can- not generalize to unseen new styles. Point-cloud-based style transfer [19, 37] requires a pre-trained depth estima- tion module that is prone to inaccurate geometry reconstruc- tion. Zero-shot style transfer with neural radiance fields (NeRF) [8] cannot capture detailed style patterns and tex- tures as it implicitly injects the style information into neu- ral network parameters. Optimization-based style trans- fer [17, 39, 63] suffers from slow optimization and cannot scale with new styles. In this work, we introduce StyleRF to resolve the three- way dilemma by performing style transformation in the fea- ture space of a radiance field. A radiance field is a contin- uous volume that can restore more precise geometry than point clouds or meshes. In addition, transforming a radi- ance field in the feature space is more expressive with bet- ter stylization quality than implicit methods [8], and it can also generalize to arbitrary styles. We construct a 3D scene representation with a grid of deep features to enable fea- ture transformation. In addition, multi-view consistent style transformation in the feature space could be achieved by either transforming the whole feature grid or transforming the sampled 3D points. We adopt the latter as the former in- curs much more computational cost during training to styl- ize the whole feature grid in every iteration, whereas the latter can reduce computational cost through decreasing the size of training patch and the number of sampled points. However, applying off-the-shelf style transformations to a batch of sampled 3D points impairs the multi-view consis- tency as they are conditioned on the holistic statistics of the batch. Beyond that, transforming every sampled 3D point is memory-intensive since NeRF needs to query hundreds of sampled points along each ray for rendering a single pixel. We decompose the style transformation into sampling- invariant content transformation (SICT) and deferred style transformation (DST), the former eliminating the depen- dency on holistic statistics of sampled point batch and the latter deferring style transformation to 2D feature maps for better efficiency. In SICT, we introduce volume-adaptive normalization that learns the mean and variance of thewhole volume instead of computing them from a sampled batch. In addition, we apply channel-wise self-attention to transform each 3D point independently to make it condi- tioned on the feature of that point regardless of the holis- tic statistics of the sampled batch. In DST, we defer the style transformation to the volume-rendered 2D feature maps based on the observation that the style transforma- tion of each point is the same. By formulating the style transformation by pure matrix multiplication and adaptive bias addition, transforming 2D feature maps is mathemat- ically equivalent to transforming 3D point features but it saves computation and memory greatly. Thanks to the memory-efficient representation of 3D scenes and deferred style transformation, our network can train with 256×256 patches directly without requiring sub-sampling like previ- ous NeRF-based 3D style transfer methods [8, 11, 22]. The contributions of this work can be summarized in three aspects. First , we introduce StyleRF, an innovative zero-shot 3D style transfer framework that can generate zero-shot high-quality 3D stylization via style transforma- tion within the feature space of a radiance field. Second , we design sampling-invariant content transformation and deferred style transformation, the former achieving multi- view consistent transformation by eliminating dependency on holistic statistics of sampled point batch while the lat- ter greatly improves stylization efficiency by deferring style transformation to 2D feature maps. Third , extensive experi- ments show that StyleRF achieves superior 3D style transfer with accurate geometry reconstruction, high-quality styliza- tion, and great generalization to new styles.
Liu_Fine-Grained_Face_Swapping_via_Regional_GAN_Inversion_CVPR_2023	Abstract We present a novel paradigm for high-ﬁdelity face swap- ping that faithfully preserves the desired subtle geometry and texture details. We rethink face swapping from the per- spective of ﬁne-grained face editing, i.e., “editing for swap- ping” (E4S), and propose a framework that is based on the explicit disentanglement of the shape and texture of facial components. Following the E4S principle, our framework enables both global and local swapping of facial features, as well as controlling the amount of partial swapping spec- iﬁed by the user. Furthermore, the E4S paradigm is in- herently capable of handling facial occlusions by means of facial masks. At the core of our system lies a novel Re- gional GAN Inversion (RGI) method, which allows the ex- plicit disentanglement of shape and texture. It also allows face swapping to be performed in the latent space of Style- GAN. Speciﬁcally, we design a multi-scale mask-guided en- coder to project the texture of each facial component into regional style codes. We also design a mask-guided injec- tion module to manipulate the feature maps with the style Work done when Zhian Liu was an intern at Tencent AI Lab †Equal contribution. ⇤Corresponding author: [email protected]. Based on the disentanglement, face swapping is re- formulated as a simpliﬁed problem of style and mask swap- ping. Extensive experiments and comparisons with current state-of-the-art methods demonstrate the superiority of our approach in preserving texture and shape details, as well as working with high resolution images. The project page is https://e4s2022.github.io	1. Introduction Face swapping aims at transferring the identity infor- mation ( e.g., shape and texture of facial components) of a source face to a given target face, while retaining the identity-irrelevant attribute information of the target ( e.g., expression, head pose, background, etc.). It has immense potential applications in the entertainment and ﬁlm produc- tion industry, and thus has drawn considerable attention in the ﬁeld of computer vision and graphics. The ﬁrst and foremost challenge in face swapping is identity preservation ,i.e., how to faithfully preserve the unique facial characteristics of the source image. Most existing methods [ 9,24,38] rely on a pre-trained 2D face recognition network [ 12] or a 3D morphable face model (3DMM) [ 7,13] to extract the global identity-related 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. 8578 tures, which are then injected into the face generation pro- cess. However, these face models are mainly designed for classiﬁcation rather than generation, thus some informative and important visual details related to facial identity may not be captured. Furthermore, the 3D face model built from a single input image can hardly meet the requirement of ro- bust and accurate facial shape recovery. Consequently, re- sults from previous methods often exhibit the “in-between effect”: i.e., the swapped face resembles both the source and the target faces, which looks like a third person instead of faithfully preserving the source identity. A related prob- lem is skin color , where we argue that skin color is some- times an important aspect of the source identity and should be preserved, while previous methods will always maintain the skin color of the target face, resulting in unrealistic re- sults when swapping faces with distinct skin tones. Another challenge is how to properly handle facial oc- clusion. In real applications, for example, it is a common situation that some face regions are occluded by hair in the input images. An ideal swapped result should maintain the hair from the target, meaning that the occluded part should be recovered in the source image. To handle occlusion, FS- GAN [ 25] designs an inpainting sub-network to estimate the missing pixels of the source, but their inpainted faces are blurry. A reﬁnement network is carefully designed in FaceShifter [ 24] to maintain the occluded region in the tar- get; however, the reﬁnement network may bring back some identity information of the target. To address the above challenges more effectively, we re- think face swapping from a new perspective of ﬁne-grained face editing, i.e.,“editing for swapping” (E4S) . Given that both the shape and texture of individual facial components are correlated with facial identity, we consider to disentan- gle shape and texture explicitly for better identity preserva- tion. Instead of using a face recognition model or 3DMMs to extract global identity features, inspired by ﬁne-grained face editing [ 23], we exploit component masks for local fea- ture extraction. With such disentanglement, face swapping can be transformed as the replacement of local shape and texture between two given faces. The locally-recomposed shapes and textures are then fed into a mask-guided gener- ator to synthesize the ﬁnal result. One additional advantage of our E4Sframework is that the occlusion challenge can be naturally handled by the masks, as the current face parsing network [ 44] can provide the pixel-wise label of each face region. The generator can ﬁll out the missing pixels with the swapped texture features adaptively according to those labels. It requires no additional effort to design a dedicated module as in previous methods [ 24,25]. The key to our E4S is the disentanglement of shape and texture of facial components. Recently, StyleGAN [ 18] has been applied to various image synthesis tasks due to its amazing performance on high-quality image generation,which inspires us to exploit a pre-trained StyleGAN for the disentanglement. This is an ambitious goal because current GAN inversion methods [ 30,33,36] only focus on global attribute editing (age, gender, expression, etc.) in the global style space of StyleGAN, and provide no mechanism for local shape and texture editing. To solve this, we propose a novel Regional GAN Inver- sion (RGI) method that resides in a new regional-wise W+ space, denoted as Wr+. Speciﬁcally, we design a mask- guided multi-scale encoder to project an input face into the style space of StyleGAN. Each facial component has a set of style codes for different layers of the StyleGAN gener- ator. We also design a mask-guided injection module that uses the style codes to manipulate the feature maps in the generator according to the given masks. In this way, the shape and texture of each facial component are fully disen- tangled, where the texture is represented by the style codes while the shape is by the masks. Moreover, this new in- version latent space supports the editing of each individual face component in shape and texture, enabling various ap- plications such as face beautiﬁcation, hairstyle transfer, and controlling the swapping extent of face swapping. To sum up, our contributions are: •We tackle face swapping from a new perspective of ﬁne-grained editing, i.e.,editing for swapping , and propose a novel framework for high-ﬁdelity face swap- ping with identity preservation and occlusion handling. •We propose a StyleGAN-based Regional GAN Inver- sion (RGI) method that resides in a novel Wr+space, for the explicit disentanglement of shape and texture. It simpliﬁes face swapping as the swapping of the cor- responding style codes and masks. •The extensive experiments on face swapping, face edit- ing, and other extension tasks demonstrate the effec- tiveness of our E4S framework and RGI.
Khurana_Point_Cloud_Forecasting_as_a_Proxy_for_4D_Occupancy_Forecasting_CVPR_2023	Abstract Predicting how the world can evolve in the future is cru- cial for motion planning in autonomous systems. Classi- cal methods are limited because they rely on costly human annotations in the form of semantic class labels, bounding boxes, and tracks or HD maps of cities to plan their mo- tion — and thus are difﬁcult to scale to large unlabeled datasets. One promising self-supervised task is 3D point cloud forecasting [11, 18–20] from unannotated LiDAR se- quences. We show that this task requires algorithms to im- plicitly capture (1) sensor extrinsics (i.e., the egomotion of the autonomous vehicle), (2) sensor intrinsics (i.e., the sampling pattern speciﬁc to the particular LiDAR sensor), and (3) the shape and motion of other objects in the scene. But autonomous systems should make predictions about the world and not their sensors! To this end, we factor out (1) and (2) by recasting the task as one of spacetime (4D) oc- Equal contributioncupancy forecasting. But because it is expensive to obtain ground-truth 4D occupancy, we “render” point cloud data from 4D occupancy predictions given sensor extrinsics and intrinsics, allowing one to train and test occupancy algo- rithms with unannotated LiDAR sequences. This also al- lows one to evaluate and compare point cloud forecasting algorithms across diverse datasets, sensors, and vehicles.	1. Introduction Motion planning in a dynamic environment requires au- tonomous agents to predict the motion of other objects. Standard solutions consist of perceptual modules such as mapping, object detection, tracking, and trajectory fore- casting. Such solutions often rely on human annotations in the form of HD maps of cities, or semantic class labels, bounding boxes, and object tracks, and therefore are difﬁ- cult to scale to large unlabeled datasets. One promising self- supervised task is 3D point cloud forecasting [11, 18–20]. 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. 1116 Figure 2. Points depend on the intersection of rays from the depth sensor and the environment. Therefore, accurately predict- ing points requires accurately predicting sensor extrinsics (sensor egomotion) and intrinsics (ray sampling pattern). But we want to understand dynamics of the environment, not our LiDAR sensor! Since points appear where lasers from the sensor and scene intersect, the task of forecasting point clouds requires algo- rithms to implicitly capture (1) sensor extrinsics ( i.e., the ego-motion of the autonomous vehicle), (2) sensor intrin- sics ( i.e., the sampling pattern speciﬁc to the LiDAR sen- sor), and (3) the shape and motion of other objects in the scene. This task can be non-trivial even in a static scene (Fig. 2). We argue that autonomous systems should focus on making predictions about the world and not themselves, since an ego-vehicle has access to its future motion plans (extrinsics) and calibrated sensor parameters (intrinsics). We factor out these (1) sensor extrinsics and (2) intrin- sics by recasting the task of point cloud forecasting as one of spacetime (4D) occupancy forecasting. This disentan- gles and simpliﬁes the formulation of point cloud forecast- ing, which now focuses solely on forecasting the central quantity of interest, the 4D occupancy. Because it is ex- pensive to obtain ground-truth 4D occupancy, we “render” point cloud data from 4D occupancy predictions given sen- sor extrinsics and intrinsics. In some ways, our approach can be seen as the spacetime analog of novel-view syn- thesis from volumetric models such as NeRFs [12]; rather than rendering images by querying a volumetric model with rays from a known camera view, we render a LiDAR scan by querying a 4D model with rays from known sensor in- trinsics and extrinsics. This allows one to train and test 4D occupancy forecasting algorithms with un-annotated Li- DAR sequences. This also allows one to evaluate and compare point cloud forecasting algorithms across diverse datasets, sensors, and vehicles. We ﬁnd that our approach to 4D occupancy forecasting, which can also render point clouds, performs drastically better than SOTAs in point cloud forecasting, both quantitatively (by up to 3.26m L1 error, Tab. 1) and qualitatively (Fig. 6). Our method beats prior art with zero-shot cross-sensor generalization (Tab. 2). To our knowledge, these are ﬁrst results that generalize across train/test sensor rigs, illustrating the power of dis- entangling sensor motion from scene motion.
Liu_FlatFormer_Flattened_Window_Attention_for_Efficient_Point_Cloud_Transformer_CVPR_2023	Abstract Transformer, as an alternative to CNN, has been proven effective in many modalities ( e.g., texts and images). For 3D point cloud transformers, existing efforts focus primarily on pushing their accuracy to the state-of-the-art level. However, their latency lags behind sparse convolution-based models (3×slower ), hindering their usage in resource-constrained, latency-sensitive applications (such as autonomous driving). This inefficiency comes from point clouds’ sparse and irreg- ular nature, whereas transformers are designed for dense, regular workloads. This paper presents FlatFormer to close this latency gap by trading spatial proximity for better com- putational regularity. We first flatten the point cloud with window-based sorting and partition points into groups of equal sizes rather than windows of equal shapes . This effec- tively avoids expensive structuring and padding overheads. We then apply self-attention within groups to extract local features, alternate sorting axis to gather features from differ- ent directions, and shift windows to exchange features across groups. FlatFormer delivers state-of-the-art accuracy on Waymo Open Dataset with 4.6×speedup over (transformer- based) SST and 1.4×speedup over (sparse convolutional) CenterPoint. This is the first point cloud transformer that achieves real-time performance on edge GPUs and is faster than sparse convolutional methods while achieving on-par or even superior accuracy on large-scale benchmarks.	1. Introduction Transformer [75] has become the model of choice in nat- ural language processing (NLP), serving as the backbone of many successful large language models (LLMs) [2, 17]. Recently, its impact has further been expanded to the vision community, where vision transformers (ViTs) [18, 45, 74] have demonstrated on-par or even superior performance com- pared with CNNs in many visual modalities ( e.g., image and video). 3D point cloud, however, is not yet one of them. ∗indicates equal contributions. Mean mAPH L2646668707274 Latency (ms)01326395265 FlatFormer (Ours) CenterPointSST (Center)SSTVoxSetCenterPoint++4x faster3x fasterTransformerConvolutionPoint Cloud TransformersSparse Convolutional ModelsFigure 1. Previous point cloud transformers ( ⋆) are 3-4×slower than sparse convolution-based models ( •) despite achieving similar detection accuracy. FlatFormer is the first point cloud transformer that is faster than sparse convolutional methods with on-par accu- racy. Latency is measured on an NVIDIA Quadro RTX A6000. Different from images and videos, 3D point clouds are intrinsically sparse and irregular. Most existing point cloud models [94] are based on 3D sparse convolution [24], which computes convolution only on non-zero features. They re- quire dedicated system support [14, 71, 91] to realize high utilization on parallel hardware ( e.g., GPUs). Many efforts have been made toward point cloud trans- formers (PCTs) to explore their potential as an alternative to sparse convolution. Global PCTs [26] benefit from the reg- ular computation pattern of self-attention but suffer greatly from the quadratic computational cost (w.r.t. the number of points). Local PCTs [50, 98] apply self-attention to a local neighborhood defined in a similar way to point-based mod- els [57] and are thus bottlenecked by the expensive neighbor gathering [47]. These methods are only applicable to single objects or partial indoor scans (with <4k points) and cannot be efficiently scaled to outdoor scenes (with >30k points). Inspired by Swin Transformer [45], window PCTs [19,70] compute self-attention at the window level, achieving 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. 1200 pressive accuracy on large-scale 3D detection benchmarks. Despite being spatially regular, these windows could have drastically different numbers of points (which differ by more than 80×) due to the sparsity. This severe imbalance results in redundant computation with inefficient padding and par- titioning overheads. As a result, window PCTs can be 3× slower than sparse convolutional models (Figure 1), limiting their applications in resource-constrained, latency-sensitive scenarios ( e.g., autonomous driving, augmented reality). This paper introduces FlatFormer to close this huge la- tency gap. Building upon window PCTs, FlatFormer trades spatial proximity for better computational regularity by par- titioning 3D point cloud into groups of equal sizes instead ofwindows of equal shapes . It applies self-attention within groups to extract local features, alternates the sorting axis to aggregate features from different orientations, and shifts windows to exchange features across groups. Benefit from the regular computation pattern, FlatFormer achieves 4.6× speedup over (transformer-based) SST and 1.4×speedup over (sparse convolutional) CenterPoint while delivering the state-of-the-art accuracy on Waymo Open Dataset. To the best of our knowledge, FlatFormer is the first point cloud transformer that achieves on-par or superior accuracy than sparse convolutional methods with lower latency. It is also the first to achieve real-time performance on edge GPUs. With better hardware support for transformers ( e.g., NVIDIA Hopper), point cloud transformers will have a huge potential to be the model of choice in 3D deep learning. We believe our work will inspire future research in this direction.
Kobayashi_Two-Way_Multi-Label_Loss_CVPR_2023	Abstract A natural image frequently contains multiple classiﬁ- cation targets, accordingly providing multiple class labels rather than a single label per image. While the single-label classiﬁcation is effectively addressed by applying a softmax cross-entropy loss, the multi-label task is tackled mainly in a binary cross-entropy (BCE) framework. In contrast to the softmax loss, the BCE loss involves issues regarding imbal- ance as multiple classes are decomposed into a bunch of binary classiﬁcations; recent works improve the BCE loss to cope with the issue by means of weighting. In this pa- per, we propose a multi-label loss by bridging a gap be- tween the softmax loss and the multi-label scenario. The proposed loss function is formulated on the basis of relative comparison among classes which also enables us to fur- ther improve discriminative power of features by enhanc- ing classiﬁcation margin. The loss function is so ﬂexible as to be applicable to a multi-label setting in two ways for discriminating classes as well as samples. In the exper- iments on multi-label classiﬁcation, the proposed method exhibits competitive performance to the other multi-label losses, and it also provides transferrable features on single- label ImageNet training. Codes are available at https: //github.com/tk1980/TwowayMultiLabelLoss .	1. Introduction Deep neural networks are successfully applied to super- vised learning [12, 26, 38] through back-propagation based on a loss function exploiting plenty of annotated samples. In the supervised learning, classiﬁcation is one of primary tasks to utilize as annotation a class label to which an image sample belongs. As in ImageNet [7], most of image datasets provide a single class label per image, and a softmax loss is widely employed to deal with the single-label annotation, producing promising performance on various tasks. The single-label setting, however, is a limited scenario from practical viewpoints. An image frequently contains multiple classiﬁcation targets [39], such as objects, requir-ing laborious cropping to construct single-label annotations. There are also targets, such as visual attributes [25], which are hard to be disentangled and thereby incapable of produc- ing single-label instances. Those realistic situations pose so-called multi-label classiﬁcation where an image sample is equipped with multiple labels beyond a single label. While a softmax loss works well in a single-label learn- ing, the multi-label tasks are addressed mainly by applying a binary cross-entropy (BCE) loss. Considering multiple labels are drawn from Cclass categories, the multi-label classiﬁcation can be decomposed into Cbinary classiﬁca- tion tasks, each of which focuses on discriminating samples in a target class category [28]; the BCE loss is well coupled with the decomposition approach. Such a decomposition, however, involves an imbalance issue. Even in a case of balanced class distribution, the number of positive samples is much smaller than that of negatives, as small portion of wholeC-class categories are assigned to each sample as an- notation (positive) labels. The biased distribution is prob- lematic in a naive BCE loss. To cope with the imbalance issue in BCE, a simple weighting approach based on class frequencies [25] is commonly applied and in recent years it is further sophisticated by incorporating adaptive weighting scheme such as in Focal loss [18] and its variant [2]. On the other hand, the softmax loss naturally copes with multi- ple classes without decomposition nor bringing the above- mentioned imbalance issue; it actually works well in the balanced (single-label) class distribution. The softmax loss is intrinsically based on relative comparison among classes (3) which is missed in the BCE-based losses, though being less applicable to multi-label classiﬁcation. In this paper, we propose a multi-label loss to effec- tively deal with multiple labels in a manner similar to the softmax loss. Through analyzing the intrinsic loss func- tion of the softmax loss, we formulate an efﬁcient multi- label loss function to exploit relative comparison between positive and negative classes. The relative comparison is related to classiﬁcation margin between positive and neg- ative classes, and we propose an approach to enlarge the margin by s
Liao_A_Light_Weight_Model_for_Active_Speaker_Detection_CVPR_2023	Abstract Active speaker detection is a challenging task in audio- visual scenarios, with the aim to detect who is speaking in one or more speaker scenarios. This task has received con- siderable attention because it is crucial in many applica- tions. Existing studies have attempted to improve the per- formance by inputting multiple candidate information and designing complex models. Although these methods have achieved excellent performance, their high memory and computational power consumption render their application to resource-limited scenarios difficult. Therefore, in this study, a lightweight active speaker detection architecture is constructed by reducing the number of input candidates, splitting 2D and 3D convolutions for audio-visual feature extraction, and applying gated recurrent units with low computational complexity for cross-modal modeling. Ex- perimental results on the AVA-ActiveSpeaker dataset reveal that the proposed framework achieves competitive mAP per- formance (94.1% vs. 94.2%), while the resource costs are significantly lower than the state-of-the-art method, partic- ularly in model parameters (1.0M vs. 22.5M, approximately 23×) and FLOPs (0.6G vs. 2.6G, approximately 4×). Ad- ditionally, the proposed framework also performs well on the Columbia dataset, thus demonstrating good robustness. The code and model weights are available at https: //github.com/Junhua-Liao/Light-ASD .	1. Introduction Active speaker detection is a multi-modal task that aims to identify active speakers from a set of candidates in ar- bitrary videos. This task is crucial in speaker diariza- tion [7, 42], speaker tracking [28, 29], automatic video edit- ing [10, 20], and other applications, and thus has attracted considerable attention from both the industry and academia. *Corresponding author Figure 1. mAP vs. FLOPs, size ∝parameters. This figure shows the mAP of different methods [1,2,19,23,37,45] on the benchmark and the FLOPs required to predict one frame containing three can- didates. The size of the blobs is proportional to the number of model parameters. The legend shows the size of blobs correspond- ing to the model parameters from 1×106to30×106. Research on active speaker detection dates back more than two decades [8, 35]. However, the lack of reliable large-scale data has delayed the development of this field. With the release of the first large-scale active speaker detec- tion dataset, A V A-ActiveSpeaker [33], significant progress has been made in this field [15, 37, 38, 40, 47], following the rapid development of deep learning for audio-visual tasks [22]. These studies improved the performance of ac- tive speaker detection by inputting face sequences of mul- tiple candidates simultaneously [1, 2, 47], extracting visual features with 3D convolutional neural networks [3, 19, 48], and modeling cross-modal information with complex atten- tion modules [9,44,45], etc, which resulted in higher mem- ory and computation requirements. Therefore, applying the existing methods to scenarios requiring real-time process- ing with limited memory and computational resources, such as automatic video editing and live television, is difficult. 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. 22932 This study proposes a lightweight end-to-end architec- ture designed to detect active speakers in real time, where improvements are made from the three aspects of: (a) Sin- gle input: inputting a single candidate face sequence with the corresponding audio; (b) Feature extraction: split- ting the 3D convolution of visual feature extraction into 2D and 1D convolutions to extract spatial and temporal infor- mation, respectively, and splitting the 2D convolution for audio feature extraction into two 1D convolutions to ex- tract the frequency and temporal information; (c) Cross- modal modeling: using gated recurrent unit (GRU) [6] with less calculation, instead of complex attention mod- ules, for cross-modal modeling. Based on the character- istics of the lightweight architecture, a novel loss function is designed for training. Figure 1 visualizes multiple met- rics of different active speaker detection approaches. The experimental results reveal that the proposed active speaker detection method (1.0M params, 0.6G FLOPs, 94.1% mAP) significantly reduces the model size and computational cost, and its performance is still comparable to that of the state-of-the-art method [23] (22.5M params, 2.6G FLOPs, 94.2% mAP) on the benchmark. Moreover, the proposed method demonstrates good robustness in cross-dataset test- ing. Finally, the single-frame inference time of the proposed method ranges from 0.1ms to 4.5ms, which is feasible for deployment in real-time applications. The major contributions can be summarized as follows: • A lightweight design is developed from the three as- pects of information input, feature extraction, and cross-modal modeling; subsequently, a lightweight and effective end-to-end active speaker detection framework is proposed. In addition, a novel loss func- tion is designed for training. • Experiments on A V A-ActiveSpeaker [33], a bench- mark dataset for active speaker detection released by Google, reveal that the proposed method is comparable to the state-of-the-art method [23], while still reducing model parameters by 95.6% and FLOPs by 76.9%. • Ablation studies, cross-dataset testing, and qualitative analysis demonstrate the state-of-the-art performance and good robustness of the proposed method.
Li_Adversarially_Masking_Synthetic_To_Mimic_Real_Adaptive_Noise_Injection_for_CVPR_2023	Abstract This paper considers the synthetic-to-real adaptation of point cloud semantic segmentation, which aims to segment the real-world point clouds with only synthetic labels avail- able. Contrary to synthetic data which is integral and clean, point clouds collected by real-world sensors typi- cally contain unexpected and irregular noise because the sensors may be impacted by various environmental condi- tions. Consequently, the model trained on ideal synthetic data may fail to achieve satisfactory segmentation results on real data. Influenced by such noise, previous adversar- ial training methods, which are conventional for 2D adap- tation tasks, become less effective. In this paper, we aim to mitigate the domain gap caused by target noise via learn- ing to mask the source points during the adaptation pro- cedure. To this end, we design a novel learnable masking module, which takes source features and 3D coordinates as inputs. We incorporate Gumbel-Softmax operation into the masking module so that it can generate binary masks and be trained end-to-end via gradient back-propagation. With the help of adversarial training, the masking module can learn to generate source masks to mimic the pattern of irregular target noise, thereby narrowing the domain gap. We name our method “Adversarial Masking” as adversarial training and learnable masking module depend on each other and cooperate with each other to mitigate the domain gap. Ex- periments on two synthetic-to-real adaptation benchmarks verify the effectiveness of the proposed method.	1. Introduction Recently, point cloud semantic segmentation task at- tracts increasing attention because of its important role in various real-world applications, e.g., autonomous driving, augmented reality, etc. Despite remarkable progress [5, *Work done during an internship at Baidu. †Corresponding author: Guoliang Kang LiDARscanfromSynLiDAR LiDARscanformSemKITTI LiDARscanfromasyntheticdataset(SynLiDAR) LiDARscancollectedfromrealworld(SemKITTI)Figure 1. Comparison between a synthetic LiDAR scan (upper) and a real scan (lower). Both original point clouds and projected LiDAR images are given. Black points denote noise and other col- ors denote points from different classes. Compared with synthetic data which is integral and clean, point clouds collected from the real world typically contain unexpected and irregular noise which may impede the adaptation. 19, 30, 33, 39, 40, 62, 63], most algorithms are designed for the fully-supervised setting, where massive annotated data is available. In the real world, it is costly and time- consuming to annotate large amounts of data, especially for labeling each point in the segmentation task. Syn- 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. 20464 thetic data is easy to obtain and its label can be automat- ically generated, which largely reduces the human effort of annotating data. However, it is usually infeasible to di- rectly apply networks trained on synthetic data to real-world data due to the apparent domain gap between them. In this paper, we consider the synthetic-to-real domain adap- tation [7, 10, 29, 38, 42, 45, 56, 68] for point cloud segmen- tation. Specifically, we aim to utilize the fully-annotated synthetic point clouds (source domain) and unlabeled point clouds collected from imperfect real-world sensors (target domain) to train a network to support the segmentation of real-world point clouds (target domain). Domain adaptation solutions [8, 9, 20, 21] aim to dis- cover and mitigate the domain shift from source to target domain. Through comparing the synthetic and real-world point clouds, we observe that the domain shift can be largely attributed to the unexpected and irregular noise existing in the target domain data. As with [53], we consider “noise” to be the missing points of certain instances/objects, where all pixel channels are zero. Such noise may be caused by various factors such as non-reflective surfaces ( e.g., glass). As shown in Fig. 1, the synthetic point cloud is integral and clean, but the real one contains large amounts of noisy points. A model trained on clean source data may find it hard to understand the scene context under the distraction of noises and thus cannot achieve satisfactory segmentation results on target point clouds. Previous domain adaptation methods [4, 13, 14, 18, 28, 31, 32, 38, 61] ( e.g., adversarial training), which have been proven effective in the 2D visual tasks, can be applied to this 3D segmentation setting. For example, Squeeze- SegV2 [54] employs geodesic correlation alignment [37] to align the point-wise feature distributions of two domains. However, without explicitly modeling and dealing with the noise, these methods bring quite weak benefits to the adap- tation performance. Recently, several works attempt to deal with the target noise to mitigate the domain gap. Rochan et al. [43] randomly select target noise masks and apply the selected mask to source samples. Wu et al. [53] compute one dataset-level mask and apply it to all source samples. Zhao et al. [67] use CycleGAN [69] to perform noise in- painting which is then used to learn synthetic noise gen- eration module. The issues of these previous works are two-fold: 1) they cannot adaptively determine the injected noises according to the context of source samples; 2) the generated mask cannot be guaranteed to reduce the domain shift. Thus, these methods may achieve sub-optimal results. In this paper, we aim to mitigate the domain shift caused by the target noise by learning to adaptively mask the source points during the adaptation procedure. To reach this goal, we need to deal with two problems: 1) how to learn a spa- tial mask that can be adaptively determined according to the specific context of a source sample, and 2) how to guaran-tee the learned masks help narrow the domain gap. To solve the first problem, we design a learnable masking module named “Adaptive Spatial Masking (ASM)” module, which takes source Cartesian coordinates and features as input, to generate point-wise source masks. We incorporate Gumbel- Softmax operation into the masking module so that it can generate binary masks and be trained end-to-end via gra- dient back-propagation. To solve the second problem, we incorporate adversarial training into the masking module learning process. Specifically, during training, we add an additional domain discriminator on top of the feature ex- tractor. By encouraging features from two domains (fea- tures of masked source samples and those of normal tar- get samples) to be indistinguishable, the masking module is able to learn to generate masks mimicking the pattern of tar- get noise and narrow the domain gap. Note that these two designs cooperate with each other to better align features across domains and improve the adaptation performance. In a nutshell, our contributions can be summarized as: • We notice that the pattern of target noise is unexpected and irregular. Thus, we propose to model the target noise in a learnable way. Previous works, which don’t explicitly model the target noise or ignore such char- acteristics, are less effective. • We propose to adversarially mask source samples to mimic the target noise patterns. In detail, we design a novel learnable masking module and incorporate ad- versarial training. Both components cooperate with each other to promote the adaptation. • Experiments on two synthetic-to-real adaptation benchmarks, i.e. SynLiDAR →SemKITTI and Syn- LiDAR →nuScenes, demonstrate that our method can effectively improve the adaptation performance.
Liu_Learned_Image_Compression_With_Mixed_Transformer-CNN_Architectures_CVPR_2023	Abstract Learned image compression (LIC) methods have exhib- ited promising progress and superior rate-distortion per- formance compared with classical image compression stan- dards. Most existing LIC methods are Convolutional Neural Networks-based (CNN-based) or Transformer-based, which have different advantages. Exploiting both advantages is a point worth exploring, which has two challenges: 1) how to effectively fuse the two methods? 2) how to achieve higher performance with a suitable complexity? In this pa- per, we propose an efficient parallel Transformer-CNN Mix- ture (TCM) block with a controllable complexity to incor- porate the local modeling ability of CNN and the non-local modeling ability of transformers to improve the overall ar- chitecture of image compression models. Besides, inspired by the recent progress of entropy estimation models and at- tention modules, we propose a channel-wise entropy model with parameter-efficient swin-transformer-based attention (SWAtten) modules by using channel squeezing. Experi- mental results demonstrate our proposed method achieves state-of-the-art rate-distortion performances on three dif- ferent resolution datasets (i.e., Kodak, Tecnick, CLIC Pro- fessional Validation) compared to existing LIC methods. The code is at https://github.com/jmliu206/ LIC_TCM .	1. Introduction Image compression is a crucial topic in the field of im- age processing. With the rapidly increasing image data, lossy image compression plays an important role in storing and transmitting efficiently. In the passing decades, there were many classical standards, including JPEG [31], WebP [11], and VVC [29], which contain three steps: transform, quantization, and entropy coding, have achieved impres- sive Rate-Distortion (RD) performance. On the other hand, different from the classical standards, end-to-end learned *Heming Sun is the corresponding author. VVC (VTM- 12.1) 0.131bpp\|34.69dB\|15.04dB Ground Truth Ours [MSE] 0.127bpp\|35.78dB\|16.25dB Ours [MS- SSIM] 0.114bpp\|30.81dB\|17.07dB WebP 0.180bpp\|30.77dB\|12.47dBFigure 1. Visualization of decompressed images of kodim 23from Kodak dataset based on different methods (The subfigure is titled as “Method \|Bit rate \|PSNR\|MS-SSIM”). image compression (LIC) is optimized as a whole. Some very recent LIC works [5, 13, 32, 34, 36, 37] have outper- formed VVC which is the best classical image and video coding standards at present, on both Peak signal-to-noise ratio (PSNR) and Multi-Scale Structural Similarity (MS- SSIM). This suggests that LIC has great potential for next- generation image compression techniques. Most LIC methods are CNN-based methods [6, 10, 20, 21, 33] using the variational auto-encoder (V AE) which is proposed by Ball ´eet al. [3]. With the development of vision transformers [9,22] recently, some vision transformer-based LIC methods [23, 36, 37] are also investigated. For CNN- based example, Cheng et al. [6] proposed a residual block- based image compression model. For transformer-based ex- ample, Zou et al. [37] tried a swin-transformer-based image compression model. These two kinds of methods have dif- ferent advantages. CNN has the ability of local modeling, while transformers have the ability to model non-local in- formation. It is still worth exploring whether the advantages of these two methods can be effectively combined with a suitable complexity. In our method, we try to efficiently incorporate both advantages of CNN and transformers by proposing an efficient parallel Transformer-CNN Mixture (TCM) block under a controllable complexity to improve 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. 14388 the RD performance of LIC. In addition to the type of the neural network, the design of entropy model is also an important technique in LIC. The most common way is to introduce extra latent variables as hyper-prior to convert the probability model of compact coding-symbols to a joint model [3]. On that basis, many methods spring up. Minnen et al. [26] utilized the masked convolutional layer to capture the context information. Fur- thermore, they [27] proposed a parallel channel-wise auto- regressive entropy model by splitting the latent to 10 slices. The results of encoded slices can assist in the encoding of remaining slices in a pipeline manner. Recently, many different attention modules [6, 21, 37] were designed and proposed to improve image compres- sion. Attention modules can help the learned model pay more attention to complex regions. However, many of them are time-consuming, or can only capture local informa- tion [37]. At the same time, these attention modules are usually placed in both the main and the hyper-prior path of image compression network, which will further intro- duce large complexity because of a large input size for the main path. To overcome that problem, we try to move at- tention modules to the channel-wise entropy model which has1/16input size compared with that of main path to re- duce complexity. Nevertheless, if the above attention mod- ules are directly added to the entropy model, a large num- ber of parameters will be introduced. Therefore, we pro- pose a parameter-efficient swin-transformer-based attention module (SWAtten) with channel squeezing for the channel- wise entropy model. At the same time, to avoid the latency caused by too many slices, we reduce the number of slices from 10 to 5 to achieve the balance between running speed and RD-performance. As Fig. 1 shows, our method can get pleasant results compared with other methods. The contributions of this paper can be summarized as follows: • We propose a LIC framework with parallel transformer-CNN mixture (TCM) blocks that ef- ficiently incorporate the local modeling ability of CNN and the non-local modeling ability of transform- ers, while maintaining controllable complexity. • We design a channel-wise auto-regressive entropy model by proposing a parameter-efficient swin- transformer-based attention (SWAtten) module with channel squeezing. • Extensive experiments demonstrate that our approach achieves state-of-the-art (SOTA) performance on three datasets (i.e., Kodak, Tecnick, and CLIC datasets) with different resolutions. The method outperforms VVC (VTM-12.1) by 12.30%, 13.71%, 11.85% in Bjøntegaard-delta-rate (BD-rate) [4] on Kodak, Tec- nick, and CLIC datasets, respectively.
Liu_COT_Unsupervised_Domain_Adaptation_With_Clustering_and_Optimal_Transport_CVPR_2023	Abstract Unsupervised domain adaptation (UDA) aims to trans- fer the knowledge from a labeled source domain to an unlabeled target domain. Typically, to guarantee desir- able knowledge transfer, aligning the distribution between source and target domain from a global perspective is widely adopted in UDA. Recent researchers further point out the importance of local-level alignment and propose to construct instance-pair alignment by leveraging on Optimal Transport (OT) theory. However, existing OT-based UDA approaches are limited to handling class imbalance chal- lenges and introduce a heavy computation overhead when considering a large-scale training situation. To cope with two aforementioned issues, we propose a Clustering-based Optimal Transport (COT) algorithm, which formulates the alignment procedure as an Optimal Transport problem and constructs a mapping between clustering centers in the source and target domain via an end-to-end manner. With this alignment on clustering centers, our COT eliminates the negative effect caused by class imbalance and reduces the computation cost simultaneously. Empirically, our COT achieves state-of-the-art performance on several authorita- tive benchmark datasets.	1. Introduction Benefiting from the availability of large-scale data, deep learning has achieved tremendous success over the past few years. However, directly applying a well-trained con- volution neural network on a new domain frequently suf- fers from the domain gap/discrepancy challenge, resulting in spurious predictions on the new domain. To remedy this, Unsupervised Domain Adaptation (UDA) has attracted many researchers’ attention, which can transfer the knowl- edge from a labeled domain to an unlabeled domain. A major line of UDA approaches [1,1,28,42,49,53] aim Email: [email protected] *Equal Contribution †Corresponding Authorto learn a global domain shift by aligning the global source and target distribution while ignoring the local-level align- ment between two domains. By leveraging on global do- main adaptation, the global distributions of source and tar- get domain are almost the same, thus losing the fine-grained information for each class (class-structure) on the source and target domain. Recently, to preserve class structure in both domains, several works [6,15,23,30,38,40,44,51,54] adopt optimal transport (OT) to minimize the sample-level transportation cost between source and target domain, achieving a signifi- cant performance on UDA. However, there exist two issues on recent OT-based UDA approaches. (i) When considering a realistic situation, i.e. the class imbalance1phenomenon occurs between the source and target domain, samples be- longing to the same class in the target domain are assigned with false pseudo labels due to the mechanism of optimal transport, which requires each sample in source domain can be mapped to target samples. As a result, existing OT-based UDA methods provide poor pair-wise matching when fac- ing class imbalance challenges. (ii) OT-based UDA meth- ods tend to find a sample-level optimal counterpart, which requires a large amount of computation overhead, especially training on large-scale datasets. To solve two aforementioned issues, we propose a Clustering-based Optimal Transport algorithm, termed COT, to construct a clustering-level instead of sample-level mapping between source and target domain. Clusters in the source domain are obtained from the classifiers supervised by the labeled source domain data. While for the target do- main, COT utilizes a set of learnable clusters to represent the feature distribution of the target domain, which can de- scribe the sub-domain information [50,57]. For instance, in many object recognition tasks [13,20] an object could contain many attributes. Each attribute can be viewed as a sub-domain. To this end, the clusters on the source and target domain can represent the individual sub-domain in- formation, respectively, such that optimal transport between clusters intrinsically provides a local mapping from the sub- domain in the source domain to those in the target domain. Moreover, we provide a theoretical analysis and compre- 1label distribution are different in two domains, Ps(y)̸=Pt(y) 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. 19998 hensive experimental results to guarantee that (i) COT can alleviate the negative effect caused by class imbalance; (ii) Compared to existing OT-based UDA approaches, our COT economizes much computation head. In summary, our main contributions include: •We propose a novel Clustering-based Optimal Trans- port module as well as a specially designed loss de- rived from the discrete type of Kantorovich dual form, which resolves two aforementioned challenges on the existing OT-based UDA algorithms, facilitating the de- velopment of OT-based UDA community. •We provide a theoretical analysis to guarantee the ad- vantages of our COT. •Our COT achieves state-of-the-art performance on sev- eral UDA benchmark datasets.
Liu_What_You_Can_Reconstruct_From_a_Shadow_CVPR_2023	Abstract 3D reconstruction is a fundamental problem in computer vision, and the task is especially challenging when the object to reconstruct is partially or fully occluded. We introduce a method that uses the shadows cast by an unobserved object in order to infer the possible 3D volumes under occlusion. We create a differentiable image formation model that allows us to jointly infer the 3D shape of an object, its pose, and the position of a light source. Since the approach is end-to- end differentiable, we are able to integrate learned priors of object geometry in order to generate realistic 3D shapes of different object categories. Experiments and visualizations show that the method is able to generate multiple possible solutions that are consistent with the observation of the shadow. Our approach works even when the position of the light source and object pose are both unknown. Our approach is also robust to real-world images where ground- truth shadow mask is unknown.	1. Introduction Reconstructing the 3D shape of objects is a fundamental challenge in computer vision, with a number of applications in robotics, graphics, and data science. The task aims toestimate a 3D model from one or more camera views, and re- searchers over the last twenty years have developed excellent methods to reconstruct visible objects [1, 13 –15, 24, 25, 43]. However, objects are often occluded, with the line of sight obstructed either by another object in the scene, or by them- selves (self-occlusion). Reconstruction from a single image is an under-constrained problem, and occlusions further re- duce the number of constraints. To reconstruct occluded objects, we need to rely on additional context. One piece of evidence that people use to uncover occlu- sions is the shadow cast on the floor by the hidden object. For example, figure 1 shows a scene with an object that has become fully occluded. Even though no appearance features are visible, the shadow reveals that another object exists be- hind the chair, and the silhouette constrains the possible 3D shapes of the occluded object. What hidden object caused that shadow? In this paper, we introduce a framework for reconstructing 3D objects from their shadows. We formulate a generative model of objects and their shadows cast by a light source, which we use to jointly infer the 3D shape, its pose, and the location of the light source. Our model is fully differen- tiable, which allows us to use gradient descent to efficiently search for the best shapes that explain the observed shadow. Our approach integrates both learned empirical priors about 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. 17059 the geometry of typical objects and the geometry of cam- eras in order to estimate realistic 3D volumes that are often encountered in the visual world. Since we model the image formation process, we are able to jointly reason over the object geometry and the parame- ters of the light source. When the light source is unknown, we recover multiple different shapes and multiple different positions of the light source that are consistent with each other. When the light source location is known, our approach can make use of that information to further refine its outputs. We validate our approach for a number of different object categories on a new ground truth dataset. The primary contribution of this paper is a method to use the shadows in a scene to infer the 3D structure, and the rest of the paper will analyze this technique in detail. Section 2 provides a brief overview of related work for using shadows. Section 3 formulates a generative model for objects and how they cast shadows, which we are able to invert in order to in- fer shapes from shadows. Section 4 analyzes the capabilities of this approaches with a known and unknown light source. We believe the ability to use shadows to estimate the spatial structure of the scene will have a large impact on computer vision systems’ ability to robustly handle occlusions.
Khattak_MaPLe_Multi-Modal_Prompt_Learning_CVPR_2023	Abstract Pre-trained vision-language (V-L) models such as CLIP have shown excellent generalization ability to downstream tasks. However, they are sensitive to the choice of input text prompts and require careful selection of prompt templates to perform well. Inspired by the Natural Language Processing (NLP) literature, recent CLIP adaptation approaches learn prompts as the textual inputs tone-tune CLIP for down- stream tasks. We note that using prompting to adapt repre- sentations in a single branch of CLIP (language or vision) is sub-optimal since it does not allow theexibility to dynam- ically adjust both representation spaces on a downstream task. In this work, we propose Multi-modal Prompt Learn- ing (MaPLe) forbothvision and language branches to im- prove alignment between the vision and language represen- tations. Our design promotes strong coupling between the vision-language prompts to ensure mutual synergy and dis- courages learning independent uni-modal solutions. Fur- ther, we learn separate prompts across different early stages to progressively model the stage-wise feature relationships to allow rich context learning. We evaluate the effectiveness of our approach onthreerepresentative tasks of generaliza- tion to novel classes, new target datasets and unseen do- main shifts. Compared with the state-of-the-art method Co- CoOp, MaPLe exhibits favorable performance and achieves an absolute gain of 3.45% on novel classes and 2.72% on overall harmonic-mean, averaged over 11 diverse image recognition datasets. Our code and pre-trained models are available at https://github.com/muzairkhattak/multimodal- prompt-learning .	1. Introduction Foundational vision-language (V-L) models such as CLIP (Contrastive Language-Image Pretraining) [ 32] have shown excellent generalization ability to downstream tasks. Such models are trained to align language and vision modali- ties on web-scale datae.g., 400 million text-image pairs in CLIP. These models can reason about open-vocabulary vi- sual concepts, thanks to the rich supervision provided bynatural language. During inference, hand-engineered text prompts are usede.g., ‘a photo of a<category>’ as a query for text encoder. The output text embeddings are matched with the visual embeddings from an image encoder to predict the output class. Designing high quality contex- tual prompts have been proven to enhance the performance of CLIP and other V-L models [ 17,42]. Despite the effectiveness of CLIP towards generalization to new concepts, its massive scale and scarcity of training data (e.g., few-shot setting) makes it infeasible tone-tune the full model for downstream tasks. Suchne-tuning can also forget the useful knowledge acquired in the large-scale pretraining phase and can pose a risk of overtting to the downstream task. To address the above challenges, exist- ing works propose language prompt learning to avoid man- ually adjusting the prompt templates and providing a mech- anism to adapt the model while keeping the original weights frozen [ 14,25,29,48,49]. Inspired from Natural Language Processing (NLP), these approaches only explore prompt learning for the text encoder in CLIP (Fig. 1:a) while adap- tation choices together with an equally important image en- coder of CLIP remains an unexplored topic in the literature. Our motivation derives from the multi-modal nature of CLIP, where a text and image encoder co-exist andboth contribute towards properly aligning the V-L modalities. We argue that any prompting technique should adapt the model completely and therefore, learning prompts only for the text encoder in CLIP is not sufcient to model the adap- tations needed for the image encoder. To this end, we set out to achieve completeness in the prompting approach and pro- poseMulti-modalPromptLearning (MaPLe) to adequately ne-tune the text and image encoder representations such that their optimal alignment can be achieved on the down- stream tasks (Fig. 1:b). Our extensive experiments on three key representative settings including base-to-novel gener- alization, cross-dataset evaluation, and domain generaliza- tion demonstrate the strength of MaPLe. On base-to-novel generalization, our proposed MaPLe outperforms existing prompt learning approaches across 11 diverse image recog- nition datasets (Fig. 1:c) and achieves absolute average gain of 3.45% on novel classes and 2.72% on harmonic-mean 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. 19113 Figure 1. Comparison of MaPLe with standard prompt learning methods. (a)Existing methods adopt uni-modal prompting techniques tone-tune CLIP representations as prompts are learned only in a single branch of CLIP (language or vision). (b)MaPLe introduces branch-aware hierarchical prompts that adapt both language and vision branches simultaneously for improved generalization. (c)MaPLe surpasses state-of-the-art methods on 11 diverse image recognition datasets for novel class generalization task. over the state-of-the-art method Co-CoOp [ 48]. Further, MaPLe demonstrates favorable generalization ability and robustness in cross-dataset transfer and domain generaliza- tion settings, leading to consistent improvements compared to existing approaches. Owing to its streamlined architec- tural design, MaPLe exhibits improved efciency during both training and inference without much overhead, as com- pared to Co-CoOp which lacks efciency due to its image instance conditioned design. In summary, the main contri- butions of this work include: • We proposemulti-modalprompt learning in CLIP to favourably align its vision-language representations. To the best of our knowledge, this is therst multi- modal prompting approach forne-tuning CLIP. • To link prompts learned in text and image encoders, we propose acoupling functionto explicitly condition vi- sion prompts on their language counterparts. It acts as a bridge between the two modalities and allows mutual propagation of gradients to promote synergy. • Our multi-modal prompts are learned across multi- ple transformer blocks in both vision and language branches toprogressivelylearn the synergistic be- haviour of both modalities. This deep prompting strat- egy allows modeling the contextual relationships inde- pendently, thus providing moreexibility to align the vision-language representations.
Li_MAGE_MAsked_Generative_Encoder_To_Unify_Representation_Learning_and_Image_CVPR_2023	Abstract Generative modeling and representation learning are two key tasks in computer vision. However, these models are typically trained independently, which ignores the po- tential for each task to help the other, and leads to training and model maintenance overheads. In this work, we pro- pose MAsked Generative Encoder (MAGE), the ﬁrst frame- work to unify SOTA image generation and self-supervised representation learning. Our key insight is that using vari- able masking ratios in masked image modeling pre-training can allow generative training (very high masking ratio) and representation learning (lower masking ratio) under the same training framework. Inspired by previous gen- erative models, MAGE uses semantic tokens learned by a vector-quantized GAN at inputs and outputs, combining this with masking. We can further improve the represen- tation by adding a contrastive loss to the encoder output. We extensively evaluate the generation and representation learning capabilities of MAGE. On ImageNet-1K, a single MAGE ViT-L model obtains 9.10 FID in the task of class- unconditional image generation and 78.9% top-1 accuracy for linear probing, achieving state-of-the-art performance in both image generation and representation learning. Code is available at https://github.com/LTH14/mage .	1. Introduction In recent years, we have seen rapid progress in both gen- erative models and representation learning of visual data. Generative models have demonstrated increasingly spectac- ular performance in generating realistic images [3,7,15,46], while state-of-the-art self-supervised representation learn- ing methods can extract representations at a high seman- tic level to achieve excellent performance on a number of This work was done when Tianhong Li interned at Google Research, and was partially supported by the MIT-IBM Watson Research Collabo- ration grant. Correspondence to: Tianhong Li <[email protected]> , Huiwen Chang <[email protected]> . Figure 1. Linear probing and class unconditional generation per- formance of different methods trained and evaluated on ImageNet- 1K. MAGE achieves SOTA performance in linear probing and es- tablishes a new SOTA in class unconditional generation. downstream tasks such as linear probing and few-shot trans- fer [2, 6, 8, 13, 25, 26]. Currently, these two families of models are typically trained independently. Intuitively, since generation and recognition tasks require both visual and semantic under- standing of data, they should be complementary when com- bined in a single framework. Generation beneﬁts represen- tation by ensuring that both high-level semantics and low- level visual details are captured; conversely, representation beneﬁts generation by providing rich semantic guidance. Researchers in natural language processing have observed this synergy: frameworks such as BERT [14] have both high-quality text generation and feature extraction. Another example is DALLE-2 [43], where latents conditioned on apre-trained CLIP representation are used to create high- quality text-to-image generations. However, in computer vision, there are currently no widely adopted models that unify image generation and representation learning in the same framework. Such uni- 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. 2142 Original MAGE seed=0 iter=6 MAGE seed=1 iter=6 Mask MAE MAGE seed=0 iter=1 MAGE seed= 2iter=6 Figure 2. Reconstruction results using MAE and MAGE with 75% masking ratio. MAE reconstructs blurry images with low quality, while MAGE can reconstruct high-quality images with detail, and further improves quality through iterative decoding (see subsection 3.2 for details). With the same mask, MAGE generates diverse reconstruction results with different random seeds. Note that the mask for MAGE is on semantic tokens whereas that of MAE is on patches in the input image. ﬁcation is non-trivial due to the structural difference be- tween these tasks: in generative modeling, we output high-dimensional data, conditioned on low-dimension in- puts such as class labels, text embeddings, or random noise. In representation learning, we input a high-dimensional im- age and create a low-dimensional compact embedding use- ful for downstream tasks. Recently, a number of papers have shown that represen- tation learning frameworks based on masked image mod- eling (MIM) can obtain high-quality representations [2, 18, 26,31], often with very high masking ratios (e.g. 75%) [26]. Inspired by NLP, these methods mask some patches at the input, and the pre-training task is to reconstruct the origi- nal image by predicting these masked patches. After pre- training, task-speciﬁc heads can be added to the encoder to perform linear probe or ﬁne-tuning. These works inspire us to revisit the uniﬁcation question. Our key insight in this work is that generation is viewed as “reconstructing” images that are 100% masked, while rep- resentation learning is viewed as “encoding” images that are 0%masked. We can therefore enable a uniﬁed architecture by using a variable masking ratio during MIM pre-training. The model is trained to reconstruct over a wide range of masking ratios covering high masking ratios that enable generation capabilities, and lower masking ratios that en- able representation learning. This simple but very effec- tive approach allows a smooth combination of generative training and representation learning in the same framework : same architecture, training scheme, and loss function. However, directly combining existing MIM methodswith a variable masking ratio is insufﬁcient for high qual- ity generation because such methods typically use a simple reconstruction loss on pixels, leading to blurry outputs. For example, as a representative of such methods, the recon- struction quality of MAE [27] is poor: ﬁne details and tex- tures are missing (Figure 2). A similar issue exists in many other MIM methods [11, 36]. This paper focuses on bridging this gap. We propose MAGE, a framework that can both generate realistic im- ages and extract high-quality representations from images. Besides using variable masking ratio during pre-training, unlike previous MIM methods whose inputs are pixels, both the inputs and the reconstruction targets of MAGE are semantic tokens . This design improves both generation and representation learning, overcoming the issue described above. For generation, as shown in Figure 2, operating in token space not only allows MAGE to perform image gen- eration tasks iteratively (subsection 3.2), but also enables MAGE to learn a probability distribution of the masked to- kens instead of an average of all possible masked pixels, leading to diverse generation results. For representation learning, using tokens as inputs and outputs allows the net- work to operate at a high semantic level without losing low- level details, leading to signiﬁcantly higher linear probing performances than existing MIM methods. We evaluate MAGE on multiple generative and repre- sentation downstream tasks. As shown in Figure 1, for class- unconditional image generation on ImageNet-1K, our method surpasses state of the art with both ViT-B and ViT- L (ViT-B achieves 11.11 FID [29] and ViT-L achieves 9.10 2143 FID), outperforming the previous state-of-the-art result by a large margin (MaskGIT [7] with 20.68 FID). This signif- icantly push the limit of class-unconditional generation to a level even close to the state-of-the-art of class-conditional image generation ( 6 FID [7, 46]), which is regarded as a much easier task in the literature [38]. For linear probing on ImageNet-1K, our method with ViT-L achieves 78.9% top- 1 accuracy, surpassing all previous MIM-based represen- tation learning methods and many strong contrastive base- lines such as MoCo-v3 [13]. Moreover, when combined with a simple contrastive loss [9], MAGE-C with ViT-L can further get 80.9% accuracy, achieving state-of-the-art performance in self-supervised representation learning. We summarize our contributions: We introduce MAGE, a novel method that uniﬁes genera- tive model and representation learning by a single token- based MIM framework with variable masking ratios, in- troducing new insights to resolve the uniﬁcation problem. MAGE establishes a new state of the art on the task of class-unconditional image generation on ImageNet-1K. MAGE further achieves state of the art in different down- stream tasks, such as linear probing, few-shot learning, transfer learning, and class-conditional image generation.
Lan_Vision_Transformers_Are_Good_Mask_Auto-Labelers_CVPR_2023	Abstract We propose Mask Auto-Labeler (MAL), a high-quality Transformer-based mask auto-labeling framework for in- stance segmentation using only box annotations. MAL takes box-cropped images as inputs and conditionally generates their mask pseudo-labels. We show that Vision Transform- ers are good mask auto-labelers. Our method significantly reduces the gap between auto-labeling and human annota- tion regarding mask quality. Instance segmentation models trained using the MAL-generated masks can nearly match the performance of their fully-supervised counterparts, re- taining up to 97.4% performance of fully supervised mod- els. The best model achieves 44.1% mAP on COCO in- stance segmentation (test-dev 2017), outperforming state- of-the-art box-supervised methods by significant margins. Qualitative results indicate that masks produced by MAL are, in some cases, even better than human annotations.	1. Introduction Computer vision has seen significant progress over the last decade. Tasks such as instance segmentation have made it possible to localize and segment objects with pixel-level accuracy. However, these tasks rely heavily on expan- sive human mask annotations. For instance, when creat-ing the COCO dataset, about 55k worker hours were spent on masks, which takes about 79% of the total annotation time [1]. Moreover, humans also make mistakes. Human annotations are often misaligned with actual object bound- aries. On complicated objects, human annotation quality tends to drop significantly if there is no quality control. Due to the expensive cost and difficulty of quality control, some other large-scale detection datasets such as Open Images [2] and Objects365 [3], only contain partial or even no instance segmentation labels. In light of these limitations, there is an increasing in- terest in pursuing box-supervised instance segmentation, where the goal is to predict object masks from bounding box supervision directly. Recent box-supervised instance segmentation methods [4–8] have shown promising perfor- mance. The emergence of these methods challenges the long-held belief that mask annotations are needed to train instance segmentation models. However, there is still a non- negligible gap between state-of-the-art approaches and their fully-supervised oracles. Our contributions: To address box-supervised instance segmentation, we introduce a two-phase framework consist- ing of a mask auto-labeling phase and an instance segmenta- tion training phase (see Fig. 2). We propose a Transformer- based mask auto-labeling framework, Mask Auto-Labeler (MAL), that takes Region-of-interest (RoI) images as inputs 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. 23745 Box-sup LossCroppedRegionsSupervisedMask LossMask LabelsMALGenerateMasksPhase 1: Mask Auto-labelingInst SegImagePhase 2: Instance Segmentation TrainingTrain Figure 2. An overview of the two-phase framework of box- supervised instance segmentation. For the first phase, we train Mask Auto-Labeler using box supervision and conditionally gen- erate masks of the cropped regions in training images (top). We then train the instance segmentation models using the generated masks (bottom). and conditionally generates high-quality masks (demon- strated in Fig. 1) within the box. Our contributions can be summarized as follows: • Our two-phase framework presents a versatile design compatible with any instance segmentation architecture. Unlike existing methods, our framework is simple and agnostic to instance segmentation module designs. • We show that Vision Transformers (ViTs) used as image encoders yield surprisingly strong auto-labeling results. We also demonstrate that some specific designs in MAL, such as our attention-based decoder, multiple-instance learning with box expansion, and class-agnostic training, crucial for strong auto-labeling performance. Thanks to these components, MAL sometimes even surpasses hu- mans in annotation quality. • Using MAL-generated masks for training, instance seg- mentation models achieve up to 97.4% of their fully supervised performance on COCO and LVIS. Our re- sult significantly narrows down the gap between box- supervised and fully supervised approaches. We also demonstrate the outstanding open-vocabulary general- ization of MAL by labeling novel categories not seen during training. Our method outperforms all the existing state-of-the- art box-supervised instance segmentation methods by large margins. This might be attributed to good representations of ViTs and their emerging properties such as meaningful grouping [9], where we observe that the attention to objects might benefit our task significantly (demonstrated in Fig. 6). We also hypothesize that our class-agnostic training de- sign enables MAL to focus on learning general grouping instead of focusing on category information. Our strong re- sults pave the way to remove the need for expensive human annotation for instance segmentation in real-world settings.
Li_DANI-Net_Uncalibrated_Photometric_Stereo_by_Differentiable_Shadow_Handling_Anisotropic_Reflectance_CVPR_2023	Abstract Uncalibrated photometric stereo (UPS) is challenging due to the inherent ambiguity brought by the unknown light. Although the ambiguity is alleviated on non-Lambertian ob- jects, the problem is still difficult to solve for more general objects with complex shapes introducing irregular shad- ows and general materials with complex reflectance like anisotropic reflectance. To exploit cues from shadow and reflectance to solve UPS and improve performance on gen- eral materials, we propose DANI-Net, an inverse render- ing framework with differentiable shadow handling and anisotropic reflectance modeling. Unlike most previous methods that use non-differentiable shadow maps and as- sume isotropic material, our network benefits from cues of shadow and anisotropic reflectance through two differen- tiable paths. Experiments on multiple real-world datasets demonstrate our superior and robust performance.	1. Introduction Photometric stereo (PS) [48] aims at recovering the sur- face normal from several images captured under varying light conditions with a fixed viewpoint. It has been ap- plied to many fields ( e.g., movies production [6], indus- trial quality inspection [47], and biometrics [53]) due to its advantage in recovering fine-detailed surfaces over other approaches [10, 16] ( e.g., multi-view stereo [38], active sensor-based solutions [61]). Light calibration is crucial to the performance [52]. However, it is also tedious, restricting the applicability of PS. Therefore, uncalibrated photometric stereo (UPS) methods estimating surface normal with un- known lights have been widely studied in the literature. *Corresponding authorUncalibrated photometric stereo suffers from General Bas-Relief (GBR) ambiguity [4] for an integrable, Lam- bertian surface. However, GBR ambiguity is alleviated on a non-Lambertian surface [13]. Therefore, recent ad- vances in UPS ( e.g., [26,56]) adopt the isotropic reflectance model accounting for non-Lambertian effects to solve UPS. Nonetheless, such a model restricts methods’ performance on objects with more general ( e.g., anisotropic) materials, while modeling general reflectance is challenging due to extra unknowns, which eventually make UPS intractable. Other works ( e.g., [25, 56, 57]) notice the benefits of the shadow cues in utilizing global shape-light information to solve PS/UPS because the shadow reflects the interaction of shape and light [24, 59]. However, these methods either fail to exploit the shadow cues due to the lack of a differ- entiable path from the shadow to the concerned unknowns like shape [25], or the shadow cues have limited effects on the visible shape reconstruction due to the implicit shape representation [56, 57]. To this end, this paper proposes the DANI-Net , which solves UPS by Differentiable shadow handling, Anisotropic reflectance modeling, and Neural Inverse Rendering. DANI-Net builds the differentiable path in the sequence of inverse rendering errors, shadow maps, and surface nor- mal maps (or light conditions) (Fig. 1) to fully exploit the shadow cues to solve UPS. Since those cues facilitate solv- ing extra unknowns introduced by a more sophisticated reflectance model, DANI-Net manages to build up such a model (Fig. 1) to improve the performance on general materials. During optimization, DANI-Net propagates in- verse rendering errors via two paths of shadow cues and anisotropic reflectance, respectively, and simultaneously optimizes the shape ( i.e., the depth map and surface nor- mal map), anisotropic reflectance model, shadow map, and light conditions ( i.e., direction and intensity). As a result, 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. 8381 OtherPSMethods DANI-Net Shadow HandlingReflectance Modeling Depth MapInverse Rendering ErrorShadowMapRenderedImageBRDFSphereSurface Normaland LightConditionsNon-differentiable PathRegion of InterestDifferentiable PathReference Point Figure 1. The proposed DANI-Net differs from other PS or UPS methods in two aspects: 1) Shadow Handling. The path in the se- quence of inverse rendering errors, shadow maps, and surface nor- mal maps (or light conditions) of the DANI-Net is differentiable; 2)Reflectance Modeling. DANI-Net adopts an anisotropic re- flectance model. The state-of-the-art UPS method SCPS-NIR [26] is compared in this figure. As can be observed, the proposed DANI-Net produces a smoother and more realistic shadow map of copper B UNNY thanks to the differentiable shadow handling and renders a more realistic copper B ALL image and Bidirectional Re- flectance Distribution Function (BRDF) sphere (of the reference point) due to the anisotropic reflectance modeling. Data of copper BUNNY and B ALL are from D ILIGENT102[36]. DANI-Net achieves state-of-the-art performance on several real-world benchmark datasets. In a nutshell, our contribu- tions are summarized as follows: • We propose a differentiable shadow handling method that facilitates exploiting shadow cues with global shape-light information to solve UPS. Experimental results demonstrate its effectiveness in shadow map re- covery and surface normal estimation. • We introduce an anisotropic reflectance model that de- scribes both isotropic and anisotropic materials to im- prove performance on general materials. Experimental results demonstrate its effectiveness on surface normal estimation for objects with a broad range of isotropic and anisotropic materials. • We propose the DANI-Net that simultaneously opti- mizes shape, anisotropic reflectance, shadow map, and light conditions in an unsupervised manner, propagat- ing inverse rendering errors through two paths involv- ing the shadow cues and anisotropic reflectance, re- spectively. DANI-Net achieves state-of-the-art perfor- mance on several real-world benchmark datasets.
Li_3D-Aware_Face_Swapping_CVPR_2023	Abstract Face swapping is an important research topic in com- puter vision with wide applications in entertainment and privacy protection. Existing methods directly learn to swap 2D facial images, taking no account of the geometric in- formation of human faces. In the presence of large pose variance between the source and the target faces, there always exist undesirable artifacts on the swapped face. In this paper, we present a novel 3D-aware face swap- ping method that generates high-fidelity and multi-view- consistent swapped faces from single-view source and tar- get images. To achieve this, we take advantage of the strong geometry and texture prior of 3D human faces, where the 2D faces are projected into the latent space of a 3D genera- tive model. By disentangling the identity and attribute fea- tures in the latent space, we succeed in swapping faces in a 3D-aware manner, being robust to pose variations while transferring fine-grained facial details. Extensive experi- ments demonstrate the superiority of our 3D-aware face swapping framework in terms of visual quality, identity sim- ilarity, and multi-view consistency. Code is available at https://lyx0208.github.io/3dSwap . ∗Corresponding authors.	1. Introduction Face swapping aims to transfer the identity of a person in the source image to another person in the target image while preserving other attributes like head pose, expression, illumination, background, etc. It has attracted extensive at- tention recently in the academic and industrial world for its potential wide applications in entertainment [14,30,38] and privacy protection [7, 37, 48]. The key of face swapping is to transfer the geometric shape of the facial region ( i.e., eyes, nose, mouth) and detailed texture information (such as the color of eyes) from the source image to the target image while pre- serving both geometry and texture of non-facial regions (i.e., hair, background, etc). Currently, some 3D-based methods consider geometry prior of human faces by fit- ting the input image to 3D face models such as 3D Mor- phable Model (3DMM) [8] to overcome the differences of face orientation and expression between sources and tar- gets [7, 15, 34, 43]. However, these parametric face mod- els only produce coarse frontal faces without fine-grained details, leading to low-resolution and fuzzy swapping re- sults. On the other hand, following Generative Adversarial Network [24], GAN-based [6, 23, 32, 39, 40, 42] or GAN- inversion-based [44, 55, 57, 60] approaches adopt the 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. 12705 versarial training strategy to learn texture information from inputs. Despite the demonstrated photorealistic and high- resolution images, the swapped faces via 2D GANs sustain undesirable artifacts when two input faces undergo large pose variation since the strong 3D geometry prior of human faces is ignored. Moreover, learning to swap faces in 2D images makes little use of the shaped details from sources, leading to poorer performance on identity transferring. Motivated by the recent advances of 3D generative mod- els [12, 13, 20, 25, 45] in synthesizing multi-view consis- tent images and high-quality 3D shapes, it naturally raises a question: can we perform face swapping in a 3D-aware manner to exploit the strong geometry and texture priors? To answer this question, two challenges arise. First , how to infer 3D prior directly from 3D-GAN models still remains open. Current 3D-aware generative models synthesize their results from a random Gaussian noise z, so that their output images are not controllable. This increases the complexity of inferring the required prior from arbitrary input. Second , the inferred prior corresponding to input images is in the form of a high-dimension feature vector in the latent space of 3D GANs. Simply synthesizing multi-view target im- ages referring to the prior and applying 2D face swapping to them produces not only inconsistent artifacts but also a heavy computational load. To address these challenges, we systematically inves- tigate the geometry and texture prior of these 3D gener- ative models and propose a novel 3D-aware face swap- ping framework 3dSwap. We introduce a 3D GAN inver- sion framework to project the 2D inputs into the 3D latent space, motivated by recent GAN inversion approaches [46, 47, 51]. Specifically, we design a learning-based inver- sion algorithm that trains an encoding network to efficiently and robustly project input images into the latent space of EG3D [12]. However, directly borrowing the architecture from 2D approaches is not yet enough since a single-view input provides limited information about the whole human face. To further improve the multi-view consistency of la- tent code projection, we design a pseudo-multi-view train- ing strategy. This design effectively bridges the domain gap between 2D and 3D. To tackle the second problem, we design a face swapping algorithm based on the 3D la- tent codes and directly synthesize the swapped faces with the 3D-aware generator. In this way, we achieve 3D GAN- inversion-based face swapping by a latent code manipulat- ing algorithm consisting of style-mixing and interpolation, where latent code interpolation is responsible for identity transferring while style-mixing helps to preserve attributes. In summary, our contributions are threefold: • To the best of our knowledge, we first address the 3D-aware face swapping task. The proposed 3dSwap method sets a strong baseline and we hope this work will foster future research into this task.• We design a learning-based 3D GAN inversion with the pseudo-multi-view training strategy to extract ge- ometry and texture prior from arbitrary input images. We further utilize these strong prior by designing a la- tent code manipulating algorithm, with which we di- rectly synthesize the final results with the pretrained generator. • Extensive experiments on benchmark datasets demon- strate the superiority of the proposed 3dSwap over state-of-the-art 2D face swapping approaches in iden- tity transferring. Our reconstruction module for 3D- GAN inversion performs favorably over the state-of- the-art methods as well.
Li_Correlational_Image_Modeling_for_Self-Supervised_Visual_Pre-Training_CVPR_2023	Abstract We introduce Correlational Image Modeling ( CIM), a novel and surprisingly effective approach to self-supervised visual pre-training. Our CIM performs a simple pretext task: we randomly crop image regions (exemplars) from an input image (context) and predict correlation maps be- tween the exemplars and the context. Three key designs enable correlational image modeling as a nontrivial and meaningful self-supervisory task. First, to generate useful exemplar-context pairs, we consider cropping image regions with various scales, shapes, rotations, and transformations. Second, we employ a bootstrap learning framework that in- volves online and target encoders. During pre-training, the former takes exemplars as inputs while the latter converts the context. Third, we model the output correlation maps via a simple cross-attention block, within which the context serves as queries and the exemplars offer values and keys. We show thatCIM performs on par or better than the current state of the art on self-supervised and transfer benchmarks. Code is available at https://github.com/weivision/ Correlational-Image-Modeling.git .	1. Introduction Recent advances in self-supervised visual pre-training have shown great capability in harvesting meaningful representations from hundreds of millions of—often eas-ily accessible— unlabeled images. Among existing pre- training paradigms, Multi-View Self-Supervised Learning (MV-SSL) [8 –12, 21, 23] and Masked Image Modeling (MIM) [2, 22, 54, 68] are two leading methods in the self- supervised learning racetrack, thanks to their nontrivial and meaningful self-supervisory pretext tasks . MV-SSL follows an augment-and-compare paradigm (Figure 1(a)) – randomly transforming an input image into two augmented views and then comparing two different views in the representation space. Such an instance-wise discriminative task is rooted in view-invariant learning [43], i.e., changing views of data does not affect the conveyed information. On the contrary, following the success of Masked Language Modeling (MLM) [16], MIM conducts amask-and-predict pretext task within a single view (Fig- ure 1(b)) – removing a proportion of random image patches and then learning to predict the missing information. This simple patch-wise generative recipe enables Transformer- based deep architectures [17] to learn generalizable repre- sentations from unlabeled images. Beyond augment-and-compare ormask-and-predict pre- text tasks in MV-SSL and MIM, in this paper, we endeavor to investigate another simple yet effective paradigm for self- supervised visual representation learning. We take inspira- tion from visual tracking [70] in computer vision that defines the task of estimating the motion or trajectory of a target object ( exemplar ) in a sequence of scene images ( contexts ). To cope with challenging factors such as scale variations, deformations, and occlusions, one typical tracking pipeline 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. 15105 is formulated as maximizing the correlation between the spe- cific exemplar and holistic contexts [3,5,46,52]. Such simple correlational modeling can learn meaningful representations in the capability of both localization and discrimination, thus making it appealing to serve as a promising pretext task for self-supervised learning. Training a standard correlational tracking model, however, requires access to numerous labeled data, which is unavail- able in unsupervised learning. Also, the task goal of vi- sual tracking is intrinsically learning toward one-shot object detection—demanding rich prior knowledge of objectness— while less generic for representation learning. Therefore, it is nontrivial to retrofit supervised correlational modeling for visual tracking into a useful self-supervised pretext task. Driven by this revelation, we present a novel crop-and- correlate paradigm for self-supervised visual representa- tion learning, dubbed as Correlational Image Modeling (CIM). To enable correlational modeling for effectively self- supervised visual pre-training, we introduce three key de- signs. First, as shown in Figure 1(c), we randomly crop image regions (treated as exemplars ) with various scales, shapes, rotations, and transformations from an input image (context ). The corresponding correlation maps can be de- rived from the exact crop regions directly. This simple crop- ping recipe allows us to easily construct the exemplar-context pairs together with ground-truth correlation maps without human labeling cost. Second, we employ a bootstrap learn- ing framework that is comprised of two networks: an online encoder and a target encoder, which, respectively, encode exemplars andcontext into latent space. This bootstrapping effect works in a way that the model learns to predict the spatial correlation between the updated representation of exemplars and the slow-moving averaged representation of context . Third, to realize correlational learning, we introduce a correlation decoder built with a cross-attention layer and a linear predictor, which computes queries from context , with keys and values from exemplars , to predict the corresponding correlation maps. Our contributions are summarized as follows: 1)We present a simple yet effective pretext task for self-supervised visual pre-training, characterized by a novel unsupervised correlational image modeling framework ( CIM).2)We demonstrate the advantages of our CIM in learning trans- ferable representations for both ViT and ResNet models that can perform on par or better than the current state-of-the-art MIM and MV-SSL learners while improving model robust- ness and training efficiency. We hope our work can motivate future research in exploring new useful pretext tasks for self-supervised visual pre-training.
Khan_Temporally_Consistent_Online_Depth_Estimation_Using_Point-Based_Fusion_CVPR_2023	Abstract Depth estimation is an important step in many computer vision problems such as 3D reconstruction, novel view syn- thesis, and computational photography. Most existing work focuses on depth estimation from single frames. When ap- plied to videos, the result lacks temporal consistency, show- ing flickering and swimming artifacts. In this paper we aim to estimate temporally consistent depth maps of video streams in an online setting. This is a difficult problem as future frames are not available and the method must choose between enforcing consistency and correcting errors from previous estimations. The presence of dynamic objects fur- ther complicates the problem. We propose to address these challenges by using a global point cloud that is dynami- cally updated each frame, along with a learned fusion ap- proach in image space. Our approach encourages consis- tency while simultaneously allowing updates to handle er- rors and dynamic objects. Qualitative and quantitative re- sults show that our method achieves state-of-the-art quality for consistent video depth estimation.	1. Introduction Depth reconstruction is a long-standing, fundamental problem in computer vision. For decades the most popular depth estimation techniques were based on stereo match- ing [37] or structure-from-motion [36]. However, more re- cently the best results have come from learning-based ap- proaches [7]. As the overall reconstruction quality has im- proved, focus has shifted to new areas such as monocular es- timation [21,32,33], edge quality [58], and temporal consis- tency [20, 26]. The latter is particularly important for video applications in computational photography and virtual re- ality [3, 52] as inconsistent depth can cause objectionable flickering and swimming artifacts. Consistent video depth estimation, however, remains a difficult problem as even the best method will suffer from unpredictable errors and imperfections based on scene con- tent, especially in textureless and specular regions. This difficulty is aggravated by many of the aforementioned ap- plications requiring online reconstruction: future frames are Figure 1. Comparing monocular depth estimation on the ScanNet dataset across four frames. Clockwise from Top-left: Input RGB image, Ranftl et al.’s DPT [32], our result with a DPT backbone, RGB-D sensor ground truth. not known beforehand and temporal consistency must be balanced with error-correction. Furthermore, the presence of dynamic objects — which are inherently inconsistent — adds an additional layer of complication. Luo et al. [26] address these problems by assuming all frames are known beforehand and fine-tuning their method for each input video. Other approaches seek an online solution by encod- ing consistency in network weights either through a training loss [21], by using recurrent architectures [8, 30, 56], or by conditioning it on the input [24]. Each method, however, ultimately relies on the raw output of a neural network be- ing consistent, which is difficult to achieve due to camera noise and aliasing in convolutional networks [14, 44, 57]. In this work, we propose the use of a global point cloud to encourage temporal consistency in online video depth es- timation. We demonstrate how to tackle the twin problems of handling dynamic objects and updating a static — and potentially erroneous — point cloud when future frames are not known. With quantitative and qualitative results, we show our approach significantly improves the temporal consistency of both stereo and monocular depth estimation without sacrificing spatial quality. In summary, our contri- butions are as follows: 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. 9119 k Global Point CloudUpdate Global Point Cloud𝑑𝑑f𝑡𝑡Temporal Fusion Spatial Fusion Splat Point Cloud𝑐𝑐p𝑡𝑡,𝑑𝑑p𝑡𝑡 RGBD video Time….t t-1 t-2 t t-1 t-2 Temporally Consistent Depth 𝑐𝑐𝑡𝑡,𝑑𝑑𝑡𝑡𝑑𝑑o𝑡𝑡Figure 2. We generate temporally consistent depth maps for each RGB video frame tby fusing the projected depth from a prior point cloud dt pwith the estimated depth dt. This is done by temporally fusing dt pto update dynamic regions, followed by spatial fusion with dtbased on confidence estimates of accuracy. The final result dt ois used to update the point cloud for the next frame. 1. We propose point cloud-based fusion for temporally consistent video depth estimation. 2. We present a three-stage approach to encourage con- sistency in online settings, while simultaneously al- lowing updates to improve the accuracy of reconstruc- tion and handle dynamic scenes. 3. We present an image-space approach to dynamics es- timation and depth fusion that is lightweight and has low runtime overhead.
Kong_LaserMix_for_Semi-Supervised_LiDAR_Semantic_Segmentation_CVPR_2023	Abstract Densely annotating LiDAR point clouds is costly, which often restrains the scalability of fully-supervised learning methods. In this work, we study the underexplored semi- supervised learning (SSL) in LiDAR semantic segmenta- tion. Our core idea is to leverage the strong spatial cues of LiDAR point clouds to better exploit unlabeled data. We propose LaserMix to mix laser beams from different Li- DAR scans and then encourage the model to make con- sistent and confident predictions before and after mixing. Our framework has three appealing properties. 1) Generic: LaserMix is agnostic to LiDAR representations (e.g., range view and voxel), and hence our SSL framework can be uni- versally applied. 2) Statistically grounded: We provide a detailed analysis to theoretically explain the applicabil- ity of the proposed framework. 3) Effective: Comprehen- sive experimental analysis on popular LiDAR segmentation (∗)Lingdong and Jiawei contributed equally to this work. (B)Ziwei serves as the corresponding author. E-mail: [email protected] .datasets (nuScenes, SemanticKITTI, and ScribbleKITTI) demonstrates our effectiveness and superiority. Notably, we achieve competitive results over fully-supervised coun- terparts with 2×to5×fewer labels and improve the supervised-only baseline significantly by relatively 10.8%. We hope this concise yet high-performing framework could facilitate future research in semi-supervised LiDAR seg- mentation. Code is publicly available1.	1. Introduction LiDAR segmentation is one of the fundamental tasks in autonomous driving perception [41]. It enables autonomous vehicles to semantically perceive the dense 3D structure of the surrounding scenes [15, 34, 39]. However, densely an- notating LiDAR point clouds is inevitably expensive and labor-intensive [18, 23, 47], which restrains the scalability of fully-supervised LiDAR segmentation methods. Semi- supervised learning (SSL) that directly leverages the easy- to-acquire unlabeled data is hence a viable and promising 1https://github.com/ldkong1205/LaserMix . 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. 21705 solution to achieve scalable LiDAR segmentation [13, 14]. Yet, semi-supervised LiDAR segmentation is still under- explored. Modern SSL frameworks are mainly designed for image recognition [2, 3, 42] and semantic segmenta- tion [6,21,37] tasks, which only yield sub-par performance on LiDAR data due to the large modality gap between 2D and 3D. Recent research [20] proposed to consider semi- supervised point cloud semantic segmentation as a fresh task and proposed a point contrastive learning framework. However, their solutions do not differentiate indoor and out- door scenes and therefore overlook the intrinsic and impor- tant properties that only exist in LiDAR point clouds. In this work, we explore the use of spatial prior for semi- supervised LiDAR segmentation. Unlike the general 2D/3D segmentation tasks, the spatial cues are especially signif- icant in LiDAR data. In fact, LiDAR point clouds serve as a perfect reflection of real-world distributions, which is highly dependent on the spatial areas in the LiDAR- centered 3D coordinates. As shown in Fig. 1 (left), the top laser beams travel outward long distance and perceive mostly vegetation , while the middle and bottom beams tend to detect carandroad from the medium and close distances, respectively. To effectively leverage this strong spatial prior, we propose LaserMix to mix laser beams from different LiDAR scans, and then encourage the LiDAR segmenta- tion model to make consistent and confident predictions be- fore and after mixing. Our SSL framework is statistically grounded, which consists of the following components: 1)Partitioning the LiDAR scan into low-variation areas. We observe a strong distribution pattern on laser beams as shown in Fig. 1 (left) and thus propose the laser partition. 2)Efficiently mixing every area in the scan with foreign data and obtaining model predictions. We propose Laser- Mix to manipulate the laser-grouped areas from two LiDAR scans in an intertwining way as depicted in Fig. 1 (middle) and serves as an efficient LiDAR mixing strategy for SSL. 3)Encouraging models to make confident and consistent predictions on the same area in different mixing. We hence propose a mixing-based teacher-student training pipeline. Despite the simplicity of our overall pipeline, it achieves competitive results over the fully supervised counterpart us- ing2×to5×fewer labels as shown in Fig. 1 (right) and sig- nificantly outperforms all prevailing semi-supervised seg- mentation methods on nuScenes [11] (up to +5.7%mIoU) and SemanticKITTI [1] (up to +3.5%mIoU). Moreover, LaserMix directly operates on point clouds so as to be agnostic to different LiDAR representations, e.g., range view [32] and voxel [58]. Therefore, our pipeline is highly compatible with existing state-of-the-art (SoTA) LiDAR segmentation methods under various representa- tions [46, 56, 57]. Besides, our pipeline achieves compet- itive performance using very limited annotations on weak supervision dataset [47]: it achieves 54.4%mIoU on Se-manticKITTI [1] using only 0.8%labels, which is on-par with PolarNet [56] ( 54.3%), RandLA-Net [19] ( 53.9%), and RangeNet++ [32] ( 52.2%) using 100% labels. Spatial prior is proven to play a pivotal role in the success of our framework through comprehensive empirical analysis. To summarize, this work has the following key contributions: • We present a statistically grounded SSL framework that effectively leverages the spatial cues in LiDAR data to facilitate learning with semi-supervisions. • We propose LaserMix, a novel and representation- agnostic mixing technique that strives to maximize the “strength” of the spatial cues in our SSL framework. • Our overall pipeline significantly outperforms previ- ous SoTA methods in both low- and high-data regimes. We hope this work could lay a solid foundation for semi-supervised LiDAR segmentation.
Lei_EFEM_Equivariant_Neural_Field_Expectation_Maximization_for_3D_Object_Segmentation_CVPR_2023	Abstract We introduce Equivariant Neural Field Expectation Maximization ( EFEM ), a simple, effective, and robust ge- ometric algorithm that can segment objects in 3D scenes without annotations or training on scenes. We achieve such unsupervised segmentation by exploiting single ob- ject shape priors. We make two novel steps in that direc- tion. First, we introduce equivariant shape representations to this problem to eliminate the complexity induced by the variation in object configuration. Second, we propose a novel EM algorithm that can iteratively refine segmenta- tion masks using the equivariant shape prior. We collect a novel real dataset Chairs and Mugs that contains vari- ous object configurations and novel scenes in order to verify the effectiveness and robustness of our method. Experimen- tal results demonstrate that our method achieves consis- tent and robust performance across different scenes where the (weakly) supervised methods may fail. Code and data available at https://www.cis.upenn.edu/ ˜leijh/ projects/efem	1. Introduction Learning how to decompose 3D scenes into object in- stances is a fundamental problem in visual perception sys- tems. Past developments in 3D computer vision have made huge strides on this problem by training neural net- works on 3D scene datasets with segmentation masks [55, 63, 67]. However, these works heavily rely on large la- beled datasets [3, 15] that require laborious 3D annotation based on special expertise. Few recent papers alleviate this problem by reducing the need to either sparse point label- ing [24, 60] or bounding boxes [12]. In this work, we follow an object-centric approach in- spired by the Gestalt school of perception that captures an object as a whole shape [32, 47] invariant to its pose and scale [31]. A holistic approach builds up a prior for each object category, that then enables object recognition in dif- ferent complex scenes with varying configurations. Directly learning object-centric priors instead of analyzing each 3D Figure 1. We present EFEM, an unsupervised 3D object segmen- tation method applicable to real-world scenes (results on the right) by only training on ShapeNet single object reconstruction. scene inspires a more efficient way of learning instance seg- mentation: both a mug on the table and a mug in the dish- washer are mugs, and one does not have to learn to seg- ment out a mug in all possible environmental contexts if we have a unified shape concept for mugs. Such object- centric recognition facilitates a robust scene analysis for au- tonomous systems in many interactive real-world environ- ments with a diversity of object configurations: Imagine a scenario where a robot is doing the dishes in the kitchen. Dirty bowls are piled in the sink and the robot is clean- ing them and placing them into a cabinet. Objects of the same category appear in the scene repeatedly under differ- ent configurations (piles, neat lines in the cabinet). What is even more challenging is that even within this one single task (doing dishes) the scene configuration can drastically change when objects are moved. We show that such scenar- ios cannot be addressed by the state-of-the-art strongly or weakly supervised methods that struggle under such scene configuration variations. In this paper, we introduce a method that can segment 3D object instances from 3D static scenes by learning pri- ors of single object shapes (ShapeNet [4]) without using any scene-level labels. Two main challenges arise when we re- move the scene-level annotation. First, objects in the scene can have a different position, rotation, and scale than the canonical poses where the single object shapes were trained. Second, the shape encoder which is trained on object-level input cannot be directly applied to the scene observations unless the object masks are known. We address the first challenge by introducing equivariance to this problem. By 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. 4902 learning a shape prior that is equivariant to the similitude group SIM(3), the composition of a rotation, a translation, and a uniform scaling in 3D (Sec. 3.1), we address the com- plexity induced by the SIM(3) composition of objects. For the second challenge, we introduce a simple and effective iterative algorithm, Equivariant neural Field Expectation Maximization ( EFEM ), that refines the object segmentation mask, by alternately iterating between mask updating and shape reconstruction (Sec. 3.2). The above two steps en- able us to directly exploit the learned single instance shape prior to perform segmentation in real-world scenes. We col- lected and annotated a novel real-world test set (240 scenes) (Sec. 4.4) that contains diverse object configurations and novel scenes to evaluate the generalizability and robustness to novel object instances and object configuration changes. Experiments on both synthetic data (Sec. 4.3) and our novel real dataset (Sec. 4.4) give us an insight to the effectiveness of the method. Compared to weakly supervised methods, when the testing scene setup is similar to the training setup, our method has a small performance gap to the (weakly) supervised baselines. However, when the testing scenes are drawn from novel object configurations, our method consis- tently outperforms the (weakly) supervised baselines. Our paper makes the following novel contributions to the 3D scene segmentation problem: (1) a simple and effective iterative EM algorithm that can segment objects from the scenes using only single object shape priors. (2) addressing the diversity of object composition in 3D scenes by combin- ing representations equivariant to rotation, translation, and scaling of the objects. (3) an unsupervised pipeline for 3D instance segmentation that works in real-world data and can generalize to novel setups. (4) a novel real-world test set Chairs and Mugs that contains diverse object configura- tions and scenes.
Ke_Neural_Preset_for_Color_Style_Transfer_CVPR_2023	Abstract In this paper, we present a Neural Preset technique to address the limitations of existing color style transfer meth- ods, including visual artifacts, vast memory requirement, and slow style switching speed. Our method is based on two core designs. First, we propose Deterministic Neural Color Mapping (DNCM) to consistently operate on each pixel via an image-adaptive color mapping matrix, avoiding artifacts and supporting high-resolution inputs with a small memory footprint. Second, we develop a two-stage pipeline by divid- ing the task into color normalization and stylization, which allows efficient style switching by extracting color styles as presets and reusing them on normalized input images. Due to the unavailability of pairwise datasets, we describe how to train Neural Preset via a self-supervised strategy. Vari- ous advantages of Neural Preset over existing methods are demonstrated through comprehensive evaluations. Besides, we show that our trained model can naturally support mul- tiple applications without fine-tuning, including low-light image enhancement, underwater image correction, image dehazing, and image harmonization. The project page is: https://ZHKKKe.github.io/NeuralPreset.	1. Introduction With the popularity of social media ( e.g., Instagram and Facebook), people are increasingly willing to share pho- tos in public. Before sharing, color retouching becomes an indispensable operation to help express the story cap- tured in images more vividly and leave a good first impres- sion. Photo editing tools usually provide color style presets, such as image filters or Look-Up Tables (LUTs), to help users explore efficiently. However, these filters/LUTs are handcrafted with pre-defined parameters, and are not able to generate consistent color styles for images with diverse appearances. Therefore, careful adjustments by the users is still necessary. To address this problem, color style trans- fer techniques have been introduced to automatically map the color style from a well-retouched image ( i.e., the style image) to another ( i.e., the input image). Earlier color style transfer methods [41–43, 49] focus on retouching the input image according to low-level fea- ture statistics of the style image. They disregard high-level information, resulting in unexpected changes in image in- †Corresponding author. This project is in part supported by a General Research Fund from RGC of Hong Kong (RGC Ref.: 11205620). 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. 14173 herent colors. Although recent deep learning based mod- els [1,6,19,34,36,54] give promising results, they typically suffer from three obvious limitations in practice (Fig. 1 (a)). First, they produce unrealistic artifacts ( e.g., distorted tex- tures or inharmonious colors) in the stylized image since they perform color mapping based on convolutional mod- els, which operate on image patches and may have incon- sistent outputs for pixels with the same value. Although some auxiliary constraints [36] or post-processing strate- gies [34] have been proposed, they still fail to prevent ar- tifacts robustly. Second, they cannot handle high-resolution (e.g., 8K) images due to their huge runtime memory foot- print. Even using a GPU with 24GB of memory, most recent models suffer from the out-of-memory problem when pro- cessing 4K images. Third, they are inefficient in switching styles because they carry out color style transfer as a single- stage process, requiring to run the whole model every time. In this work, we present a Neural Preset technique with two core designs to overcome the above limitations: (1)Neural Preset leverages Deterministic Neural Color Mapping (DNCM) as an alternative to the color mapping process based on convolutional models. By multiplying an image-adaptive color mapping matrix, DNCM converts pixels of the same color to a specific color, avoiding un- realistic artifacts effectively. Besides, DNCM operates on each pixel independently with a small memory footprint, supporting very high-resolution inputs. Unlike adaptive 3D LUTs [7, 55] that need to regress tens of thousands of pa- rameters or automatic filters [23, 27] that perform particu- lar color mappings, DNCM can model arbitrary color map- pings with only a few hundred learnable parameters. (2)Neural Preset carries out color style transfer in two stages to enable fast style switching. Specifically, the first stage builds a nDNCM from the input image for color nor- malization, which maps the input image to a normalized color style space representing the “image content”; the sec- ond stage builds a sDNCM from the style image for color stylization, which transfers the normalized image to the tar- get color style. Such a design has two advantages in terms of efficiency: the parameters of sDNCM can be stored as color style presets and reused by different input images, while the input image can be stylized by diverse color style presets after normalized once with nDNCM . In addition, since there are no pairwise datasets avail- able, we propose a new self-supervised strategy for Neu- ral Preset to be trainable. Our comprehensive evaluations demonstrate that Neural Preset outperforms state-of-the-art methods significantly in various aspects. Notably, Neural Preset can produce faithful results for 8K images (Fig. 1 (b)) and can provide consistent color style transfer results across video frames without post-processing. Compared to re- cent deep learning based models, Neural Preset achieves ∼28×speedup on a Nvidia RTX3090 GPU, supportingreal-time performances at 4K resolution. Finally, we show that our trained model can be applied to other color map- ping tasks without fine-tuning, including low-light image enhancement [30], underwater image correction [52], im- age dehazing [16], and image harmonization [37].
Ling_Learning_Optical_Expansion_From_Scale_Matching_CVPR_2023	Abstract This paper address the problem of optical expansion (OE). OE describes the object scale change between two frames, widely used in monocular 3D vision tasks. Previ- ous methods estimate optical expansion mainly from opti- cal flow results, but this two-stage architecture makes their results limited by the accuracy of optical flow and less ro- bust. To solve these problems, we propose the concept of 3D optical flow by integrating optical expansion into the 2D optical flow, which is implemented by a plug-and-play module, namely TPCV . TPCV implements matching features at the correct location and scale, thus allowing the simul- taneous optimization of optical flow and optical expansion tasks. Experimentally, we apply TPCV to the RAFT optical flow baseline. Experimental results show that the baseline optical flow performance is substantially improved. More- over, we apply the optical flow and optical expansion re- sults to various dynamic 3D vision tasks, including motion- in-depth, time-to-collision, and scene flow, often achiev- ing significant improvement over the prior SOTA. Code is available at https://github.com/HanLingsgjk/ TPCV .	1. Introduction Optical expansion (OE) is a fundamental and important concept in monocular dynamic 3D vision tasks [1, 3, 22, 23, 32]. OE describes the scale change of an object between two frames, which can be translated into motion in the depth direction. It has essential applications in time-to-collision, scene flow, and motion-in-depth estimation. OE schemes have unique advantages in 3D motion estimation tasks, re- quiring only a single camera and enabling dense and fixed baseline independent results. In this paper, we discuss a ro- bust and novel approach for OE estimation. *Corresponding author †Corresponding author 𝐹rame1 𝑀𝑢𝑙𝑡𝑖 𝑆𝑐𝑎𝑙𝑒 𝐹rame2 𝑓𝑟𝑎𝑚𝑒 1 𝑠ൌ 1 𝑓𝑟𝑎𝑚𝑒 2 𝑠ൌ 1 𝑠ൌ 0.7 𝑠ൌ 0.5Figure 1. Scale matching idea. Left: multi-scale matching be- tween two frames. Right: texture around the matching point, where sis the size of the image scaling. The core idea of scale matching is to match texture features at the correct location and scale. As seen above, the texture at the license plate can be better matched when the second frame is scaled 0.7 times in size, where the scaling magnification also reflects the optical expansion of that local pixel between the two frames. Furthermore, scale matching can better handle the motion in the depth direction and contains potential 3D motion information. Prior works In early time-to-collision (TTC) studies [4, 22, 23], OE was obtained from motion modeling, where the motion estimation was provided by optical flow or SIFT [10]. Such algorithms relied on optical flow results and specific model assumptions, often yielding only sparse and low-accuracy results. Some recent methods [32] regress OE based on existing optical flow results and achieve better out- comes. However, these two-stage estimation methods rely on accurate optical flow results and decrease computational efficiency. Instead, we consider optical flow and expan- sion estimation as two complementary tasks. Introduc- ing OE in optical flow can realize matching features at the correct location and scale. As shown in Fig. 1, this fusion 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. 5445 𝑠ൌ 1.25 𝑠ൌ 1 𝑠ൌ 1/1.25 𝑠ൌ 0.5𝑎 𝑖𝑚𝑎𝑔𝑒 𝑡𝑜 𝑖𝑚𝑎𝑔𝑒 𝑏 𝑖𝑚𝑎𝑔𝑒 𝑡𝑜 𝑝𝑦𝑟𝑎𝑚𝑖𝑑𝑓𝑟𝑎𝑚𝑒 1 𝑓𝑟𝑎𝑚𝑒 2 𝑠ൌ 0.75 𝑠ൌ 1 𝑠ൌ 0.5 𝑠ൌ 0.75 𝑐 𝑇𝑃𝐶𝑉Figure 2. Three different matching modes . We match objects between two consecutive frames, where the cat is away from the camera and the car is close to the camera. (a) The 2D optical flow matches the cat and car in the original size image. (b) Match the second frame with the first frame after multiscale scaling, we found that cat can be better matched when magnified by 1.25 times, and car can be better matched when shrunk by 0.75 times. However, obtaining an accurate enlarged picture of a cat is impos- sible. (c) Transpose matching, where the cat zoomed in in frame 1 to
Liu_MMVC_Learned_Multi-Mode_Video_Compression_With_Block-Based_Prediction_Mode_Selection_CVPR_2023	Abstract Learning-based video compression has been extensively studied over the past years, but it still has limitations in adapting to various motion patterns and entropy models. In this paper, we propose multi-mode video compression (MMVC), a block wise mode ensemble deep video com- pression framework that selects the optimal mode for fea- ture domain prediction adapting to different motion pat- terns. Proposed multi-modes include ConvLSTM-based fea- ture domain prediction, optical flow conditioned feature do- main prediction, and feature propagation to address a wide range of cases from static scenes without apparent mo- tions to dynamic scenes with a moving camera. We parti- tion the feature space into blocks for temporal prediction in spatial block-based representations. For entropy coding, we consider both dense and sparse post-quantization resid- ual blocks, and apply optional run-length coding to sparse residuals to improve the compression rate. In this sense, our method uses a dual-mode entropy coding scheme guided by a binary density map, which offers significant rate reduc- tion surpassing the extra cost of transmitting the binary se- lection map. We validate our scheme with some of the most popular benchmarking datasets. Compared with state-of- the-art video compression schemes and standard codecs, our method yields better or competitive results measured with PSNR and MS-SSIM.	1. Introduction Over the past several years, with the emergence and booming of short videos and video conferences across the world, video has become the major container of informa- tion and interaction among people on a daily basis. Conse- quently, we have been witnessing a vast demand increase on transmission bandwidth and storage space, together with the vibrant growth and discovery of handcrafted codecs such as A VC/H.264 [23], HEVC [23], and the recently released *Equally contributed authors.VVC [22], along with a number of learning based meth- ods [1, 7, 9, 11, 12, 15–17, 21, 27, 30, 31]. Prior works in deep video codecs have underlined the importance of utilizing and benefiting from deep neural net- work models, which can exploit complex spatial-temporal correlations and have the ability of ‘learning’ contextual and motion features. The main objective of deep video com- pression is to predict the next frame from previous frames or historical data, which results in the reduction of amount of residual information that needs to be encoded and trans- mitted. This has so far led to two directions: (1) to build ef- ficient prediction or estimation models, and extract motion information by leveraging the temporal correlation across the frames [1, 9, 16, 31]; (2) to make accurate estimation of the distribution of residual data and push down the in- formation entropy statistically by appropriate conditioning [7, 11, 30]. The existing works usually fall in one or a com- bination of the above two realms. In the light of learning capability that deep neural networks can offer, we argue and demonstrate in this work that some measures of adaptively selecting the right mode among different available models in the encoding path can be advantageous on top of the ex- isting schemes, especially when the adaptive model selec- tion is applied at the block level in the feature space. Drawing wisdom from conventional video codec stan- dards that typically address various types of motions (in- cluding the unchanged contextual information) in the unit of macroblocks, we present a learning-based, block wise video compression scheme that applies content-driven mode se- lection on the fly. Our proposed method consists of four modes targeting different scenarios: •Skip mode (S) aims to utilize the frame buffer on the decoder and find the most condensed representation to transmit unchanged blocks to achieve the best possi- ble bitrate. This mode is particularly useful for static scenes where same backgrounds are captured by a fixed view camera. •Optical Flow Conditioned Feature Prediction mode (OFC) leverages the temporal locality of motions. 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. 18487 Figure 1. Overview of our proposed multi-mode video coding method. The current and previous frames are fed into the feature extractor and then go through branches of prediction modes followed by residual channel removal, quantization, and entropy coding process. We then select the optimal prediction and entropy coding schemes for each block that lead to the smallest code size. this mode, we capture the optical flow [24] between the past two frames, and the warped new frame is treated as a preliminary prediction to the current frame. This warping serves as the condition to provide guidance to the temporal prediction DNN model. •Feature Propagation mode (FPG) applies to blocks where changes are detected, but there is no better pre- diction mode available. This mode copies the previ- ously reconstructed feature block as the prediction, and encodes the residual from there. • For other generic cases, we propose the Feature Pre- diction mode (FP) for feature domain inter-frame pre- diction with a ConvLSTM network to produce a pre- dicted current frame (block). Prior to the mode selection step, The transmitter pro- duces the optimal low-dimensional representation of each frame using a learned encoder and decoder pair based on the image compression framework in [14] for the mapping from pixel to feature space. The block by block difference between the previous frame and the current frame repre- sents the block wise motion. Unlike some state-of-the-art video compression frameworks that separately encode mo- tions and residuals, our method does not encode the motion as it is automatically generated by the prediction using the information available on both the transmitter and receiver. To adapt to different dynamics that may exist even within a single frame for different blocks, our method evaluatesmultiple prediction modes that are listed above at the block level. As a result, we can always obtain residuals that have the highest sparsity thereby the shortest code length per block. Furthermore, we propose a residual channel remov- ing strategy to mask out residual channels that are inessen- tial to frame reconstruction, exploiting favorable tradeoffs between noticeably higher compression ratio and negligible quality degradation. Technical contributions of this work are summarized as follows: • We present MMVC, a dynamic mode selection-based video compression approach that adapts to different motion and contextual patterns with Skip mode and different feature-domain prediction paths in the unit of block. • To improve the residual sparsity without losing much quality while minimizing the bitrate, we propose a block wise channel removal scheme and a density- adaptive entropy coding strategy. • We perform extensive experiments and a compara- tive study to showcase that MMVC exhibits supe- rior or similar performance compared to state-of-the- art learning-based methods and conventional codecs. In the ablation study, we show the effectiveness of our scheme by quantifying the utilization of different modes that varies by video contents and scenes. 18488
Lin_DynamicDet_A_Unified_Dynamic_Architecture_for_Object_Detection_CVPR_2023	Abstract Dynamic neural network is an emerging research topic in deep learning. With adaptive inference, dynamic mod- els can achieve remarkable accuracy and computational efficiency. However, it is challenging to design a power- ful dynamic detector, because of no suitable dynamic ar- chitecture and exiting criterion for object detection. To tackle these difficulties, we propose a dynamic framework for object detection, named DynamicDet. Firstly, we care- fully design a dynamic architecture based on the nature of the object detection task. Then, we propose an adaptive router to analyze the multi-scale information and to de- cide the inference route automatically. We also present a novel optimization strategy with an exiting criterion based on the detection losses for our dynamic detectors. Last, we present a variable-speed inference strategy, which helps to realize a wide range of accuracy-speed trade-offs with only one dynamic detector. Extensive experiments conducted on the COCO benchmark demonstrate that the proposed DynamicDet achieves new state-of-the-art accuracy-speed trade-offs. For instance, with comparable accuracy, the inference speed of our dynamic detector Dy-YOLOv7-W6 surpasses YOLOv7-E6 by 12%, YOLOv7-D6 by 17%, and YOLOv7-E6E by 39%. The code is available at https: //github.com/VDIGPKU/DynamicDet .	1. Introduction Object detection is an essential topic in computer vision, as it is a fundamental component for other vision tasks, e.g., autonomous driving [26, 40, 56], multi-object track- ing [52,57], intelligent transportation [36,55], etc. In recent years, tremendous progress has been made toward more ac- curate and faster detectors, such as Network Architecture Search (NAS)-based detectors [10,25,48] and YOLO series models [2, 9, 11, 21, 44, 45]. However, these methods need to design and train multiple models to achieve a few good trade-offs between accuracy and speed, which is not flexible enough for various application scenarios. To alleviate this †Corresponding author. 12.5 15.0 17.5 20.0 22.5 25.0 27.5 30.0 V100 batch 1 inference time (ms)525354555657COCO APtest (%) Dy-YOLOv7 (Ours) YOLOv7 PP-YOLOE+ YOLOv5 (r6.2) YOLOv6Figure 1. Comparison of the proposed dynamic detectors and other efficient object detectors. Our method can achieve a wide range of state-of-the-art trade-offs between accuracy and speed with a single model. “Easy” image“Hard” image Figure 2. Examples of “easy” and “hard” images for the object detection task. problem, we focus on dynamic inference for the object de- tection task, and attempt to use only one dynamic detector to achieve a wide range of good accuracy-speed trade-offs, as shown in Fig. 1. The human brain inspires many fields of deep learning, and the dynamic neural network [12] is a typical one. As two examples shown in Fig. 2, we can quickly identify all objects on the left “easy” image, while we need more time to achieve the same effect for the right one. In other words, the processing speeds of images are different in our brains [18, 34], which depend on the difficulties of the im- ages. This property motivates the image-wise dynamic neu- ral network, and many exciting works have been proposed 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. 6282 (e.g., Branchynet [43], MSDNet [17], DVT [50]). Although these approaches have achieved remarkable performance, they are all designed specifically for the image classifica- tion task and are not suitable for other vision tasks, espe- cially for the object detection [12]. The main difficulties in designing an image-wise dynamic detector are as follows. Dynamic detectors cannot utilize the existing dy- namic architectures. Most existing dynamic architectures are cascaded with multiple stages ( i.e., a stack of multiple layers) [17, 20, 33, 54], and predict whether to stop the in- ference at each exiting point. Such a paradigm is feasible in image classification but is ineffective in object detection, since an image has multiple objects and each object usu- ally has different categories and scales, as shown in Fig. 2. Hence, almost all detectors depend heavily on multi-scale information, utilizing the features on different scales to de- tect objects of different sizes (which are obtained by fusing the multi-scale features of the backbone with a detection neck, i.e., FPN [27]). In this case, the exiting points for detectors can only be placed behind the last stage. Con- sequently, the entire backbone module has to be run com- pletely [58], and it is impossible to achieve dynamic infer- ence on multiple cascaded stages. Dynamic detectors cannot exploit the existing exiting criteria for image classification. For the image classi- fication task, the threshold of top-1 accuracy is a widely used criterion for decision-making [17, 50]. Notably, it only needs one fully connected layer to predict the top-1 accuracy at intermediate layer, which is easy and costless. However, object detection task requires the neck and the head to predict the categories and locations of the object in- stances [3, 14, 27, 39]. Hence, the existing exiting criteria for image classification is not suitable for object detection. To deal with the above difficulties, we propose a dynamic framework to achieve dynamic inference for object detec- tion, named DynamicDet. Firstly, We design a dynamic ar- chitecture for the object detection task, which can exit with multi-scale information during the inference. Then, we pro- pose an adaptive router to choose the best route for each image automatically. Besides, we present the correspond- ing optimization and inference strategies for the proposed DynamicDet. Our main contributions are as follows: • We propose a dynamic architecture for object detec- tion, named DynamicDet, which consists of two cas- caded detectors and a router. This dynamic architec- ture can be easily adapted to mainstream detectors, e.g., Faster R-CNN and YOLO. • We propose an adaptive router to predict the difficulty scores of the images based on the multi-scale features, and achieve automatic decision-making. In addition, we propose a hyperparameter-free optimization strat- egy and a variable-speed inference strategy for our dy-namic architecture. • Extensive experiments show that DynamicDet can ob- tain a wide range of accuracy-speed trade-offs with only one dynamic detector. We also achieve new state- of-the-art trade-offs for real-time object detection ( i.e., 56.8% AP at 46 FPS).
Lentsch_SliceMatch_Geometry-Guided_Aggregation_for_Cross-View_Pose_Estimation_CVPR_2023	Abstract This work addresses cross-view camera pose estimation, i.e., determining the 3-Degrees-of-Freedom camera pose of a given ground-level image w.r.t. an aerial image of the lo- cal area. We propose SliceMatch, which consists of ground and aerial feature extractors, feature aggregators, and a pose predictor. The feature extractors extract dense features from the ground and aerial images. Given a set of can- didate camera poses, the feature aggregators construct a single ground descriptor and a set of pose-dependent aerial descriptors. Notably, our novel aerial feature aggregator has a cross-view attention module for ground-view guided aerial feature selection and utilizes the geometric projec- tion of the ground camera’s viewing frustum on the aerial image to pool features. The efficient construction of aerial descriptors is achieved using precomputed masks. Slice- Match is trained using contrastive learning and pose es- timation is formulated as a similarity comparison between the ground descriptor and the aerial descriptors. Compared to the state-of-the-art, SliceMatch achieves a 19% lower median localization error on the VIGOR benchmark using the same VGG16 backbone at 150 frames per second, and a 50% lower error when using a ResNet50 backbone.	1. Introduction Cross-view camera pose estimation aims to estimate the 3-Degrees-of-Freedom (3-DoF) ground camera pose, i.e., planar location and orientation, by comparing the captured ground-level image to a geo-referenced overhead aerial im- age containing the camera’s local surroundings. In prac- tice, the local aerial image can be obtained from a reference database using any rough localization prior, e.g., Global Navigation Satellite Systems (GNSS), image retrieval [19], or dead reckoning [13]. However, this prior is not nec- essarily accurate, for example, GNSS can contain errors up to tens of meters in urban canyons [2, 48, 49]. The cross-view formulation provides a promising alternative to *indicates equal contribution. (a) Ground image (b) Aerial image 1 0°8 234567 90° 180° 180° 270° (d) Pose heatmap (c) Aerial slicing Cross -view attention Slice masks for a candidate pose… … 4 2 4 2Figure 1. SliceMatch identifies for a ground-level image (a) its camera’s 3-DoF pose within a corresponding aerial image (b). It divides the camera’s Horizontal Field-of-View (HFoV) into ‘slices’, i.e., vertical regions in (a). After self-attention, our novel aggregation step (c) applies cross-view attention to create ground slice-specific aerial feature maps. To efficiently test many candi- date poses, the slice features are aggregated using pose-dependent aerial slice masks that represent the camera’s sliced HFoV at that pose. The slice masks for each pose are precomputed. All aerial pose descriptors are compared to the ground descriptor, resulting in a dense scoring map (d). Our output is the best-scoring pose. ground-level camera pose estimation techniques that require detailed 3D point cloud maps [31] or semantic maps [3,43], since the aerial imagery provides continuous coverage of the Earth’s surface including the area where accurate point clouds are difficult to collect. Moreover, acquiring up-to- date aerial imagery is less costly than maintaining and up- dating large-scale 3D point clouds or semantics maps. Recently, several works have addressed cross-view cam- era localization [55] or 3-DoF pose estimation [33, 36, 44, 50]. Roughly, those methods can be categorized into global image descriptor-based [50, 55] and dense pixel- level feature-based [33, 36, 44] methods. Global descriptor- based methods take advantage of the compactness of the image representation and often have relatively fast infer- ence time [50, 55]. Dense pixel-level feature-based meth- 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. 17225 ods [33, 36, 44] are potentially more accurate as they pre- serve more details in the image representation. They use the geometric relationship between the ground and aerial view to project features across views and estimate the cam- era pose via computationally expensive iterations. Aiming for both accurate and efficient camera pose estimation, in this work, we improve the global descriptor-based approach and enforce feature locality in the descriptor. We observe several limitations in existing global descriptor-based cross-view camera pose estimation meth- ods [50,55]. First, they rely on the aerial encoder to encode all spatial context and the aerial encoder has to learn how to aggregate local information, e.g., via the SAFA mod- ule [34], into the global descriptor, without accessing the information in the ground view or exploiting geometric con- straints between the ground-camera viewing frustum and the aerial image. Second, existing global descriptor-based methods for cross-view localization [50, 55] do not explic- itly consider the orientation of the ground camera in their descriptor construction. As a result, they either do not esti- mate the orientation [55] or require multiple forward passes on different rotated samples to infer the orientation [50]. Third, existing global descriptors-based methods [50, 55] are not trained discriminatively against different orienta- tions. Therefore, the learned features may be less discrimi- native for orientation prediction. To address the observed gaps, we devise a novel, accu- rate, and efficient method for cross-view camera pose es- timation called SliceMatch (see Figure 1). Its novel aerial feature aggregation explicitly encodes directional informa- tion and pools features using known camera geometry to ag- gregate the extracted aerial features into an aerial global de- scriptor. The proposed aggregation step ‘slices’ the ground Horizontal Field-of-View (HFoV) into orientation-specific descriptors. For each pose in a set of candidates, it aggre- gates the extracted aerial features into corresponding aerial slice descriptors. The aggregation uses cross-view attention to weigh aerial features w.r.t. to the ground descriptor, and exploits the geometric constraint that every vertical slice in the ground image corresponds to an azimuth range extrud- ing from the projected ground camera position in the aerial image. The feature extraction is done only once for con- structing the descriptors for all pose candidates, resulting in fast training and inference speed. We contrastively train the model by pairing the ground image descriptor with aerial descriptors at different locations and orientations. Hence, the model learns to extract discriminative features for both localization and orientation estimation. Contributions: i) A novel aerial feature aggregation step that uses a cross-view attention module for ground- view guided aerial feature selection, and the geometric rela- tionship between the ground camera’s viewing frustum and the aerial image to construct pose-dependent aerial descrip-tors. ii)SliceMatch’s design allows for efficient implemen- tation, which runs significantly faster than previous state-of- the-art methods. Namely, for an input ground-aerial image pair, SliceMatch extracts dense features only once, aggre- gates aerial descriptors at a set of poses without extra com- putation, and compares the aerial descriptor of each pose with the ground descriptor. iii)Compared to the previous state-of-the-art global descriptor-based cross-view camera pose estimation method, SliceMatch constructs orientation- aware descriptors and adopts contrastive learning for both locations and orientations. Powered by the above designs, SliceMatch sets the new state-of-the-art for cross-view pose estimation on two commonly used benchmarks.
Li_Center_Focusing_Network_for_Real-Time_LiDAR_Panoptic_Segmentation_CVPR_2023	Abstract LiDAR panoptic segmentation facilitates an autonomous vehicle to comprehensively understand the surrounding ob- jects and scenes and is required to run in real time. The recent proposal-free methods accelerate the algorithm, but their effectiveness and efficiency are still limited owing to the difficulty of modeling non-existent instance centers and the costly center-based clustering modules. To achieve ac- curate and real-time LiDAR panoptic segmentation, a novel center focusing network (CFNet) is introduced. Specifically, the center focusing feature encoding (CFFE) is proposed to explicitly understand the relationships between the origi- nal LiDAR points and virtual instance centers by shifting the LiDAR points and filling in the center points. More- over, to leverage the redundantly detected centers, a fast center deduplication module (CDM) is proposed to select only one center for each instance. Experiments on the Se- manticKITTI and nuScenes panoptic segmentation bench- marks demonstrate that our CFNet outperforms all existing methods by a large margin and is 1.6 times faster than the most efficient method.	1. Introduction Panoptic segmentation [18] combines both semantic seg- mentation and instance segmentation in a single framework. It predicts semantic labels for the uncountable stuff classes (e.g.road,sidewalk ), while it simultaneously provides se- mantic labels and instance IDs for the countable things classes ( e.g.car,pedestrian ). The LiDAR panoptic segmen- tation is one of the bases for the safety of autonomous driv- ing, which employs the point clouds collected by the Light Detection and Ranging (LiDAR) sensors to effectively de- pict the surroundings. Existing LiDAR panoptic segmenta- tion methods first conduct semantic segmentation, and then achieve instance segmentation for the things categories in Figure 1. PQ vs. runtime on the SemanticKITTI test set. Runtime measurements are taken on a single NVIDIA RTX 3090 GPU. The panoptic quality (PQ) is introduced in section 4.1. two ways, the proposal-based and proposal-free methods. The proposal-based methods [17, 31, 37] adopt a two- stage process similar to the well-known Mask R-CNN [14] in the image domain. It first generates object proposals for thethings points by using 3D detection networks [19, 30] and then refines the instance segmentation results within each proposal. As shown in Fig. 1, these methods are usu- ally complicated and hardly achieve real-time processing, owing to their sequential multi-stage pipelines. The proposal-free frameworks [13, 15, 21, 22, 29, 35, 39] are more compact. To associate the things points with in- stance IDs, these methods usually leverage the instance cen- ters. Specifically, they regress the offsets from the points to their corresponding centers, and then adopt the class- agnostic center-based clustering modules [13, 15, 29] or the bird’s-eye view (BEV) center heatmap [22, 35, 39]. How- ever, two problems exist in these methods. First, for center feature extracting and center modeling, the non-existent 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. 13425 (a) CFNet without CFFE (b) CFNet with CFFEFigure 2. Instance segmentation of a car. Without our CFFE, the car is split into parts (a), while the CFFE significantly alleviates this problem (b). Different colors represent different instances. stance centers increase the difficulty, considering that the LiDAR points are usually surface-aggregated [35] and an instance center is imaginary in most cases. As shown in Fig. 2(a), the difficulty often results in the fault that one instance is incorrectly split into several parts. Second, for exploiting the redundantly detected centers, the clus- tering modules ( e.g. MeanShift, DBSCAN) are too time- consuming to support the real-time autonomous driving per- ception systems, while the BEV center heatmap cannot dis- tinguish objects with different altitudes in the same BEV grid. For accurate and fast LiDAR panoptic segmentation, a proposal-free center focusing network (CFNet) is proposed. For better encoding center features, a novel center focusing feature encoding (CFFE) is proposed to generate center- focusing feature maps by shifting the things points to fill in the non-existent instance centers for more accurate pre- dictions (as shown in Fig. 2(b)). For center modeling, the CFNet not only decomposes the panoptic segmentation task into the widely-used semantic segmentation and center off- set regression, but also proposes a new confidence score prediction for indicating the accuracy of the center offset re- gression. Subsequently, for the detected centers exploiting, a novel center deduplication module (CDM) is designed to select one center for a single instance. The CDM keeps the predicted centers with higher confidence scores, while suppressing the ones with lower confidence. Finally, in- stance segmentation is achieved by assigning the shifted things points to the closest center. For efficiency, the pro- posed CFNet is built on the 2D projection-based segmenta- tion paradigm. Our contributions are as follows: • A proposal-free CFNet is proposed to achieve accu- rate and fast LiDAR panoptic segmentation by solving the bottleneck problems of center modeling and center- based clustering in previous methods. • The CFFE is proposed to alleviate the difficulty of modeling the non-existent instance centers and the CDM is designed to efficiently keep one center for each instance.• The proposed CFNet is evaluated on the nuScenes and SemanticKITTI LiDAR panoptic segmentation bench- marks. Our CFNet achieves the state-of-the-art perfor- mance with a real-time inference speed.
Lin_Deep_Frequency_Filtering_for_Domain_Generalization_CVPR_2023	Abstract Improving the generalization ability of Deep Neural Net- works (DNNs) is critical for their practical uses, which has been a longstanding challenge. Some theoretical studies have uncovered that DNNs have preferences for some fre- quency components in the learning process and indicated that this may affect the robustness of learned features. In this paper, we propose Deep Frequency Filtering (DFF) for learning domain-generalizable features, which is the first endeavour to explicitly modulate the frequency components of different transfer difficulties across domains in the latent space during training. To achieve this, we perform Fast Fourier Transform (FFT) for the feature maps at different layers, then adopt a light-weight module to learn attention masks from the frequency representations after FFT to en- hance transferable components while suppressing the com- ponents not conducive to generalization. Further, we empir- ically compare the effectiveness of adopting different types of attention designs for implementing DFF . Extensive exper- iments demonstrate the effectiveness of our proposed DFF and show that applying our DFF on a plain baseline out- performs the state-of-the-art methods on different domain generalization tasks, including close-set classification and open-set retrieval.	1. Introduction Domain Generalization (DG) seeks to break through the i.i.d. assumption that training and testing data are identi- cally and independently distributed. This assumption does not always hold in reality since domain gaps are commonly seen between the training and testing data. However, col- lecting enough training data from all possible domains is costly and even impossible in some practical environments. Thus, learning generalizable feature representations is of *This work was done when Shiqi Lin and Zhipeng Huang were interns at Microsoft Research Asia.high practical value for both industry and academia. Recently, a series of research works [78] analyze deep learning from the frequency perspective. These works, represented by the F-Principle [75], uncover that there are different preference degrees of DNNs for the infor- mation of different frequencies in their learning processes. Specifically, DNNs optimized with stochastic gradient- based methods tend to capture low-frequency components of the training data with a higher priority [74] while exploit- ing high-frequency components to trade the robustness (on unseen domains) for the accuracy (on seen domains) [66]. This observation indicates that different frequency compo- nents are of different transferability across domains. In this work, we seek to learn generalizable features from a frequency perspective. To achieve this, we conceptual- ize Deep Frequency Filtering (DFF), which is a new tech- nique capable of enhancing the transferable frequency com- ponents and suppressing the ones not conducive to general- ization in the latent space. With DFF, the frequency compo- nents of different cross-domain transferability are dynam- ically modulated in an end-to-end manner during training. This is conceptually simple, easy to implement, yet remark- ably effective. In particular, for a given intermediate fea- ture, we apply Fast Fourier Transform (FFT) along its spa- tial dimensions to obtain the corresponding frequency rep- resentations where different spatial locations correspond to different frequency components. In such a frequency do- main, we are allowed to learn a spatial attention map and multiply it with the frequency representations to filter out the components adverse to the generalization across do- mains. The attention map above is learned in an end-to-end manner using a lightweight module, which is instance- adaptive. As indicated in [66, 74], low-frequency com- ponents are relatively easier to be generalized than high- frequency ones while high-frequency components are com- monly exploited to trade robustness for accuracy. Although this phenomenon can be observed consistently over differ- ent instances, it does not mean that high-frequency com- 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. 11797 ponents have the same proportion in different samples or have the same degree of effects on the generalization abil- ity. Thus, we experimentally compare the effectiveness of task-wise filtering with that of instance-adaptive filtering. Here, the task-wise filtering uses a shared mask over all in- stances while the instance-adaptive filtering uses unshared masks. We find the former one also works but is inferior to our proposed design by a clear margin. As analyzed in [10], the spectral transform theory [32] shows that updating a sin- gle value in the frequency domain globally affects all orig- inal data before FFT, rendering frequency representation as a global feature complementary to the local features learned through regular convolutions. Thus, a two-branch architec- ture named Fast Fourier Convolution (FFC) is introduced in [32] to exploit the complementarity of features in the fre- quency and original domains with an efficient ensemble. To evaluate the effectiveness of our proposed DFF, we choose this two-branch architecture as a base architecture and apply our proposed frequency filtering mechanism to its spectral transform branch. Note that FFC provides an effective im- plementation for frequency-space convolution while we in- troduce a novel frequency-space attention mechanism. We evaluate and demonstrate our effectiveness on top of it. Our contributions can be summarized in the following: • We discover that the cross-domain generalization ability of DNNs can be significantly enhanced by a simple learn- able filtering operation in the frequency domain. • We propose an effective Deep Frequency Filtering (DFF) module where we learn an instance-adaptive spatial mask to dynamically modulate different frequency components during training for learning generalizable features. • We conduct an empirical study for the comparison of dif- ferent design choices on implementing DFF, and find that the instance-level adaptability is required when learning frequency-space filtering for domain generalization.
Lee_Multimodal_Prompting_With_Missing_Modalities_for_Visual_Recognition_CVPR_2023	Abstract In this paper, we tackle two challenges in multimodal learning for visual recognition: 1) when missing-modality occurs either during training or testing in real-world sit- uations; and 2) when the computation resources are not available to finetune on heavy transformer models. To this end, we propose to utilize prompt learning and miti- gate the above two challenges together. Specifically, our modality-missing-aware prompts can be plugged into mul- timodal transformers to handle general missing-modality cases, while only requiring less than 1%learnable param- eters compared to training the entire model. We further ex- plore the effect of different prompt configurations and an- alyze the robustness to missing modality. Extensive experi- ments are conducted to show the effectiveness of our prompt learning framework that improves the performance under various missing-modality cases, while alleviating the re- quirement of heavy model re-training. Code is available.1	1. Introduction Our observation perceived in daily life is typically mulit- modal, such as visual, linguistic, and acoustic signals, thus modeling and coordinating multimodal information is of great interest and has broad application potentials. Re- cently, multimodal transformers [13, 17, 22, 25, 35] emerge as the pre-trained backbone models in several multimodal downstream tasks, including genre classification [22], mul- timodal sentiment analysis [25, 35], and cross-modal re- trieval [13,15,17,30], etc. Though providing promising per- formance and generalization ability on various tasks, there are still challenges for multimodal transformers being ap- plied in practical scenarios: 1) how to efficiently adapt the multimodal transformers without using heavy computation resource to finetune the entire model? 2) how to ensure the robustness when there are missing modalities, e.g., incom- plete training data or observations in testing? 1https://github.com/YiLunLee/missing aware prompts Figure 1. Illustration of missing-modality scenarios in training multimodal transformers. Prior work [22] investigates the robust- ness of multimodal transformers to modality-incomplete test data, with the requirement to finetune the entire model using modality- complete training data. In contrast, our work studies a more gen- eral scenario where various modality-missing cases would occur differently not only for each data sample but also learning phases (training, testing, or both), and we adopt prompt learning to adapt the pre-trained transformer for downstream tasks without requir- ing heavy computations on finetuning the entire model. Most multimodal transformer-based methods have a common assumption on the data completeness, which may not hold in practice due to the privacy, device, or security constraints. Thus, the performance may degrade when the data is modality-incomplete (regardless of training or test- ing). On the other hand, transformers pretrained on large- scale datasets are frequently adopted as backbone and fine- tuned for addressing various downstream tasks, thanks to the strong generalizability of transformers. However, as the model size of transformers increases (e.g., up to billions of parameters [5,26,27]), finetuning becomes significantly ex- pensive (e.g., up to millions of A100-GPU-hours [31]) and is even not feasible for practitioners due to the limited com- putation resources in most real-world applications. In addi- tion, finetuning a transformer on relatively small-scale tar- get datasets can result in restricted generalizability [9, 10] and stability [24], thus hindering it from being reused 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. 14943 further learning with new data or in other tasks. This motivates us to design a method that allows multi- modal transformers to alleviate these two real-world chal- lenges. One pioneer work [22] investigates the sensitiv- ity of vision-language transformers against the presence of modal-incomplete test data (i.e., either texts or images are missing). However, they only consider the case of miss- ing a specific modality for all the data samples, while in real-world scenarios the missing modality for each input data could not be known in advance. Moreover, [22] in- troduces additional task tokens to handle different missing- modal scenarios (e.g., text-only token when missing visual modality) and requires to optimize cross-modal features in the model. Hence finetuning the entire transformer becomes inevitable, leading to significant computation expense. In this paper, we study multimodal transformers under a more general modality-incomplete scenario, where var- ious missing-modality cases may occur in any data sam- ples, e.g., there can be both text-only and image-only data during training or testing. In particular, we also focus on alleviating the requirement of finetuning the entire trans- formers. To this end, we propose a framework stemmed from prompt learning techniques for addressing the afore- mentioned challenges. Basically, prompt learning meth- ods [2,5,8,16,18,32,42] emerge recently as efficient and ef- fective solutions for adapting pre-trained transformers to the target domain via only training very few parameters (i.e., prompts), and achieve comparable performance with fine- tuning the whole heavy model. As motivated by [29] which shows that prompts are good indicators for different distri- butions of input, we propose to regard different situations of missing modalities as different types of input and adopt the learnable prompts to mitigate the performance drop caused by missing modality. As a result, the size of our learnable prompts can be less than 1% of the entire transformer, and thus the computation becomes more affordable compared to holistic finetuning. The key differences between our work and [22] are illustrated in Figure 1. In order to further explore the prompt designs for multimodal transformers to tackle the general modality- incomplete scenario, we investigate two designs of inte- grating our missing-aware prompts2into pre-trained mul- timodal transformers: 1) input-level, and 2) attention-level prompt learning. We find that, the location of attaching prompts to transformers is crucial for the studied missing- modality cases in this paper, which also aligns the findings in [36], though under a different problem setting. We conduct experiments to explore different prompt configurations and have observations of the impact on the length and location of prompts: 1) As the number of prompting layers increases, the model performs better in- 2In this paper, we use “missing-aware prompts” and “modality- missing-aware prompts” interchangeably.tuitively but it is not the most important factor; 2) Attach- ing prompts to the layers near the data input achieves better performance; 3) The prompts’ length has slight impact on model performance for attention-level prompts but may in- fluence input-level prompts more on certain datasets. More- over, we show extensive results to validate the effective- ness of adopting our prompting framework to alleviate the missing-modality issue under various cases, while reducing the learnable parameters to less than 1%compared to the entire model. Our main contributions are as follows: • We introduce a general scenario for multimodal learn- ing, where the missing modality may occur differently for each data sample, either in training or testing phase. • We propose to use missing-aware prompts to tackle the missing modality situations, while only requiring less than 1% parameters to adapt pre-trained models, thus avoiding finetuning heavy transformers. • We further study two designs of attaching prompts onto different locations of a pretrained transformer, input-level and attention-level prompting, where the input-level prompting is generally a better choice but the attention-level one can be less sensitive to certain dataset settings.
Liu_RIATIG_Reliable_and_Imperceptible_Adversarial_Text-to-Image_Generation_With_Natural_Prompts_CVPR_2023	Abstract The field of text-to-image generation has made remark- able strides in creating high-fidelity and photorealistic im- ages. As this technology gains popularity, there is a grow- ing concern about its potential security risks. However, there has been limited exploration into the robustness of these models from an adversarial perspective. Existing re- search has primarily focused on untargeted settings, and lacks holistic consideration for reliability (attack success rate) and stealthiness (imperceptibility). In this paper, we propose RIATIG, a reliable and im- perceptible adversarial attack against text-to-image mod- els via inconspicuous examples. By formulating the exam- ple crafting as an optimization process and solving it using a genetic-based method, our proposed attack can generate imperceptible prompts for text-to-image generation models in a reliable way. Evaluation of six popular text-to-image generation models demonstrates the efficiency and stealthi- ness of our attack in both white-box and black-box settings. To allow the community to build on top of our findings, we’ve made the artifacts available1.	1. Introduction The text-to-image generation has captured widespread attention from the research community with its creative and realistic image generation capability [52, 55, 56]. The abil- ity to generate text-consistent images from natural language descriptions could potentially bring tremendous benefits to many areas of life, such as multimedia editing, computer- aided design, and art creation [23, 27, 37, 43]. Driven by recent advances in models trained with large datasets [42, 57] and multimodal learning [45] (e.g., diffu- sion models [17]), text-to-image generation has made sig- nificant progress in synthesizing high-fidelity and photore- alistic images, such as DALL ·E [43], DALL ·E 2 [42] and Imagen [45]. At the same time, there are a growing num- 1Code is available at: https://github.com/WUSTL-CSPL/RIATIG a cassavas that is on the side of a towe r a holsel is on the side of evergeren a tr ee standing next to Baruntse Target Images Adversarial Images DALL E mini . DALL E 2 . Imagen Figure 1. Examples of RIATIG attacks. The top texts are the adversarial prompts, and the bottom texts are the target models. The first row represents the target images, while the second row represents the adversarial image generated by the prompt. ber of ethical concerns about the potential misuse of this technology [36,45,53]. Generative models could be used to generate synthetic video/audio/images of individuals (e.g., Deepfakes [14]), or synthetic contents with harmful stereo- types, violence, or obscenities [10, 45, 53]. To prevent the generation of such harmful content, content moderation fil- ters are deployed in public APIs (e.g., DALL ·E 2) to filter unsafe text prompts that may lead to harmful content. How- ever, despite their best intentions, existing model-based text filters remain susceptible to adaptive adversarial attacks. Deep neural networks (DNNs) have been shown to be vulnerable to adversarial examples [9, 31, 32]. By applying these techniques, it is possible to craft an adversarial text that looks natural to bypass the content filters, yet gener- ates a completely different category of potentially malicious images. However, the adversarial attacks on text-to-image generators are less explored. To the best of our knowledge, there are two closely related studies [15, 36]. Nevertheless, two challenges remain: 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. 20585 1) Reliability. One significant limitation is that the existing works [15, 36] do not offer a reliable method to find ad- versarial examples. [15] discovers that DALL ·E 2 has cer- tain hidden vocabularies that can be used to generate images with some absurd (non-natural) prompts; however, this vo- cabulary is often limited and not stealthy (natural). Built on evocative prompting, [36] crafts adversarial examples via the morphological similarity between existing words. How- ever, it is very difficult to find texts with such linguistic sim- ilarity, and as a result, it can be challenging for this method to be adopted and generalized in different scenarios. 2) Stealthiness. The existing approaches can only craft ad- versarial examples that appear to be non-natural compared to normal texts or retain similar meanings, making them easily filtered and recognized by human examiners. For ex- ample, [15] crafts Apoploe vesrreaitais to represent bugs, and [36] crafts falaiscoglieklippantilado to represent cliff. Also, [36] combines creepy and spooky into creepooky to generate an image that looks creepy and scary, yet the in- ferred meaning of this new word is highly related to the generated image, limiting its stealthiness. To address these challenges, we propose RIATIG , a re- liable and imperceptible adversarial attack against text-to- image models using natural examples. To achieve this, we first formulate the generation of the adversarial examples as an optimization problem and apply genetic-based optimiza- tion methods to solve it, thus making our methods much more reliable in finding working adversarial examples. Fur- thermore, in order to improve the stealthiness, we propose a new text mutation technique to generate adversarial text that is visually and semantically similar to its normal ver- sion (some example results are shown in Figure 1). RIATIG is evaluated on six popular text-to-image mod- els with both white-box and black-box attack settings. Ex- perimental results show that compared with the state-of- the-art text-to-image-oriented adversarial attacks, RIATIG demonstrates significantly better performance in terms of attack effectiveness and sample quality. Overall, the contri- butions of this work are summarized as follows: • We are the first to systematically analyze the adver- sarial robustness of text-to-image generation models in both the white-box and black-box settings. • We propose genetic-based optimization methods to find natural adversarial examples reliably. • We evaluate our attacks on six popular text-to-image generation models and compare our attacks with five baselines. The evaluation results show that our meth- ods achieve a much higher success rate and sample quality, raising awareness of improving and securing the robustness of text-to-image models.
Koneputugodage_Octree_Guided_Unoriented_Surface_Reconstruction_CVPR_2023	Abstract We address the problem of surface reconstruction from unoriented point clouds. Implicit neural representations (INRs) have become popular for this task, but when infor- mation relating to the inside versus outside of a shape is not available (such as shape occupancy, signed distances or surface normal orientation) optimization relies on heuris- tics and regularizers to recover the surface. These meth- ods can be slow to converge and easily get stuck in local minima. We propose a two-step approach, OG-INR, where we (1) construct a discrete octree and label what is inside and outside (2) optimize for a continuous and high-fidelity shape using an INR that is initially guided by the octree’s labelling. To solve for our labelling, we propose an en- ergy function over the discrete structure and provide an ef- ficient move-making algorithm that explores many possible labellings. Furthermore we show that we can easily inject knowledge into the discrete octree, providing a simple way to influence the result from the continuous INR. We evaluate the effectiveness of our approach on two unoriented surface reconstruction datasets and show competitive performance compared to other unoriented, and some oriented, methods. Our results show that the exploration by the move-making algorithm avoids many of the bad local minima reached by purely gradient descent optimized methods (see Figure 1).	1. Introduction Surface reconstruction from 3D point clouds has been studied extensively in computer vision and computer graph- ics. The task requires estimating a mesh of the shape’s sur- face from a point cloud sampled from the surface of the shape. We focus on the reconstruction of watertight 3D shapes, i.e., shapes that have a well defined interior and ex- terior. Such shapes are often represented as signed distance fields (SDFs) or occupancy fields, which can be efficiently encoded by a neural network [28, 30]. As these representa- tions are fields parameterized by neural networks, they are often called neural fields. Furthermore, they implicitly rep- Figure 1. Our method with a SIREN INR (OG-SIREN, left) com- pared to SIREN wo n (right) for two shapes from the ShapeNet dataset. Our octree guidance allows for consistent inside-outside determinism. On the other hand, SIREN wo n gets stuck in lo- cal minima from which it cannot escape (due to needing to com- pletely change the occupancy of certain areas) creating extraneous surfaces (often called ghost geometries in the literature). resent the shape by a level set of the field, thus they are also often referred to as implicit neural representations (INRs). The broader class of neural fields, including INRs, have been very popular over the last few years as they can handle arbitrary topology, are memory efficient, and are continu- ous with potentially infinite resolution [39, 46]. Among the INRs for 3D shapes, SDFs are the most popular as they pro- vide more useful information (distances not just occupancy) and are required for downstream graphics algorithms such as sphere tracing and approximate soft shadows [29, 35]. When learning an implicit representation of a shape, a major difficulty is predicting whether points in space are inside or outside the shape. Many methods require ori- ented surface normals for the input points or signed dis- tances from the surface, which usually are not given by raw data from scans. While this can be estimated using the line of sight information [17] or algorithms [4, 6, 15, 21], they yield noisy predictions that after postprocessing still can lead to bad results (see Section 4.4). We consider the task of unoriented surface reconstruction, where such informa- tion is not available, and only the sampled surface points are given. We demonstrate that our method performs com- petitively and sometimes better than oriented methods, even when they are given the ground truth (GT) normals. 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. 16717 Figure 2. An illustration of our method, OG-INR. Given an unoriented point cloud (left), we progressively build and label an octree around the points (middle). The octree at depths 3-7 are shown. Surface leaves of the octree (yellow) are leaves that contain points from the point cloud, other leaves are labelled as inside (blue) or outside (transparent) by minimizing an energy function. We then train an INR model to obtain an SDF, using the labelling as supervision for the initial training, after which we can extract a mesh (right). We propose OG-INR, which uses a discrete representa- tion in conjunction with the continuous INR representation (see Figure 2). Given an input unoriented point cloud, we progressively build an octree from the input points and de- termine which leaf nodes are surface leaves ( i.e., leaves that contain a point). We also label all other leaves as inside or outside leaves (if they should be within or outside the shape respectively). To do this we minimize an energy function that trades off the watertight surface property, that every surface point should border with the inside and outside of the shape, with a minimal surface constraint. We then train standard INR architectures, initially guiding the training by the octree labelling, and show that the INR converges much faster and is less prone to large failure regions. Our main contributions are: • We introduce a novel method to guide the initial stage of INR training. It uses a labelled octree structure to allow the INR to converge significantly faster and alle- viates the local minima problem of INRs. • We propose an energy function over octree labels that captures the task of surface reconstruction. It balances the constraint of maintaining known surface regions with minimising the overall surface area. • We provide an efficient move-making algorithm to optimize the energy function, which explores many inside-outside possibilities in a structured manner. • Our discrete representation is easily understandable and human-modifiable, giving an intuitive method for applying changes to the resulting SDF.
Lei_RGBD2_Generative_Scene_Synthesis_via_Incremental_View_Inpainting_Using_RGBD_CVPR_2023	Abstract We address the challenge of recovering an underlying scene geometry and colors from a sparse set of RGBD view observations. In this work, we present a new solutiontermed RGBD 2that sequentially generates novel RGBD views along a camera trajectory, and the scene geometryis simply the fusion result of these views. More speciﬁcally, we maintain an intermediate surface mesh used for render-ing new RGBD views, which subsequently becomes com- plete by an inpainting network; each rendered RGBD view is later back-projected as a partial surface and is supple- mented into the intermediate mesh. The use of intermediate mesh and camera projection helps solve the tough problem of multi-view inconsistency. We practically implement the RGBD inpainting network as a versatile RGBD diffusionmodel, which is previously used for 2D generative model- ing; we make a modiﬁcation to its reverse diffusion processto enable our use. We evaluate our approach on the task of 3D scene synthesis from sparse RGBD inputs; extensiveexperiments on the ScanNet dataset demonstrate the supe- riority of our approach over existing ones. Project page: https://jblei.site/proj/rgbd-diffusion .	1. Introduction Scene synthesis is an essential requirement for many practical applications. The resulting scene representationcan be readily utilized in diverse ﬁelds, such as virtual re- ality, augmented reality, computer graphics, and game de- velopment. Nevertheless, conventional approaches to scenesynthesis usually involve reconstructing scenes (e.g., indoor scenes with varying sizes) by ﬁtting given observations, such as multi-view images or point clouds. The increas-ing prevalence of RGB/RGBD scanning devices has estab- lished multi-view data as a favored input modality, driving and promoting technical advancements in the realm of scenereconstruction from multi-view images. Neural Radiance Fields (NeRFs) [ 42] have demonstrated †Correspondence to Kui Jia: <[email protected] >. ݐൌͲ ݐൌͳ ݐൌʹ ݐൌ͵ Final Result… Figure 1. Illustration of Our Generative Scene Synthesis. We incrementally reconstruct the scene geometry by inpainting RGBD views as the camera moves in the scene. potential in this regard, yet they are not exempt from limi- tations. NeRFs are designed to reconstruct complete scenesby ﬁtting multi-view images, and they cannot generate or infer missing parts when the input is inevitably incomplete or missing. While recently some studies [ 3,5,8,56,63]h a v e attempted to equip NeRFs with generative and extrapola-tion capabilities, this functionality relies on a comparativelyshort representation with limited elements (e.g. typically,the length of a global latent code is much shorter than that of an image: (F= 512)/lessmuch(H×W= 128×128 = 16 ,384) ) that signiﬁcantly constrains their capacity to accurately cap-ture ﬁne-grained details in the observed data. Consequently, the effectiveness of these methods has only been established for certain categories of canonical objects, such as faces or cars [ 8,63], or relatively small toy scenes [ 5]. We introduce a novel task of generative scene synthesis from sparse RGBD views, which involves learning across multiple scenes to later enable scene synthesis from a sparse set of multi-view RGBD images. This task presents a chal- lenging setting wherein a desired solution should simulta-neously (1) preserve observed regions, hallucinate missingparts of the scene, (2) eliminate additional computational costs during inference for each individual test scene, (3) en- 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. 8422 sure exact 3D consistency, and (4) maintain scalability to scenes with unﬁxed scales. We will elaborate on them in detail as follows. Firstly, to maximize the preservation of intricate details while simulta-neously hallucinating potentially absent parts that may be-come more pronounced when views are exceedingly sparse,we perform straightforward reconstruction whose detailscome from images that can describe ﬁne structures using a maximum of H×Welements (i.e. an image size) in a view completion manner. This is particularly compatible with diffusion models that operate at full image resolutionwith an inpainting mechanism. We also found that RGBDdiffusion models greatly simplify the training complexityof a completion model, thanks to their versatile generativeability to inpaint missing RGBD pixels while preserving the integrity of known regions through a convenient train-ing process solely operated on complete RGBD data. Sec- ondly, our method employs back-projection that requiresno optimization, thus eliminating the necessity for test-time training for each individual scene, ultimately leading to a signiﬁcant enhancement in test-time efﬁciency. Thirdly, to ensure consistency among multi-view images, an interme- diate mesh representation is utilized as a means of bridging the 2D domain (i.e. multi-view RGBD images) with the 3D domain (i.e. the 3D intermediate mesh) through the aid ofcamera projection. Fourthly, to enable our method to handle scenes of indeterminate sizes, we utilize images with freelydesignated poses as the input representation. Such manner naturally ensures SE(3) equivariance, and thus offers scala- bility due to the ease with which the range of the generated content can be controlled by simply specifying their cameraextrinsic matrices. Our proposal involves generating multi-view consistent RGBD views along a predetermined camera trajectory, us- ing an intermediate mesh to render novel RGBD images that are subsequently inpainted using a diffusion model, and transforming each RGBD view into a 3D partial meshvia back-projection, and ﬁnally merging it with the inter- mediate scene mesh to produce the ﬁnal output. Speciﬁ-cally, our proposed approach initiates by ingesting multipleposed RGBD images as input and utilizing back-projectionto construct an intermediate scene mesh. This mesh encom- passes color attributes that facilitate the rendering of RGBD images from the representation under arbitrarily speciﬁed camera viewpoints. Once a camera pose is selected from the test-time rendering trajectory, the intermediate mesh isrendered to generate a new RGBD image for this pose. No- tably, the test-time view typically exhibits only slight over-lap with the known cameras, leading to naturally partially rendered RGBD images. To ﬁll the gaps in the incom- plete view, we employ an inpainting network implemented as an RGBD diffusion model with minor modiﬁcations toits reverse sampling process. The resulting inpainted out-put is then back-projected into 3D space, forming a partial mesh that complements the entire intermediate scene mesh. We iterate these steps until all test-time camera viewpointsare covered, and the intermediate scene mesh gradually be-comes complete during this process. The ﬁnal output of ourpipeline is the mesh outcome acquired from the last step. Extensive experiments on ScanNet [ 12] dataset demon- strate the superiority of our approach over existing solutionson the task of scene synthesis from sparse RGBD inputs.
Lin_Harmonious_Feature_Learning_for_Interactive_Hand-Object_Pose_Estimation_CVPR_2023	Abstract Joint hand and object pose estimation from a single image is extremely challenging as serious occlusion often occurs when the hand and object interact. Existing ap- proaches typically first extract coarse hand and object fea- tures from a single backbone, then further enhance them with reference to each other via interaction modules. How- ever, these works usually ignore that the hand and ob- ject are competitive in feature learning, since the backbone takes both of them as foreground and they are usually mu- tually occluded. In this paper, we propose a novel Har- monious Feature Learning Network (HFL-Net). HFL-Net introduces a new framework that combines the advantages of single- and double-stream backbones: it shares the pa- rameters of the low- and high-level convolutional layers of a common ResNet-50 model for the hand and object, leav- ing the middle-level layers unshared. This strategy enables the hand and the object to be extracted as the sole targets by the middle-level layers, avoiding their competition in feature learning. The shared high-level layers also force their features to be harmonious, thereby facilitating their mutual feature enhancement. In particular, we propose to enhance the feature of the hand via concatenation with the feature in the same location from the object stream. A sub- sequent self-attention layer is adopted to deeply fuse the concatenated feature. Experimental results show that our proposed approach consistently outperforms state-of-the- art methods on the popular HO3D and Dex-YCB databases. Notably, the performance of our model on hand pose esti- mation even surpasses that of existing works that only per- form the single-hand pose estimation task. Code is avail- able at https://github.com/lzfff12/HFL-Net.	1. Introduction When humans interact with the physical world, they pri- marily do so by using their hands. Thus, an accurate un- derstanding of how hands interact with objects is essen- *Corresponding author. Image Front View Other View Figure 1. HFL-Net predicts the 3D hand and object poses from single monocular RGB images accurately, even in serious occlu- sion scenarios. tial to the understanding of human behavior. It can be widely applied to a range of fields, including the devel- opment of virtual reality [36], augmented reality [33, 34], and imitation-based robot learning [35], among others. Re- cently, hand pose estimation [12–16] and 6D object pose estimation [17–19] based on monocular RGB images have respectively achieved remarkable results. However, the re- search into joint hand-object pose estimation under circum- stances of interaction remains in its infancy [2,3,23,26–28]. As illustrated in Figure 1, joint hand-object pose esti- mation from a single image is extremely challenging. The main reason for this is that when the hand and object inter- act with each other, serious occlusion occurs; occlusion, in turn, results in information loss, increasing the difficulty of each task. One mainstream solution to this problem is to utilize context. Due to physical constraints, the interacting hand and object tend to be highly correlated in terms of their poses, meaning that the appearance of one can be useful context for the other [1–3]. Methods that adopt this solu- tion typically employ a single backbone to extract features for the hand and object, respectively [2, 22, 27]. This uni- 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. 12989 fied backbone model ensures that the hand and object fea- tures are in the same space, which facilitates the subsequent mutual feature enhancement between hand and object via attention-based methods [2]. However, the hand and object pose estimation tasks are competitive in feature learning if a single backbone model is utilized. In more detail, when the hand and object are close to each other, the backbone model treats them both as foreground, and may thus be unable to differentiate the hand features from those of the object. A straightforward solution is to utilize two backbones [1, 3, 23], one for the hand and the other one for the object; when this approach is adopted, each backbone has only one target as the foreground. The main downsides of this strategy include large model size and (more importantly) the different feature spaces between backbones, which introduce difficulties with regard to mu- tual feature enhancement between the hand and object. To solve the aforementioned problems, we propose a novel Harmonious Feature Learning Network (HFL-Net). HFL-Net introduces a new framework that combines the advantages of single- and double-stream backbones. In more detail, our backbone shares the parameters of the low- and high-level convolutional layers of a common ResNet-50 model [4] for the hand and object, leaving the middle-level layers unshared. Feature maps produced by low-level layers are fed into the two sets of middle-level layers, which regard the hand and object respectively as the sole foreground tar- get. As a result, feature learning for the hand and object is no longer competitive. Finally, through sharing the param- eters of the high-level convolutional layers, the hand and object features are forced to be in similar feature spaces. In this way, our backbone realizes harmonious feature learning for the hand and object pose estimation. We further enhance the representation power of the hand and object features through the use of efficient attention models. Several existing methods have successfully real- ized hand-to-object feature enhancement via cross-attention operations [1, 2]; however, object-to-hand feature enhance- ment usually turns out to be difficult [1, 2]. Motivated by the observation that when one pixel on the hand is occluded, the object feature in the same location usually provides use- ful cues, we propose a simple but effective strategy for fa- cilitating object-to-hand feature enhancement. Specifically, we adopt ROIAlign [6] to extract fixed-size feature maps from the two output streams of our backbone respectively according to the hand bounding box. We then concatenate the two feature maps along the channel dimension and feed the obtained feature maps into a self-attention module [7]. Object-to-hand feature enhancement is automatically real- ized via the fully-connected and multi-head attention layers in the self-attention module. Finally, we split the output feature maps by the self-attention layer along the channel dimension, and take the first half as the enhanced hand fea-ture maps. We demonstrate the effectiveness of HFL-Net through comprehensive experiments on two benchmarks: HO3D [9] and Dex-YCB [10], and find that our method consistently outperforms state-of-the-art works on the joint hand-object pose estimation task. Moreover, benefiting from the learned harmonious hand and object features, the hand and object pose estimation tasks in HFL-Net are mutually beneficial rather than competitive. In our experiments, the perfor- mance of HFL-Net on the hand pose estimation task sur- passes even recent works [12,15,32] that only estimate hand poses in both the training and testing stages.
Kan_Self-Correctable_and_Adaptable_Inference_for_Generalizable_Human_Pose_Estimation_CVPR_2023	Abstract A central challenge in human pose estimation, as well as in many other machine learning and prediction tasks, is the generalization problem. The learned network does not have the capability to characterize the prediction error, gen- erate feedback information from the test sample, and cor- rect the prediction error on the fly for each individual test sample, which results in degraded performance in general- ization. In this work, we introduce a self-correctable and adaptable inference (SCAI) method to address the general- ization challenge of network prediction and use human pose estimation as an example to demonstrate its effectiveness and performance. We learn a correction network to correct the prediction result conditioned by a fitness feedback er- ror. This feedback error is generated by a learned fitness feedback network which maps the prediction result to the original input domain and compares it against the original input. Interestingly, we find that this self-referential feed- back error is highly correlated with the actual prediction error. This strong correlation suggests that we can use this error as feedback to guide the correction process. It can be also used as a loss function to quickly adapt and opti- mize the correction network during the inference process. Our extensive experimental results on human pose estima- tion demonstrate that the proposed SCAI method is able to significantly improve the generalization capability and per- formance of human pose estimation.	1. Introduction Human pose estimation (HPE) aims to correctly predict and localize human body joints. A variety of downstream applications are based on human pose estimation, such as motion capture [7, 27], activity recognition [1, 6, 37], per- son tracking [36, 41] and video surveillance [18]. Recently, deep learning-based methods for human pose estimation *Corresponding author.have achieved remarkable success [2,4,12,26,28,30]. How- ever, in complex or unseen scenarios, pose estimation re- mains very challenging due to occlusions, cluttered back- ground, and large variations of appearance and scenes, es- pecially for those distal keypoints at the end locations of body parts, such as wrists and ankles, which have large de- grees of motion freedom and often suffer from severe oc- clusions [13, 39]. We recognize that one major challenge in current human pose estimation, as well as in many other prediction tasks, is generalization. Network models, which have been well learned on the training set, often experience significant per- formance degradation on the test samples which are col- lected from different environments or scenarios. For exam- ple, in human pose estimation, there are different types of occlusions of body parts due to complex scene structures and free-style motions of human bodies. More importantly, the occlusion scenarios of the test samples could be much different from those in the training samples. This often leads to the significant performance degradation of human pose estimation from the training data to the test data. For example, in our experiment, the average prediction accuracy on the training samples is 95.5%. However, on the test set, this accuracy drops to 67%. For those distal keypoints at tip locations of body parts which often experience more signif- icant occlusions, their average performance drop is much more significant, from 95.3% to 57%. To address this performance degradation or generaliza- tion problem, there are two major questions that need to be carefully answered: (1) how can we tell if the prediction is accurate or not during testing and how to characterize the prediction error? This is difficult because the ground truth values of the test samples are not available during test- ing. Specifically, in pose estimation, we do not have the labeled ground truth locations of the body keypoints. (2) How to correct the prediction error based on the specific characteristics of the test sample? Current network models, once successfully trained with labeled samples at the train- ing side, remain fixed during testing, performing the feed- 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. 5537 forward-only inference process to generate the prediction result. There is no mechanism for us to examine the spe- cific characteristics of the test sample and use them as feed- back to correct the prediction error or adjust the network model. We believe that this unique capability of sample- specific prediction error characterization, error correction, and model optimization is very important for the general- ization performance of learned network models. It also has the potential to significantly improve the prediction accu- racy of test samples. To address these two challenging issues, in this work, we propose to explore a learning-based feedback-control or correction method for prediction, with applications to hu- man pose prediction. Specifically, let ˆ v=Φ(u)be the prediction network which is tasked to predict the true value ofvfrom input u. To answer the first question, we design and learn a fitness feedback network Γwhich compares the prediction result ˆ v=Φ(u)of the prediction network Φ against the original input uand generate a self-referential feedback error. Very interestingly, in this work, we find that this self-referential feedback error is highly correlated with the prediction error of the network Φ. Note that, when com- puting the self-referential error, we do not need the ground truth data. It can be directly computed on the input sample using the prediction-feedback networks. This allows us to characterize the prediction error of test samples. Under the guidance of self-referential error feedback, we train a prediction error correction network Cto adjust the inference results during the prediction process to improve the prediction accuracy for the test samples. Besides, we find that the self-referential error and the fitness feedback network (FFN) can be used to construct a self-referential loss function on the test samples to quickly adapt and opti- mize the network model during the inference stage, making the model learnable on the test side. We apply the above self-correctable and adaptable inference (SCAI) method to human pose estimation. Our extensive experimental results on benchmark datasets demonstrate that the proposed SCAI method is able to significantly improve the generalization capability of the underlying prediction algorithm. It out- performs the existing state-of-the-art methods on human pose estimation by large margins. For example, on the MS COCO-testdev dataset, our method improves upon the cur- rent best method by up to 1.4%, which is quite significant.
Kim_PartMix_Regularization_Strategy_To_Learn_Part_Discovery_for_Visible-Infrared_Person_CVPR_2023	Abstract Modern data augmentation using a mixture-based tech- nique can regularize the models from overfitting to the training data in various computer vision applications, but a proper data augmentation technique tailored for the part-based Visible-Infrared person Re-IDentification (VI-ReID) models remains unexplored. In this paper, we present a novel data augmentation technique, dubbed PartMix , that synthesizes the augmented samples by mixing the part descriptors across the modalities to improve the performance of part-based VI-ReID models. Especially, we synthesize the positive and negative samples within the same and across different identities and regularize the backbone model through contrastive learning. In addition, we also present an entropy-based mining strat- egy to weaken the adverse impact of unreliable positive and negative samples. When incorporated into existing part-based VI-ReID model, PartMix consistently boosts the performance. We conduct experiments to demonstrate the effectiveness of our PartMix over the existing VI-ReID methods and provide ablation studies.	1. Introduction Person Re-IDentification (ReID), aiming to match per- son images in a query set to ones in a gallery set cap- tured by non-overlapping cameras, has recently received substantial attention in numerous computer vision applica- tions, including video surveillance, security, and persons analysis [64, 76]. Many ReID approaches [1, 2, 25, 26, 32, 41, 59, 73, 78] formulate the task as a visible-modality re- trieval problem, which may fail to achieve satisfactory re- sults under poor illumination conditions. To address this, most surveillance systems use an infrared camera that can capture the scene even in low-light conditions. However, *Corresponding author This research was supported by the National Research Founda- tion of Korea (NRF) grant funded by the Korea government (MSIP) (NRF2021R1A2C2006703). Vis -ID 1 Vis -ID 2Mix Mix Mix Inf-ID 2 Vis -ID 1Mix Vis -ID 1 Vis -ID 2Mix Mix Inf-ID 2 Vis -ID 1Mix Mix Inf-ID 2 Vis -ID 1Mix Vis -ID 1 Vis -ID 2Mix (a) MixUp [69] Vis -ID 1 Vis -ID 2Mix Mix Mix Inf-ID 2 Vis -ID 1Mix Vis -ID 1 Vis -ID 2Mix Mix Inf-ID 2 Vis -ID 1Mix Mix Inf-ID 2 Vis -ID 1Mix Vis -ID 1 Vis -ID 2Mix (b) CutMix [68] Vis -ID 1 Vis -ID 2Mix Mix Mix Inf-ID 2 Vis -ID 1Mix Vis -ID 1 Vis -ID 2Mix Mix Inf-ID 2 Vis -ID 1Mix Mix Inf-ID 2 Vis -ID 1Mix Vis -ID 1 Vis -ID 2Mix (c) PartMix (Ours) Figure 1. Comparison of data augmentation methods for VI- ReID. (a) MixUp [69] using a global image mixture and (b) Cut- Mix [68] using a local image mixture can be used to regularize a model for VI-ReID, but these methods provide limited perfor- mances because they yield unnatural patterns or local patches with only background or single human part. Unlike them, we present (c) PartMix using a part descriptor mixing strategy, which boosts the VI-ReID performance (Best viewed in color). directly matching these infrared images to visible ones for ReID poses additional challenges due to an inter-modality variation [60, 62, 65]. To alleviate these inherent challenges, Visible-Infrared person Re-IDentification (VI-ReID) [5–7, 12, 24, 35, 48, 54, 60–62, 66] has been popularly proposed to handle the large intra- and inter-modality variations between visible images and their infrared counterparts. Formally, these ap- proaches first extract a person representation from whole 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. 18621 visible and infrared images, respectively, and then learn a modality-invariant feature representation using feature alignment techniques, e.g., triplet [5,7,24,60–62] or ranking criterion [12, 66], so as to remove the inter-modality varia- tion. However, these global feature representations solely focus on the most discriminative part while ignoring the di- verse parts which are helpful to distinguish the person iden- tity [53, 63]. Recent approaches [53, 57, 63] attempted to further en- hance the discriminative power of person representation for VI-ReID by capturing diverse human body parts across dif- ferent modalities. Typically, they first capture several hu- man parts through, e.g., horizontal stripes [63], cluster- ing [53], or attention mechanisms [57] from both visible and infrared images, extract the features from these human parts, and then reduce inter-modality variation in a part- level feature representation. Although these methods re- duce inter-modality variation through the final prediction (e.g., identity probability), learning such part detector still leads to overfitting to the specific part because the model mainly focuses on the most discriminative part to classify the identity, as demonstrated in [4, 15, 31, 52]. In addition, these parts are different depending on the modality, it accu- mulates errors in the subsequent inter-modality alignment process, which hinders the generalization ability on unseen identity in test set. On the other hand, many data augmentation [19, 22, 38, 45, 68, 69] enlarge the training set through the image mix- ture technique [69]. They typically exploit the samples that linearly interpolate the global [38, 44, 69] or local [19, 68] images and label pairs for training, allowing the model to have smoother decision boundaries that reduce overfit- ting to the training samples. This framework also can be a promising solution to reduce inter-modality variation by mixing the different modality samples to mitigate overfit- ting to the specific modality, but directly applying these techniques to part-based VI-ReID models is challenging in that they inherit the limitation of global and local image mixture methods ( e.g., ambiguous and unnatural patterns, and local patches with only background or single human part). Therefore, the performance of part-based VI-ReID with these existing augmentations would be degraded. In this paper, we propose a novel data augmentation tech- nique for VI-ReID task, called PartMix , that synthesizes the part-aware augmented samples by mixing the part de- scriptors. Based on the observation that learning with the unseen combination of human parts may help better regu- larize the VI-ReID model, we randomly mix the inter- and intra-modality part descriptors to generate positive and neg- ative samples within the same and across different identi- ties, and regularize the model through the contrastive learn- ing. In addition, we also present an entropy-based mining strategy to weaken the adverse impact of unreliable posi-tive and negative samples. We demonstrate the effective- ness of our method on several benchmarks [33,55]. We also provide an extensive ablation study to validate and analyze components in our model.
Lee_BAAM_Monocular_3D_Pose_and_Shape_Reconstruction_With_Bi-Contextual_Attention_CVPR_2023	Abstract 3D trafﬁc scene comprises various 3D information about car objects, including their pose and shape. However, most recent studies pay relatively less attention to recon- structing detailed shapes. Furthermore, most of them treat each 3D object as an independent one, resulting in losses of relative context inter-objects and scene context reﬂect- ing road circumstances. A novel monocular 3D pose and shape reconstruction algorithm, based on bi-contextual at- tention and attention-guided modeling (BAAM), is proposed in this work. First, given 2D primitives, we reconstruct 3D object shape based on attention-guided modeling that considers the relevance between detected objects and ve- hicle shape priors. Next, we estimate 3D object pose through bi-contextual attention, which leverages relation- context inter objects and scene-context between an object and road environment. Finally, we propose a 3D non- maximum suppression algorithm to eliminate spurious ob- jects based on their Bird-Eye-View distance. Extensive experiments demonstrate that the proposed BAAM yields state-of-the-art performance on ApolloCar3D. Also, they show that the proposed BAAM can be plugged into any mature monocular 3D object detector on KITTI and sig- niﬁcantly boost their performance. Code is available at https://github.com/gywns6287/BAAM.	1. Introduction 3D trafﬁc scene understanding provides enriched de- scriptions of the dynamic objects, e.g., 3D shape, pose, and location, compared to representing objects as bounding boxes. 3D visual perception is crucial for the autonomous driving system to develop downstream tasks such as mo- tion prediction and planning, and aids to faithfully recon- Corresponding author yThese authors contributed equally to this work. 3D Scene Input image Figure 1. Reconstructed 3D scene with rough bounding box (right up) and with detailed shape (right down). For better 3D recon- struction, detailed 3D shapes are needed rather than the simple 3D bounding boxes. struct the trafﬁc scene from recorded data. To acquire pre- cise 3D information, some prior arts have relied on speciﬁc devices such as LiDAR [3,10,42] and stereo vision [26,44]. However, as the system becomes complex and costly, it quickly reaches the limit to scalability. To contrary, areas of study about 3D perception using monocular vision have been receiving attention due to its simplicity and cost efﬁ- ciency [4, 7, 19, 29, 33, 50, 51]. Monocular 3D perception is an ill-posed problem in that projective geometry inherently loses depth information. In particular, trafﬁc scene contains partially observable ob- jects, and shows ﬁne-grained classes which are visually confusing. Pseudo-LiDAR [49] presents a feasible solution of the image based 3D object detection. To reconstruct 3D poses of the objects, many studies [25, 27, 30, 33, 40, 41, 50, 51] focus on using geometry constraints between 2D and 3D. Yet, it is less studied in the line of research that leverage relative context among the objects and global scene context depending on road environment. Figure 1 compares the reconstructed 3D scene with 3D bounding boxes and detailed 3D shapes. With a detailed 3D shape, we render the trafﬁc scene in realistic and provide intuitive representations of the objects. Despite scale ambi- guity of the monocular 3D perception, 3D mesh provides a 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. 9011 strong clue to align instances’ scales and orientations. Con- currently, there have been many attempts [8, 20, 22, 31, 45, 46] to reconstruct the 3D shape of human objects. These methods mainly focus on learning PCA-basis to represent human shapes. Inspired by human shape reconstruction, recent methods [21, 24] also design PCA-basis for vehicle shape reconstruction. However, as pointed out in [1, 34], PCA-basis often loses object details and thus leads to un- satisfactory reconstruction. In this work, we propose a novel 3D pose and shape estimation algorithm, utilizing bi-contextual attention and attention-guided modeling (BAAM). Given a monocular RGB image, the proposed BAAM ﬁrst extracts various 2D primitive features such as appearance, position, and size. And it constructs object features to embed internal object structures by aggregating primitive features. For detailed object shapes, we introduce shape priors consisting of the mean shape and various template offsets to represent de- tails of vehicle shapes. Then, BAAM reconstructs objects’ 3D shapes as mesh structures with attention-guided mod- eling, which combines shape prior and individual object features based on their relevance. For accurate pose es- timation, we present the notion of bi-contextual attention consisting of relation-context and scene-context, which de- scribe the relationship inter objects and between object and road environment, respectively. Based on this rich infor- mation, BAAM integrates object features to predict ob- jects’ 3D poses through a carefully designed bi-contextual attention module. Finally, we proposed a novel 3D non- maximum suppression (NMS) algorithm that effectually re- moves spurious objects based on Bird-Eye-view (BEV) ge- ometry. Extensive experiments on Apollocar3D [43] and KITTI [12] datasets demonstrate the effectiveness of the proposed BAAM algorithm. Also, experiments show that the proposed method signiﬁcantly outperforms state-of-the arts [21, 43] in both pose and shape estimation. The main contributions of our work are four folds: We propose the attention-guided modeling that recon- structs objects’ shapes based on the relevance between objects and vehicle shape priors. We proposed the bi-contextual attention module that estimates objects’ pose by exploiting relation-context inter objects and scene-context between an object and road environment. We also develop the novel 3D non-maximum suppres- sion algorithm to remove spurious objects based on their Bird-Eye-view distance. The proposed BAAM algorithm achieves the state-of the art performance on ApolloCar3D [43]. Also, ex- periments on KITTI [12] show that the proposed al- gorithm can signiﬁcantly improve the performance of existing monocular 3D detectors.
Li_Diffusion-SDF_Text-To-Shape_via_Voxelized_Diffusion_CVPR_2023	Abstract With the rising industrial attention to 3D virtual mod- eling technology, generating novel 3D content based on specified conditions ( e.g. text) has become a hot issue. In this paper, we propose a new generative 3D model- ing framework called Diffusion-SDF for the challenging task of text-to-shape synthesis. Previous approaches lack flexibility in both 3D data representation and shape gen- eration, thereby failing to generate highly diversified 3D shapes conforming to the given text descriptions. To ad- dress this, we propose a SDF autoencoder together with the Voxelized Diffusion model to learn and generate rep- resentations for voxelized signed distance fields (SDFs) of 3D shapes. Specifically, we design a novel UinU-Net ar- chitecture that implants a local-focused inner network in- side the standard U-Net architecture, which enables better reconstruction of patch-independent SDF representations. We extend our approach to further text-to-shape tasks in- cluding text-conditioned shape completion and manipula- tion. Experimental results show that Diffusion-SDF gen- erates both higher quality and more diversified 3D shapes that conform well to given text descriptions when compared to previous approaches. Code is available at: https: //github.com/ttlmh/Diffusion-SDF . †Corresponding author.	1. Introduction Exploring data representations for 3D shapes has been a fundamental and critical issue in 3D computer vision. Ex- plicit 3D representations including point clouds [38, 39], polygon meshes [17, 24] and occupancy voxel grids [9, 53] have been widely applied in various 3D downstream ap- plications [1, 35, 56]. While explicit 3D representations achieve encouraging performance, there are some primary limitations including not being suitable for generating wa- tertight surfaces ( e.g. point clouds), or being subject to topo- logical constraints ( e.g. meshes). On the other hand, im- plicit 3D representations have been widely studied more recently [3, 14, 37], with representative works including DeepSDF [37], Occupancy Network [32] and IM-Net [8]. In general, implicit functions encode the shapes by the iso- surface of the function, which is a continuous field and can be evaluated at arbitrary resolution. In recent years, numerous explorations have been con- ducted for implicit 3D generative models, which show promising performance on several downstream applications such as single/multi-view 3D reconstruction [23, 54] and shape completion [12, 34]. Besides, several studies have also explored the feasibility of directly generating novel 3D shapes based on implicit representations [15, 21]. How- ever, these approaches are incapable of generating specified 3D shapes that match a given condition, e.g. a short text describing the shape characteristics as shown in Figure 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. 12642 Text-based visual content synthesis has the advantages of the flexibility and generality [41, 42]. Users may generate rich and diverse 3D shapes based on easily obtained natural language descriptors. In addition to generating 3D shapes directly based on text descriptions, manipulating 3D data with text guidance can be further utilized for iterative 3D synthesis and fine-grained 3D editing, which can be benefi- cial for non-expert users to create 3D visual content. In the literature, there have been few attempts on the challenging task of text-to-shape generation based on im- plicit 3D representations [29, 34, 48]. For example, Au- toSDF [34] introduced a vector quantized SDF autoen- coder together with an autoregressive generator for shape generation. While encouraging progress has been made, the quality and diversity of generated shapes still requires improvement. The current approaches struggle to gener- ate highly diversified 3D shapes that both guarantee gen- eration quality and conform to the semantics of the input text. Motivated by the success of denoising diffusion mod- els in 2D image [13, 19, 36] and even explicit 3D point cloud [31, 58, 59] generation, we find that DMs achieve high-quality and highly diversified generation while being robust to model training. To this end, we aim to design an implicit 3D representation-based generative diffusion pro- cess for text-to-shape synthesis that can achieve better gen- eration flexibility and generalization performance. In this paper, we propose the Diffusion-SDF framework for text-to-shape synthesis based on truncated signed dis- tance fields (TSDFs). Considering that 3D shapes share structural similarities at local scales, and the cubic data vol- ume of 3D voxels may lead to slow sampling speed for dif- fusion models, we propose a two-stage separated generation pipeline. First, we introduce a patch-based SDF autoen- coder that map the original signed distance fields into patch- independent local Gaussian latent representations. Sec- ond, we introduce the Voxelized Diffusion model that cap- tures the intra-patch information along with both patch-to- patch and patch-to-global relations. Specifically, we de- sign a novel UinU -Net architecture to replace the standard U-Net [46] for the noise estimator in the reverse process. UinU -Net implants a local-focused inner network inside the outer U-Net backbone, which takes into account the patch-independent prior of SDF representations to better re- construct local patch features from noise. Our work digs deeper into the further potential of diffusion model-based approaches towards text-conditioned 3D shape synthesis based on voxelized TSDFs. Experiments on the largest ex- isting text-shape dataset [6] show that our Diffusion-SDF approach achieves promising generation performance on text-to-shape tasks compared to existing state-of-the-art ap- proaches, in both qualitative and quantitative evaluations. Strictly speaking, our approach employs a combined explicit-implicit representation in the form of voxelized signed distance fields.
Liu_Diversity-Measurable_Anomaly_Detection_CVPR_2023	Abstract Reconstruction-based anomaly detection models achieve their purpose by suppressing the generalization ability for anomaly. However, diverse normal patterns are conse- quently not well reconstructed as well. Although some ef- forts have been made to alleviate this problem by modeling sample diversity, they suffer from shortcut learning due to undesired transmission of abnormal information. In this paper, to better handle the tradeoff problem, we propose Diversity-Measurable Anomaly Detection (DMAD) frame- work to enhance reconstruction diversity while avoid the undesired generalization on anomalies. To this end, we design Pyramid Deformation Module (PDM) , which mod- els diverse normals and measures the severity of anomaly by estimating multi-scale deformation fields from recon- structed reference to original input. Integrated with an in- formation compression module, PDM essentially decouples deformation from prototypical embedding and makes the fi- nal anomaly score more reliable. Experimental results on both surveillance videos and industrial images demonstrate the effectiveness of our method. In addition, DMAD works equally well in front of contaminated data and anomaly-like normal samples.	1. Introduction Visual anomaly detection is a fundamental and important problem in computer vision community, with wide applica- tions in video surveillance and industrial inspection. It aims to detect outliers from seen classes and novel patterns from unseen classes. This task is very challenging because abnor- mal data is diversely distributed and expensive to collect. So we have to construct models based on only normal samples under unsupervised setting, targeting at high discrimination between normal and abnormal samples. During the past decade, reconstruction-based methods have achieved great progress in anomaly detection. These methods use Autoencoders (AEs) [8, 9, 20, 24, 26, 29, 40] or Generative Adversarial Networks (GANs) [1, 19, 32] to re- Figure 1. Illustration of difficulty in anomaly detection in MNIST dataset. The prototype is indicated by orange triangle and the anomaly by red point. In this case, the anomaly can hardly be detected based on reconstruction error or distance in high- dimensional feature space. Our solution is illustrated in Fig. 2. construct the normal counterparts from any input images or video frames. AE-based methods firstly compress the in- puts to discard the information beyond normal prototypes, and then decode the embedding to reconstruct the inputs. According to the estimated reconstruction error, the anoma- lies can be detected. However, the performance of reconstruction-based methods for anomaly detection has long been limited by a tough problem, i.e.the tradeoff between reconstructing di- verse normals and detecting unknown anomalies . In order to discriminate anomalies more easily, previous works [8] imposes more constraints to suppress abnormal information during autoencoding, which leads to high reconstruction er- ror for diverse normal instances. For example, in Figs. 1 and 2 g, the severely deformed normal (a.k.a. anomaly-like) sample “7” has even higher error than the abnormal sam- ple “4”. To better reconstruct diverse normals, each query vector correspond to multiple prototypes in the memory, which may be combined into abnormal embedding even if abnormal projection is far away from the prototype. As a consequence, anomalies that distribute in low likelihood area between prototypical embedding are difficult to iden- tify from diverse normals. MNAD [29] introduces skip- connection for diverse reconstruction and additional con- straints to get round the incorrect combination problem. But the latter forces model transmit more unrestrained 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. 12147 Figure 2. Illustration of our diversity-measurable method in ad- dressing the detection difficulty. Numbers in white are anomaly scores. a)Original input; b)Reconstructed reference; c)Coarse deformation; d)Fine deformation; e)Measurement of diversity1; f)Deformation-augmented error map assigns lower anomaly score to the anomaly-like sample than the true anomaly; g)Pixel-wise reconstruction error yields incorrect anomaly scores. formation with abnormal part by skip-connection, resulting in shortcut learning and undesired reconstruction of anoma- lies. A key to address the above tradeoff problem is to find a proper measurement of diversity that normal and abnor- mal samples have, which is positively correlated with the severity of anomaly. With such a measure, we do not need to fight against imperfect reconstruction of normals or un- desired reconstruction of anomalies, because anomalies can be detected more accurately by the diversity measure to- gether with the reconstruction error. Note that pixel-wise reconstruction error is not an ideal measurement of diver- sity, because the high-error region often confuses anomalies with diverse normals, e.g. normals with structural deforma- tion and anomalies with colors close to the background may yield unreliable reconstruction error. In this paper, we propose a Diversity-Measurable Anomaly Detection (DMAD) framework to enhance the measurability of reconstruction diversity so as to measure abnormality more accurately. Our basic idea is to decouple the reconstruction into compact representation of prototyp- ical normals and measurable deformations of more diverse normals and anomalies. The under-estimated reconstruc- tion error can be compensated by the diversity, which can be properly measured. To this end, the DMAD framework includes a Pyramid Deformation Module (PDM) to model and measure the diversity and an Information Compression Module (ICM) to learn the prototypical normal patterns. Inspired by [4, 15], we assume anomalies ( e.g. in video surveillance) can be represented as significant deformation of appearances, including positional changes and fine mo- tions. In contrast, diverse normal samples can be repre- sented as weaker deformations thus easily distinguished 1In this case, we only count fine deformation because deformations in position and angle are considered as normal. In real-world experiments, we consider both coarse and fine deformations.from the abnormal ones. Therefore, we design PDM to model the diversity of normals as well as the severity of anomalies. More specifically, PDM learns hierarchical two- dimensional deformation fields (Fig. 2 c,d) that describe the pixel-level transformation direction and distance from ref- erence (Fig. 2 b, which is reconstructed from prototypes in memory) to original input. In ICM, we learns compressed representation as sparse prototypes. As a result, a sin- gle memory item is enough to represent each normal clus- ter. This is more compact than other memory-based works which require multiple memory items. Integrating PDM with ICM, DMAD essentially decouples the deformation information (Fig. 2 e) from class prototypes and makes the final anomaly score more discriminative (Fig. 2 f). We evaluate our anomaly detection framework in scenar- ios of video surveillance and industrial defect detection. To apply DMAD in the latter scenario, we propose a variant of PDM, PPDM, to deal with the false positive issue in texture reconstruction. Extensive experimental results verify the ef- ficacy of our approach. Moreover, our method works well even in front of contaminated data and anomaly-like nor- mals. The main contributions of our work are as follows: • We introduce diversity-measurable anomaly detection framework which allows reconstruction-based models to achieve better tradeoff between reconstructing diverse normals and detecting unknown anomalies. • We propose pyramid deformation module to implement diversity measurement, in which the deformation infor- mation is explicitly separated from compact class pro- totypes and the resulting diversity measure is positively correlated to abnormality. • Our approach outperforms previous works on video anomaly detection and industrial defect detection, and works well in front of contaminated data and anomaly- like normals, demonstrating its broad suitability and ro- bustness.
Lin_Learning_To_Detect_Mirrors_From_Videos_via_Dual_Correspondences_CVPR_2023	Abstract Detecting mirrors from static images has received signif- icant research interest recently. However, detecting mirrors over dynamic scenes is still under-explored due to the lack of a high-quality dataset and an effective method for video mirror detection (VMD). To the best of our knowledge, this is the ﬁrst work to address the VMD problem from a deep- learning-based perspective. Our observation is that there are often correspondences between the contents inside (re- ﬂected) and outside (real) of a mirror, but such correspon- dences may not always appear in every frame, e.g., due to the change of camera pose. This inspires us to propose a video mirror detection method, named VMD-Net, that can tolerate spatially missing correspondences by considering the mirror correspondences at both the intra-frame level as well as inter-frame level via a dual correspondence mod- ule that looks over multiple frames spatially and tempo- rally for correlating correspondences. We further propose a ﬁrst large-scale dataset for VMD (named VMD-D), which contains 14,987 image frames from 269 videos with corre- sponding manually annotated masks. Experimental results show that the proposed method outperforms SOTA methods from relevant ﬁelds. To enable real-time VMD, our method efﬁciently utilizes the backbone features by removing the redundant multi-level module design and gets rid of post- processing of the output maps commonly used in existing methods, making it very efﬁcient and practical for real-time video-based applications. Code, dataset, and models are available at https://jiaying.link/cvpr2023-vmd/	1. Introduction Mirrors appear everywhere. They can adversely affect the performance of computer vision tasks ( e.g., depth esti- mation [ 35], vision-and-language navigation [ 2], semantic segmentation [ 49]), due to their fundamental property that they reﬂect objects from their surroundings. Thus, it is nec- *Joint ﬁrst authors. †Corresponding author. Image VCNet Ours GTinter-frame correspondenceintra-frame correspondenceFigure 1. Although state-of-the-art single-image mirror detection method VCNet [ 36] performs well on a single image ( e.g., the ﬁrst row) by using implicitly intra-frame correspondence, it may fail when the intra-frame cue is weak or even absent in some video frames ( e.g., the second and third rows). The lack in exploiting inter-frame information causes the current mirror detection meth- ods to produce inaccurate and inconsistent results when applied to VMD. In contrast, our method can perform well in both situations by utilizing the proposed dual correspondence module to exploit intra-frame (spatial) and inter-frame (temporal) correspondences. essary to build a robust computer vision model that can dis- tinguish mirrors from their surrounding objects correctly. Existing single-image mirror detection methods exploit different cues, such as context contrast [ 42], explicity cor- respondences [ 22], semantics association [ 14], and chirality and implicit correspondences [ 36], to detect mirrors from single RGB input images. Despite these recent efforts being put into the mirror detection problem, none of them focuses on detecting mirrors from videos. However, a lot of real- world computer vision applications are video-based ( e.g., robotic navigation, autonomous driving, and surveillance), rather than image-based. Hence, solving the video mirror detection (VMD) problem can beneﬁt these applications. In this paper, we aim to address the VMD problem. There are two major challenges with this problem. First, to the best of our knowledge, there are no existing large-scale datasets that can be used for training and evaluation on the VMD problem. Second, existing mirror detection 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. 9109 Figure 2. Quantitative comparison on the performance and efﬁ- ciency between existing mirror detection methods and our method for VMD. All models are trained/tested on the proposed VMD-D dataset, under a single RTX 3090 GPU. Our model has ⇠5 times smaller network parameters and runs ⇠18⇥faster than the state- of-the-art image-based mirror detection method, VCNet [ 36], and still outperforms it by a large margin. which are all developed for the image-based mirror detec- tion task, are all based on static cues. None of them take ad- vantages of the dynamic nature of videos in the VMD prob- lem. Figure 1shows that the current state-of-the-art mirror detection method, VCNet [ 36], may fail when correspon- dences are missing in some challenging frames ( e.g., second and third rows) due to, for example, the change of camera pose, even though it may perform well in some easy cases (e.g., the ﬁrst row). Besides, as the image-based mirror de- tection task is already very challenging, existing methods for this task often adopt heavy network design and time- consuming post-processing techniques [ 19] to improve their results. Figure 2shows that existing image-based mirror detection models run at about 1fps, even on one of the lat- est GPUs, which cannot support real-time VMD. All these drawbacks motivate us to develop a large-scale dataset and an effective/efﬁcient method for video mirror detection. In this paper, we address the VMD problem in two ways. First, we construct the ﬁrst large-scale video mirror detec- tion benchmark dataset (VMD-D). It contains 14,987 im- age frames in 269 videos, coming from diverse scenes. The constructed VMD-D dataset provides large-scale and high- diversity data for training and evaluation on the VMD prob- lem. Second, we propose an effective and efﬁcient method, called VMD-Net, for the VMD problem. The proposed method exploits multi-frame correspondences at both intra- frame (spatial) and inter-frame (temporal) levels. Compared with state-of-the-art image-based mirror detection methods, which typically adopt heavy pipelines, our method uses a light-weight network architecture without the need for anypost-processing techniques. As a result, our method is efﬁ- cient for real-time applications. In particular, our method has⇠5 times fewer network parameters and runs ⇠18⇥ faster than the latest state-of-the-art image-based mirror de- tection method, VCNet [ 36]. We conduct comprehensive experiments to demonstrate the effectiveness and efﬁciency of our proposed method. Experimental results show that our method outperforms state-of-the-art methods from relevant tasks on the proposed large-scale VMD-D dataset. Our key contributions can be summarized as follows: •We construct the ﬁrst large-scale video mirror detec- tion dataset, called VMD-D. The new dataset contains 14,988 image frames from 269 videos with precise an- notated masks. •We propose a novel network, called VMD-Net, to exploit both intra-frame and inter-frame correspon- dences via a dual correspondence (DC) module. This DC module allows VMD-Net to tolerate occassionally missing correspondences in the temporal dimension. •Extensive evaluations show that our method outper- forms existing state-of-the-art methods from relevant tasks on our proposed VMD-D dataset.
Liu_Building_Rearticulable_Models_for_Arbitrary_3D_Objects_From_4D_Point_CVPR_2023	Abstract We build rearticulable models for arbitrary everyday man-made objects containing an arbitrary number of parts that are connected together in arbitrary ways via 1 degree- of-freedom joints. Given point cloud videos of such every- day objects, our method identifies the distinct object parts, what parts are connected to what other parts, and the prop- erties of the joints connecting each part pair. We do this by jointly optimizing the part segmentation, transformation, and kinematics using a novel energy minimization frame- work. Our inferred animatable models, enables retargeting to novel poses with sparse point correspondences guidance. We test our method on a new articulating robot dataset, and the Sapiens dataset with common daily objects. Ex- periments show that our method outperforms two leading prior works on various metrics.	1. Introduction Consider the sequence of points clouds observations of articulating everyday objects shown in Figure 1. As hu- *equal advising, alphabetic order Figure 2. Many man-made everyday objects can be explained with rigid parts connected in a kinematic tree with 1DOF joints. mans, we can readily infer the kinematic structure of the underlying object, i.e. the different object parts and their connectivity and articulation relative with one another [21]. This paper develops computational techniques with similar abilities. Given point cloud videos of arbitrary everyday objects (with an arbitrary number of parts) undergoing ar- ticulation, we develop techniques to build animatable 3D re- constructions of the underlying object by a) identifying the distinct object parts, b) inferring what parts are connected to what other parts, and c) the properties of the joint be- tween each connected part pair. Success at this task enables 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. 21138 Arbitrary Realistic Joint Arbitrary Parts Constraints Kinematics Category-specific e.g. people [31] no yes no quadrupeds [51, 61] no yes no cartoons [53] no yes no DeepPart [57] yes no no NPP [11] yes no no ScrewNet [17] yes yes no UnsupMotion [42] no yes no Ditto [20] yes yes no MultiBodySync [15] yes no no WatchItMove [35] yes no yes Ours yes yes yes Table 1. Most past work on inferring rearticulable models is cat- egory specific. Building rearticulable models for arbitrary every- day man-made objects requires reasoning about arbitrary part ge- ometries, arbitrary part connectivity, and realistic joint constraints (1DOF w.r.t. parent part). We situate past work along these 3 di- mensions, and discuss major trends in Sec. 2. rearticulation of objects. Given just a few user clicks spec- ifying what point goes where, we can fill in the remaining geometry as shown on the right side in Figure 1. Most past work on inferring how objects articulate tack- les it in a category-specific manner, be it for people [31, 36, 40], quadrupeds [51, 61], or even everyday objects [34]. Category-specific treatment allows the use of specialized shape models (such as the SMPL model [31] for people), or defines the output space ( e.g. location of 2 hinge joints for the category eye-glasses). This limits applicability of such methods to categories that have a canonical topology, leaving out categories with large intra-class variation ( e.g. boxes that can have 1-4 hinge joints), or in-the-wild objects which may have an arbitrary number of parts connected in arbitrary ways ( e.g. robots). Only a very few past works tackle the problem of in- ferring rearticulable models in a category-agnostic manner. Huang et al. [15] only infer part segmentations, which by itself, is insufficient for rearticulation. Jiang et al. [20] only consider a single 1-DOF joint per object, dramatically re- stricting its application (think about a humanoid robot with four limbs, but the articulable model can only articulate one). Noguchi et al. [35] present the most complete solution but instead work with visual input and don’t incorporate the 1DOF constraint, i.e. a part can only rotate or translate about a fixed axis on the parent part, common to a large number of man-made objects as can be seen in Fig. 2. Inferring 3DOF / 6DOF joints leads to unrealistic rearticulation and is thus undesirable (consider the leg of eyeglasses can freely move or rotate). Our work fills this gap in the literature. Our method extracts 3D rearticulable models for arbitrary every- day objects (containing an arbitrary number of parts that are connected together in arbitrary ways via 1DOF joint) from point cloud videos. To the best of our knowledge, this is thefirst work to tackle this specific problem. Our proposed method jointly reasons about part geome- tries and their 1-DOF inter-connectivity with one another. At the heart of our approach is a novel continuous-discrete energy formulation that seeks to jointly learn parameters of the object model ( i.e. assignments of points in the canon- ical view to parts, and the connectivity of parts to one an- other) by minimizing shape and motion reconstruction error (after appropriate articulation of the inferred model) on the given views. As it is difficult to directly optimize in the presence of continuous and discrete variables with struc- tured constraints, we first estimate a relaxed model that in- fers parts that are free to follow an arbitrary 6DOF trajec- tory over time ( i.e. doesn’t require parts to be connected in a kinematic tree with 1DOF joints). We project the es- timated relaxed model to a kinematic model and continue to optimize with the reconstruction error to further finetune the estimated joint parameters. Our joint approach leads to better models and improved rearticulation as compared to adaptations of past methods [15, 35] to this task.
Li_3D_Cinemagraphy_From_a_Single_Image_CVPR_2023	Abstract We present 3D Cinemagraphy, a new technique that mar- ries 2D image animation with 3D photography. Given a single still image as input, our goal is to generate a video that contains both visual content animation and camera mo- tion. We empirically find that naively combining existing 2D image animation and 3D photography methods leads to ob- vious artifacts or inconsistent animation. Our key insight is that representing and animating the scene in 3D space offers a natural solution to this task. To this end, we first convert the input image into feature-based layered depth images using predicted depth values, followed by unproject- ing them to a feature point cloud. To animate the scene, we perform motion estimation and lift the 2D motion into the 3D scene flow. Finally, to resolve the problem of hole emer- gence as points move forward, we propose to bidirectionally displace the point cloud as per the scene flow and synthe- size novel views by separately projecting them into target image planes and blending the results. Extensive experi- ments demonstrate the effectiveness of our method. A user study is also conducted to validate the compelling rendering results of our method. *Corresponding author.	1. Introduction Nowadays, since people can easily take images using smartphone cameras, the number of online photos has in- creased drastically. However, with the rise of online video- sharing platforms such as YouTube and TikTok, people are no longer content with static images as they have grown ac- customed to watching videos. It would be great if we could animate those still images and synthesize videos for a bet- ter experience. These living images, termed cinemagraphs, have already been created and gained rapid popularity on- line [1, 71]. Although cinemagraphs may engage people with the content for longer than a regular photo, they usu- ally fail to deliver an immersive sense of 3D to audiences. This is because cinemagraphs are usually based on a static camera and fail to produce parallax effects. We are there- fore motivated to explore ways of animating the photos and moving around the cameras at the same time. As shown in Fig. 1, this will bring many still images to life and provide a drastically vivid experience. In this paper, we are interested in making the first step towards 3D cinemagraphy that allows both realistic anima- tion of the scene and camera motions with compelling par- allax effects from a single image. There are plenty of at- tempts to tackle either of the two problems. Single-image animation methods [12, 19, 35] manage to produce a real- 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. 4595 istic animated video from a single image, but they usually operate in 2D space, and therefore they cannot create cam- era movement effects. Classic novel view synthesis meth- ods [5, 6, 9, 14, 25] and recent implicit neural representa- tions [37, 40, 58] entail densely captured views as input to render unseen camera perspectives. Single-shot novel view synthesis approaches [21, 39, 52, 66] exhibit the potential for generating novel camera trajectories of the scene from a single image. Nonetheless, these methods usually hypoth- esize that the observed scene is static without moving el- ements. Directly combining existing state-of-the-art solu- tions of single-image animation and novel view synthesis yields visual artifacts or inconsistent animation. To address the above challenges, we present a novel framework that solves the joint task of image animation and novel view synthesis. This framework can be trained to cre- ate 3D cinemagraphs from a single still image. Our key intuition is that handling this new task in 3D space would naturally enable both animation and moving cameras simul- taneously. With this in mind, we first represent the scene as feature-based layered depth images (LDIs) [50] and unpro- ject the feature LDIs into a feature point cloud. To ani- mate the scene, we perform motion estimation and lift the 2D motion to 3D scene flow using depth values predicted by DPT [45]. Next, we animate the point cloud according to the scene flow. To resolve the problem of hole emer- gence as points move forward, we are inspired by prior works [3, 19, 38] and propose a 3D symmetric animation technique to bidirectionally displace point clouds, which can effectively fill in those unknown regions. Finally, we synthesize novel views at time tby rendering point clouds into target image planes and blending the results. In this manner, our proposed method can automatically create 3D cinemagraphs from a single image. Moreover, our frame- work is highly extensible, e.g., we can augment our motion estimator with user-defined masks and flow hints for accu- rate flow estimation and controllable animation. In summary, our main contributions are: • We propose a new task of creating 3D cinemagraphs from single images. To this end, we propose a novel framework that jointly learns to solve the task of image animation and novel view synthesis in 3D space. • We design a 3D symmetric animation technique to ad- dress the hole problem as points move forward. • Our framework is flexible and customized. We can achieve controllable animation by augmenting our mo- tion estimator with user-defined masks and flow hints.
Li_Are_Data-Driven_Explanations_Robust_Against_Out-of-Distribution_Data_CVPR_2023	Abstract As black-box models increasingly power high-stakes ap- plications, a variety of data-driven explanation methods have been introduced. Meanwhile, machine learning mod- els are constantly challenged by distributional shifts. A question naturally arises: Are data-driven explanations ro- bust against out-of-distribution data? Our empirical re- sults show that even though predict correctly, the model might still yield unreliable explanations under distribu- tional shifts. How to develop robust explanations against out-of-distribution data? To address this problem, we propose an end-to-end model-agnostic learning framework Distributionally Robust Explanations (DRE). The key idea is, inspired by self-supervised learning, to fully utilizes the inter-distribution information to provide supervisory sig- nals for the learning of explanations without human anno- tation. Can robust explanations benefit the model’s general- ization capability? We conduct extensive experiments on a wide range of tasks and data types, including classification and regression on image and scientific tabular data. Our results demonstrate that the proposed method significantly improves the model’s performance in terms of explanation and prediction robustness against distributional shifts.	1. Introduction There has been an increasing trend to apply black-box machine learning (ML) models for high-stakes applications. The lack of explainability of models can have severe con- sequences in healthcare [48], criminal justice [61], and other domains. Meanwhile, ML models are inevitably ex- posed to unseen distributions that lie outside their train- ing space [28, 56]; a highly accurate model on average can fail catastrophically on out-of-distribution (OOD) data due to naturally-occurring variations, sub-populations, spurious correlations, and adversarial attacks. For example, a can- cer detector would erroneously predict samples from hospi- tals having different data acquisition protocols or equipment 1The source code and pre-trained models are available at: https: //github.com/tangli-udel/DRE . OriginalGroupDROERMIRMLocation 100Location 38Location 43DRE (ours)Location 46In-distribution (train)OOD (test) Figure 1. The explanations for in-andout-of-distribution data ofTerra Incognita [4] dataset. Note that GroupDRO [50] and IRM [2] are explicitly designed methods that can predict accu- rately across distributions. Although with correct predictions, the explanations of models trained by such methods would also high- light distribution-specific associations ( e.g., tree branches) except the object. This leads to unreliable explanations on OOD data. On the contrary, our model consistently focuses on the most discrimi- native features shared across distributions. manufacturers. Therefore, reliable explanations across dis- tributions are crucial for the safe deployment of ML models. However, existing works focus on the reliability of data- driven explanation methods [1, 64] while ignoring the ro- bustness of explanations against distributional shifts. A question naturally arises: Are data-driven explana- tions robust against out-of-distribution data? We empir- ically investigate this problem across different methods. Results of the Grad-CAM [51] explanations are shown in Fig. 1. We find that the distributional shifts would fur- ther obscure the decision-making process due to the black- box nature of ML models . As shown, the explanations fo- cus not only on the object but also spurious factors ( e.g., background pixels). Such distribution-specific associations would yield inconsistent explanations across distributions. 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. 3821 Eventually, it leads to unreliable explanations ( e.g., tree branches) on OOD data. This contradicts with human prior that the most discriminative features ought to be invariant. How to develop robust explanations against out-of- distribution data? Existing works on OOD generalization are limited to data augmentation [42, 52, 59], distribution alignment [3, 16, 31], Meta learning [12, 29, 40], or invari- ant learning [2,26]. However, without constraints on expla- nations, the model would still recklessly absorb all associ- ations found in the training data, including spurious cor- relations . To constrain the learning of explanations, ex- isting methods rely on explanation annotations [44, 54] or one-to-one mapping between image transforms [8, 19, 39]. However, there is no such mapping in general naturally- occurring distributional shifts. Furthermore, obtaining ground truth explanation annotations is prohibitively expen- sive [60], or even impossible due to subjectivity in real- world tasks [46]. To address the aforementioned limita- tions, we propose an end-to-end model-agnostic training framework Distributionally Robust Explanations (DRE) . The key idea is, inspired by self-supervised learning, to fully utilize the inter-distribution information to provide su- pervisory signals for explanation learning. Can robust explanations benefit the model’s generaliza- tion capability? We evaluate the proposed methods on a wide range of tasks in Sec. 4, including the classification and regression tasks on image and scientific tabular data. Our empirical results demonstrate the robustness of our ex- planations. The explanations of the model trained via the proposed method outperform existing methods in terms of explanation consistency, fidelity, and scientific plausibility. The extensive comparisons and ablation studies prove that our robust explanations significantly improve the model’s prediction accuracy on OOD data. As shown, the robust explanations would alleviate the model’s excessive reliance onspurious correlations , which are unrelated to the causal correlations of interest [2]. Furthermore, the enhanced ex- plainability can be generalized to a variety of data-driven explanation methods. In summary, our main contributions: • We comprehensively study the robustness of data- driven explanations against naturally-occurring distri- butional shifts. • We propose an end-to-end model-agnostic learn- ing framework Distributionally Robust Explanations (DRE). It fully utilizes inter-distribution information to provide supervisory signals for explanation learning without human annotations. • Empirical results in a wide range of tasks including classification and regression on image and scientific tabular data demonstrate superior explanation and pre- diction robustness of our model against OOD data.
Li_Adjustment_and_Alignment_for_Unbiased_Open_Set_Domain_Adaptation_CVPR_2023	Abstract Open Set Domain Adaptation (OSDA) transfers the model from a label-rich domain to a label-free one con- taining novel-class samples. Existing OSDA works over- look abundant novel-class semantics hidden in the source domain, leading to a biased model learning and trans- fer. Although the causality has been studied to remove the semantic-level bias, the non-available novel-class samples result in the failure of existing causal solutions in OSDA. To break through this barrier, we propose a novel causality- driven solution with the unexplored front-door adjustment theory, and then implement it with a theoretically grounded framework, coined A djustmen t and Alignment (ANNA), to achieve an unbiased OSDA. In a nutshell, ANNA consists of Front-Door Adjustment (FDA) to correct the biased learn- ing in the source domain and Decoupled Causal Align- ment (DCA) to transfer the model unbiasedly. On the one hand, FDA delves into fine-grained visual blocks to discover novel-class regions hidden in the base-class image. Then, it corrects the biased model optimization by implementing causal debiasing. On the other hand, DCA disentangles the base-class and novel-class regions with orthogonal masks, and then adapts the decoupled distribution for an unbiased model transfer. Extensive experiments show that ANNA achieves state-of-the-art results. The code is available at https://github.com/CityU-AIM-Group/Anna.	1. Introduction Unsupervised Domain Adaptation (UDA) [5, 8, 11, 13] has been well studied to transfer a model from a labeled domain to an unlabeled novel one, notably saving the label- ing labor for model re-implementation. However, existing UDA research follows a strong assumption that the two do- mains must share the same class space, which cannot make correct predictions for novel-class samples. This severely *Corresponding author. This work was supported by Hong Kong Research Grants Council (RGC) General Research Fund 11211221, and Innovation and Technology Commission-Innovation and Technology Fund ITS/100/20.limits real-world applications [25, 29], e.g., product recom- mendation and pathology identification with unseen classes. Aiming at addressing this issue, Open Set Domain Adap- tation (OSDA) [3, 17, 20, 29, 35] has been studied, which also needs to recognize the novel-class samples in the target domain as unknown . As shown in Figure 1(a) (top), follow- ing a similar pipeline, most existing works [3,17,20,29,35] utilize labeled base-class data to train a closed-set classi- fier in the source domain. Then, in the target domain, they adjust the model with two objectives, i.e., exploring novel samples to achieve base/novel-class separation (novel-class detection) and adapting the base-class distribution (domain alignment). Based on this pipeline, these works can suc- cessfully recognize some novel samples in the unlabelled target domain and align the base-class distribution well. While achieving great success, existing works [17, 20, 29] only consider base-class semantics in the source do- main, ignoring the novel-class spreading everywhere. This leads to a semantic-level bias between the base and novel class, further yielding a biased domain transfer for OSDA. To explore the deficiency of this bias, we visualize the base/novel-class activated regions, as shown in col. 1-2 of Figure 1(a) (bottom). It can be observed that existing ap- proaches can successfully find the base-class regions con- sistent with the image-level ground-truth chair , but can- not discover novel-class semantics, e.g., the yacht ,sea, and ground , etc. (The base and novel regions are highlighted in Figure 1(c) for better view.) Further, we conduct a per- pixel prediction on deepest features without global average pooling (col. 3), illustrating that the novel regions are mis- classified as some non-correlated base classes. These obser- vations imply that this semantic-level bias severely affects the judgment of the classifier even though the classifier can give a correct prediction for the whole image. Recently, several causality-based approaches [36,44,45] have been proposed to solve the semantic-level bias in the closed-set setting. These works [36,44,45] first conduct per- class statistics over the whole dataset to decouple the con- text, and then use decoupled components to correct the bi- ased model training in a class-balanced manner. This causal solution can successfully avoid biased model learning since 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. 24110 Figure 1. Illustration of the general pipeline (top) and observed bias (bottom) with the base/novel-class activation and per-pixel prediction (we conduct dense classification on each pixel of the 7 ×7 ResNet-50 [9] feature and highlight the pixels with the same result in the same color.) for (a) existing OSDA approaches, (b) our solution, and (c) base and novel regions in each image. the knowledge of all classes contributes to training each sample. Hence, the rational idea is to explore the causal- ity to solve the newly observed OSDA bias. However, it is intractable to implement existing causal solutions [36, 45] in OSDA since the context is unobserv- able in open-set setting [42,43]. Existing works [36,45] use backdoor adjustment theory [26] to remove the bias, which relies on the observable context with available data samples. Differently, in OSDA, the context is unobservable [42] since novel-class samples are missing in the source domain [29] and labels are non-available in the target domain, leading to the failure [42] of existing backdoor solutions [36, 44, 45]. Although the front-door adjustment [26] can break through this unobservable dilemma [26] by decoupling data samples instead of context1, it is still tricky to implement a semantic- level decoupling [36,45] on each data sample since each im- age is only assigned a single class label in classification [9]. Fortunately, as shown in Figure 1(c), we observe that each image can be decoupled into base/novel-class regions in this open-set setting, which motivates us to use the unexplored yet effective front-door adjustment [26] to remove the bias. Thus, we aim to correct the biased learning in the source domain and then align the decoupled cross-domain distribu- tion to achieve unbiased OSDA. See Sec. 3 for a theoretical analysis with Structural Causal Model. To address the problems mentioned above, we pro- pose a theoretically grounded framework, A djustmen t and Alignment (ANNA) for OSDA (see Figure 1(b) (top)) with causality, which consists of Front-Door Adjustment (FDA) to address the biased learning in the source domain, and Decoupled Causal Alignment (DCA) to transfer the model to the target domain unbiasedly. Specifically, in each base- class image, FDA delves into fine-grained visual blocks to discover novel-class regions, serving for correcting biased model learning with causal adjustment. As for the DCA module, we disentangle cross-domain images into base- class and novel-class regions with orthogonal masks, and then align the decoupled distribution free of bias. As shown 1See supplementary materials for a more detailed explanation.in Figure 1(b) (bottom), after eliminating the OSDA bias, the model can capture labeled base-class regions (col. 1) and unlabeled novel-class regions (col. 2) well. Besides, the per-pixel prediction (col. 3) gives a closer look at model inference, showing that ANNA fully considers fine-grained novel semantics like humans before making an image-level prediction. Our main contributions are as follows, • This work represents the first attempt that observes and formulates the ever-overlooked semantic-level bias in OSDA. To address this issue, we propose a theoreti- cally grounded framework, A djustmen t and Alignment (ANNA) with causality, achieving an unbiased OSDA. • We propose a Front-Door Adjustment (FDA) module to correct the biased closed-set learning, discovering and fully using novel-class regions hidden in images. • We design a Decoupled Causal Alignment (DCA) to achieve an unbiased model transfer, which decouples cross-domain images with fine-grained regions and aligns the decoupled distribution unbiasedly. • Extensive experiments on three benchmarks verify that ANNA achieves state-of-the-art performance. ANNA achieves the best HOS on all 12 sub-tasks of the chal- lenging Office-Home benchmark.
Kim_DATID-3D_Diversity-Preserved_Domain_Adaptation_Using_Text-to-Image_Diffusion_for_3D_Generative_CVPR_2023	Abstract Recent 3D generative models have achieved remarkable performance in synthesizing high resolution photorealistic images with view consistency and detailed 3D shapes, but training them for diverse domains is challenging since it re- quires massive training images and their camera distribution information. Text-guided domain adaptation methods have shown impressive performance on converting the 2D genera- tive model on one domain into the models on other domains with different styles by leveraging the CLIP (Contrastive Language-Image Pre-training), rather than collecting mas- sive datasets for those domains. However, one drawback of them is that the sample diversity in the original generative model is not well-preserved in the domain-adapted genera- tive models due to the deterministic nature of the CLIP text encoder. Text-guided domain adaptation will be even more challenging for 3D generative models not only because of catastrophic diversity loss, but also because of inferior text- image correspondence and poor image quality. Here we pro- pose DATID-3D, a domain adaptation method tailored for †Corresponding author.3D generative models using text-to-image diffusion models that can synthesize diverse images per text prompt without collecting additional images and camera information for the target domain. Unlike 3D extensions of prior text-guided domain adaptation methods, our novel pipeline was able to fine-tune the state-of-the-art 3D generator of the source domain to synthesize high resolution, multi-view consistent images in text-guided targeted domains without additional data, outperforming the existing text-guided domain adap- tation methods in diversity and text-image correspondence. Furthermore, we propose and demonstrate diverse 3D image manipulations such as one-shot instance-selected adapta- tion and single-view manipulated 3D reconstruction to fully enjoy diversity in text.	1. Introduction Recently, 3D generative models [5, 6, 13, 18, 19, 22, 31, 40–42, 59, 60, 65, 69, 74, 75] have been developed to extend 2D generative models for multi-view consistent and explic- itly pose-controlled image synthesis. Especially, some of them [5, 18, 74] combined 2D CNN generators like Style- GAN2 [28] with 3D inductive bias from the neural ren- 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. 14203 dering [38], enabling efficient synthesis of high-resolution photorealistic images with remarkable view consistency and detailed 3D shapes. These 3D generative models can be trained with single-view images and then can sample infinite 3D images in real-time, while 3D scene representation as neural implicit fields using NeRF [38] and its variants [3, 4, 8, 10, 14, 17, 20, 32 –34, 36, 45, 47, 50, 53, 54, 64, 66, 70 –73] require multi-view images and training for each scene. Training these state-of-the-art 3D generative models is challenging because it requires not only a large set of images but also the information on the camera pose distribution of those images. This requirement, unfortunately, has restricted these 3D models to the handful domains where camera pa- rameters are annotated (ShapeNetCar [7,61]) or off-the-shelf pose extractors are available (FFHQ [27], AFHQ [9, 26]). StyleNeRF [18] assumed the camera pose distribution as either Gaussian or uniform, but this assumption is valid only for a few pre-processed datasets. Transfer learning methods for 2D generative models [30, 39, 43, 44, 48, 55, 67, 68] with small dataset can widen the scope of 3D models potentially for multiple domains, but are also limited to a handful of domains with similar camera pose distribution as the source domain in practice. Text-guided domain adaptation methods [1,16] have been developed for 2D generative models as a promising approach to bypass the additional data curation issue for the target do- main. Leveraging the CLIP (Contrastive Language-Image Pre-training) models [51] pre-trained on a large number of image-text pairs with non-adversarial fine-tuning strate- gies, these methods perform text-driven domain adaptation. However, one drawback of them is the catastrophic loss of diversity inherent in a text prompt due to the determinis- tic embedding of the CLIP text encoder so that the sample diversity of the source domain 2D generative model is not preserved in the target domain 2D generative models. We confirmed this diversity loss with experiments. A text prompt “a photo of a 3D render of a face in Pixar style” should include lots of different characters’ styles in Pixar films such as Toy Story, Incredible, etc. However, CLIP-guided adapted generator can only synthesize samples with alike styles as illustrated in Figure 1 (see StyleGAN- NADA∗). Thus, we confirmed that naive extensions of these for 3D generative models show inferior text-image corre- spondence and poor quality of generated images in diversity. Optimizing with one text embedding yielded almost similar results even with different training seeds as shown in Fig- ure 2(a). Paraphrasing the text for obtaining different CLIP embeddings was also trained, but it also did not yield that many different results as illustrated in Figure 2(b). Using different CLIP encoders for a single text as in Figure 2(c) did provide different samples, but it was not an option in general since only a few CLIP encoders have been released, and retraining them requires massive servers in practice. Figure 2. Existing text-guided domain adaptation [1, 16] did not preserve the diversity in the source domain for the target domain. We propose a novel DATID-3D, a method of Domain Adaptation using Text-to-Image Diffusion tailored for 3D- aware Generative Models. Recent progress in text-to-image diffusion models enables to synthesize diverse high-quality images from one text prompt [52, 56, 58]. We first lever- age them to convert the samples from the pre-trained 3D generator into diverse pose-aware target images. Then, the target images are rectified through our novel CLIP and pose reconstruction-based filtering process. Using these filtered target images, 3D domain adaptation is performed while pre- serving diversity in the text as well as multi-view consistency. We apply our novel pipeline to the EG3D [5], a state-of-the- art 3D generator, enabling the synthesis of high-resolution multi-view consistent images in text-guided target domains as illustrated in Figure 1, without collecting additional im- ages with camera information for the target domains. Our results demonstrate superior quality, diversity, and high text- image correspondence in qualitative comparison, KID, and human evaluation compared to those of existing 2D text- guided domain adaptation methods for the 3D generative models. Furthermore, we propose one-shot instance-selected adaptation and single-view manipulated 3D reconstruction to fully enjoy diversity in the text by extending useful 2D applications of generative models.
Kulinski_StarCraftImage_A_Dataset_for_Prototyping_Spatial_Reasoning_Methods_for_Multi-Agent_CVPR_2023	Abstract Spatial reasoning tasks in multi-agent environments such as event prediction, agent type identification, or miss- ing data imputation are important for multiple applica- tions (e.g., autonomous surveillance over sensor networks and subtasks for reinforcement learning (RL)). StarCraft II game replays encode intelligent (and adversarial) multi- agent behavior and could provide a testbed for these tasks; however, extracting simple and standardized representa- tions for prototyping these tasks is laborious and hinders reproducibility. In contrast, MNIST and CIFAR10, despite their extreme simplicity, have enabled rapid prototyping and reproducibility of ML methods. Following the simplic- ity of these datasets, we construct a benchmark spatial rea- soning dataset based on StarCraft II replays that exhibit complex multi-agent behaviors, while still being as easy to use as MNIST and CIFAR10. Specifically, we carefully sum- marize a window of 255 consecutive game states to create 3.6 million summary images from 60,000 replays, includ- ing all relevant metadata such as game outcome and player races. We develop three formats of decreasing complexity: Hyperspectral images that include one channel for every unit type (similar to multispectral geospatial images), RGB images that mimic CIFAR10, and grayscale images that mimic MNIST. We show how this dataset can be used for prototyping spatial reasoning methods. All datasets, code for extraction, and code for dataset loading can be found at https://starcraftdata.davidinouye.com/.	1. Introduction Spatial tasks in multi-agent environments require rea- soning over both agents’ positions and the environmental context such as buildings, obstacles, or terrain features. These complex spatial reasoning tasks have applications in autonomous driving, autonomous surveillance over sensor networks, or reinforcement learning (RL) as subtasks of the ‡DEVCOM Army Research Laboratory †Corresponding Authors: Sean Kulinski [email protected] and David I. Inouye [email protected]. Figure 1. Two samples (one per row) showing (Blue box/left) our 64 x 64 StarCraftHyper dataset which contains all unit IDs and corresponding values for both players (color for unit IDs denotes categorical unit ids), (Green box/middle) StarCraftCIFAR10 (32 x 32) which is easy to interpret where blue is player 1, red is player 2, and green are neutral units such as terrain or resources, and (Orange box/right) StarCraftMNIST (28 x 28) which are grayscale images further simplified to show player 1 as light-gray, player 2 as dark-gray, and neutral as medium-level shades of gray. RL agent. For example, to predict a car collision, an au- tonomous driving system needs to reason about other cars, road conditions, road signs, and buildings. For autonomous surveillance over sensor networks, the system would need to reason over the positions of objects, buildings, and other agents to determine if a new agent is normal or abnormal or to impute missing sensor values. An RL system may want to predict the cumulative or final reward or impute miss- ing values given only an incomplete snapshot of the world state, i.e., partial observability. Yet, collecting large realistic datasets for these tasks is expensive and laborious. Due to the challenge of collecting real-world data, prac- titioners have turned to (semi-)synthetic sources for creat- ing large clean datasets of photo-realistic images or videos [11, 12, 24, 40]. For example, [11] leveraged the Grand Theft Auto V game engine to collect a synthetic video dataset for pedestrian detection and tracking. [4] overlays aerial images with crowd simulations to provide a crowd density estimation dataset. Yet, despite near photo-realism, these prior datasets focus on simple multi-agent environ- 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. 22004 ments (e.g., pedestrian-like simulations [11, 40]) and thus lack complex (or strategic) agent and object positioning. In sharp contrast to these prior datasets, human-based replays of the real-time strategy game StarCraft II capture complex strategic and naturally adversarial positioning of agents and objects (e.g., buildings and outposts). Indeed, the human player provides thousands of micro-commands that produce an overall intelligent and strategic positioning of agents and building units. The release of the S
Li_BBDM_Image-to-Image_Translation_With_Brownian_Bridge_Diffusion_Models_CVPR_2023	Abstract Image-to-image translation is an important and chal- lenging problem in computer vision and image process- ing. Diffusion models (DM) have shown great potentials for high-quality image synthesis, and have gained competi- tive performance on the task of image-to-image translation. However, most of the existing diffusion models treat image- to-image translation as conditional generation processes, and suffer heavily from the gap between distinct domains. In this paper, a novel image-to-image translation method based on the Brownian Bridge Diffusion Model (BBDM) is proposed, which models image-to-image translation as a stochastic Brownian Bridge process, and learns the trans- lation between two domains directly through the bidirec- tional diffusion process rather than a conditional gener- ation process. To the best of our knowledge, it is the first work that proposes Brownian Bridge diffusion process for image-to-image translation. Experimental results on various benchmarks demonstrate that the proposed BBDM model achieves competitive performance through both vi- sual inspection and measurable metrics.	1. Introduction Image-to-image translation [14] refers to building a map- ping between two distinct image domains. Numerous prob- lems in computer vision and graphics can be formulated as image-to-image translation problems, such as style trans- fer [3,9,13,22], semantic image synthesis [21,24,34,36,37, 40] and sketch-to-photo synthesis [2, 14, 43]. A natural approach to image-to-image translation is to learn the conditional distribution of the target images given the samples from the input domain. Pix2Pix [14] is one of the most popular image-to-image translation methods. It is a typical conditional Generative Adversarial Network (GAN) [26], and the domain translation is accomplished by learning a mapping from the input image to the output im- Figure 1. Comparison of directed graphical models of BBDM (Brownian Bridge Diffusion Model) and DDPM (Denoising Dif- fusion Probabilistic Model). age. In addition, a specific adversarial loss function is also trained to constrain the domain mapping. Despite the high fidelity translation performance, they are notoriously hard to train [1, 10] and often drop modes in the output distri- bution [23, 27]. In addition, most GAN-based image-to- image translation methods also suffer from the lack of di- verse translation results since they typically model the task as a one-to-one mapping. Although other generative models such as Autoregressive Models [25, 39], V AEs (Variational Autoencoders) [16,38], and Normalizing Flows [7,15] suc- ceeded in some specific applications, they have not gained the same level of sample quality and general applicability as GANs. Recently, diffusion models [12, 31] have shown compet- itive performance on producing high-quality images com- pared with GAN-based models [6]. Several conditional dif- fusion models [2, 4, 28–30] have been proposed for image- to-image translation tasks. These methods treat image-to- image translation as conditional image generation by in- tegrating the encoded feature of the reference image into the U-Net in the reverse process (the first row of Figure 1) to guide the diffusion towards the target domain. De- 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. 1952 Figure 2. Architecture of our proposed Brownian Bridge Diffusion Model (BBDM). spite some practical success, the above condition mecha- nism does not have a clear theoretical guarantee that the final diffusion result yields the desired conditional distri- bution. Therefore, most of the conditional diffusion mod- els suffer from poor model generalization, and can only be adapted to some specific applications where the conditional input has high similarity with the output, such as inpaint- ing and super-resolution [2, 4, 30]. Although LDM (Latent Diffusion Model) [28] improved the model generalization by conducting diffusion process in the latent space of cer- tain pre-trained models, it is still a conditional generation process and the multi-modal condition is projected and en- tangled via a complex attention mechanism which makes LDM much more difficult to get such a theoretical guaran- tee. Meanwhile, the performance of LDM differs greatly across different levels of latent features showing instability. In this paper, we propose a novel image-to-image trans- lation framework based on Brownian Bridge diffusion pro- cess. Compared with the existing diffusion methods, the proposed method directly builds the mapping between the input and the output domains through a Brownian Bridge stochastic process, rather than a conditional generation pro- cess. In order to speed up the training and inference pro- cess, we conduct the diffusion process in the same latent space as used in LDM [28]. However, the proposed method differs from LDM inherently in the way the mapping be- tween two image domains is modeled. The framework of BBDM is shown in the second row of Figure 1. It is easy to find that the reference image ysampled from domain B is only set as the initial point xT=yof the reverse diffu- sion, and it will not be utilized as a conditional input in the prediction network µθ(xt, t)at each step as done in related works [2, 4, 28, 30]. The main contributions of this paper include: 1. A novel image-to-image translation method based onBrownian Bridge diffusion process is proposed in this paper. As far as we know, it is the first work of Brow- nian Bridge diffusion process proposed for image-to- image translation. 2. The proposed method models image-to-image trans- lation as a stochastic Brownian Bridge process, and learns the translation between two domains directly through the bidirectional diffusion process. The proposed method avoids the conditional information leverage existing in related work with conditional dif- fusion models. 3. Quantitative and qualitative experiments demonstrate the proposed BBDM method achieves competitive per- formance on various image-to-image translation tasks.
Li_3D-Aware_Multi-Class_Image-to-Image_Translation_With_NeRFs_CVPR_2023	Abstract Recent advances in 3D-aware generative models (3D- aware GANs) combined with Neural Radiance Fields (NeRF) have achieved impressive results. However no prior works investigate 3D-aware GANs for 3D consistent multi- class image-to-image (3D-aware I2I) translation. Naively using 2D-I2I translation methods suffers from unrealistic shape/identity change. To perform 3D-aware multi-class I2I translation, we decouple this learning process into a multi-class 3D-aware GAN step and a 3D-aware I2I trans- *The corresponding author.lation step. In the first step, we propose two novel tech- niques: a new conditional architecture and an effective training strategy. In the second step, based on the well- trained multi-class 3D-aware GAN architecture, that pre- serves view-consistency, we construct a 3D-aware I2I trans- lation system. To further reduce the view-consistency prob- lems, we propose several new techniques, including a U- net-like adaptor network design, a hierarchical representa- tion constrain and a relative regularization loss. In exten- sive experiments on two datasets, quantitative and qualita- tive results demonstrate that we successfully perform 3D- aware I2I translation with multi-view consistency. Code 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. 12652 available in 3DI2I.	1. Introduction Neural Radiance Fields (NeRF) have increasingly gained attention with their outstanding capacity to synthesize high- quality view-consistent images [31,39,66]. Benefiting from the adversarial mechanism [11], StyleNeRF [12] and con- current works [4, 8, 44, 69] have successfully synthesized high-quality view-consistent, detailed 3D scenes by com- bining NeRF with StyleGAN-like generator design [22]. This recent progress in 3D-aware image synthesis has not yet been extended to 3D-aware I2I translation, where the aim is to translate in a 3D-consistent manner from a source scene to a target scene of another class (see Figure 1). A naive strategy is to use well-designed 2D-I2I trans- lation methods [15, 16, 26, 28, 46, 63, 65, 70]. These meth- ods, however, suffer from unrealistic shape/identity changes when changing the viewpoint, which are especially notable when looking at a video. Main target class characteristics, such as hairs, ears, and noses, are not geometrically realis- tic, leading to unrealistic results which are especially dis- turbing when applying I2I to translate videos. Also, these methods typically underestimate the viewpoint change and result in target videos with less viewpoint change than the source video. Another direction is to apply video-to- video synthesis methods [2, 3, 6, 30, 53]. These approaches, however, either rely heavily on labeled data or multi-view frames for each object. In this work, we assume that we only have access to single-view RGB data. To perform 3D-aware I2I translation, we extend the the- ory developed for 2D-I2I with recent developments in 3D- aware image synthesis. We decouple the learning process into a multi-class 3D-aware generative model step and a 3D-aware I2I translation step. The former can synthesize view-consistent 3D scenes given a scene label, thereby ad- dressing the 3D inconsistency problems we discussed for 2D-I2I. We will use this 3D-aware generative model to ini- tialize our 3D-aware I2I model. It therefore inherits the ca- pacity of synthesizing 3D consistent images. To train ef- fectively a multi-class 3D-aware generative model (see Fig- ure 2(b)), we provide a new training strategy consisting of: (1) training an unconditional 3D-aware generative model (i.e., StyleNeRF) and (2) partially initializing the multi- class 3D-aware generative model (i.e., multi-class StyleN- eRF) with the weights learned from StyleNeRF. In the 3D- aware I2I translation step, we design a 3D-aware I2I trans- lation architecture (Figure 2(f)) adapted from the trained multi-class StyleNeRF network. To be specific, we use the main network of the pretrained discriminator (Figure 2(b)) to initialize the encoder Eof the 3D-aware I2I translation model (Figure 2(f)), and correspondingly, the pretrained generator (Figure 2(b)) to initialize the 3D-aware I2I gen-erator (Figure 2(f)). This initialization inherits the capacity of being sensitive to the view information. Directly using the constructed 3D-aware I2I translation model (Figure 2(f)), there still exists some view-consistency problem. This is because of the lack of multi-view consis- tency regularization, and the usage of the single-view im- age. Therefore, to address these problems we introduce several techniques, including a U-net-like adaptor network design, a hierarchical representation constrain and a relative regularization loss. In sum, our work makes the following contributions : • We are the first to explore 3D-aware multi-class I2I trans- lation, which allows generating 3D consistent videos. • We decouple 3D-aware I2I translation into two steps. First, we propose a multi-class StyleNeRF. To train this multi-class StyleNeRF effectively, we provide a new training strategy. The second step is the proposal of a 3D-aware I2I translation architecture. • To further address the view-inconsistency problem of 3D- aware I2I translation, we propose several techniques: a U- net-like adaptor, a hierarchical representation constraint and a relative regularization loss. • On extensive experiments, we considerably outperform existing 2D-I2I systems with our 3D-aware I2I method when evaluating temporal consistency.
Liu_Soft_Augmentation_for_Image_Classification_CVPR_2023	Abstract Modern neural networks are over-parameterized and thus rely on strong regularization such as data augmenta- tion and weight decay to reduce overfitting and improve generalization. The dominant form of data augmentation applies invariant transforms, where the learning target of a sample is invariant to the transform applied to that sam- ple. We draw inspiration from human visual classifica- tion studies and propose generalizing augmentation with invariant transforms to soft augmentation where the learn- ing target softens non-linearly as a function of the de- gree of the transform applied to the sample: e.g., more ag- gressive image crop augmentations produce less confident learning targets. We demonstrate that soft targets allow for more aggressive data augmentation, offer more robust performance boosts, work with other augmentation poli- cies, and interestingly, produce better calibrated models (since they are trained to be less confident on aggressively cropped/occluded examples). Combined with existing ag- gressive augmentation strategies, soft targets 1) double the top-1 accuracy boost across Cifar-10, Cifar-100, ImageNet- 1K, and ImageNet-V2, 2) improve model occlusion perfor- mance by up to 4×, and 3) half the expected calibration error (ECE). Finally, we show that soft augmentation gen- eralizes to self-supervised classification tasks. Code avail- able at https://github.com/youngleox/soft_ augmentation	1. Introduction Deep neural networks have enjoyed great success in the past decade in domains such as visual understanding [42], natural language processing [5], and protein structure pre- diction [41]. However, modern deep learning models are often over-parameterized and prone to overfitting. In addi- tion to designing models with better inductive biases, strong regularization techniques such as weight decay and data augmentation are often necessary for neural networks to achieve ideal performance. Data augmentation is often a computationally cheap and effective way to regularize mod-els and mitigate overfitting. The dominant form of data aug- mentation modifies training samples with invariant trans- forms – transformations of the data where it is assumed that the identity of the sample is invariant to the transforms. Indeed, the notion of visual invariance is supported by evidence found from biological visual systems [54]. The robustness of human visual recognition has long been docu- mented and inspired many learning methods including data augmentation and architectural improvement [19, 47]. This paper focuses on the counterpart of human visual robust- ness, namely how our vision fails . Instead of maintaining perfect invariance, human visual confidence degrades non- linearly as a function of the degree of transforms such as occlusion, likely as a result of information loss [44]. We propose modeling the transform-induced information loss for learned image classifiers and summarize the contribu- tions as follows: • We propose Soft Augmentation as a generalization of data augmentation with invariant transforms. With Soft Aug- mentation, the learning target of a transformed training sample softens . We empirically compare several soften- ing strategies and prescribe a robust non-linear softening formula. • With a frozen softening strategy, we show that replac- ing standard crop augmentation with soft crop augmenta- tion allows for more aggressive augmentation, and dou- bles the top-1 accuracy boost of RandAugment [8] across Cifar-10, Cifar-100, ImageNet-1K, and ImageNet-V2. • Soft Augmentation improves model occlusion robustness by achieving up to more than 4×Top-1 accuracy boost on heavily occluded images. • Combined with TrivialAugment [37], Soft Augmentation further reduces top-1 error and improves model calibra- tion by reducing expected calibration error by more than half, outperforming 5-ensemble methods [25]. • In addition to supervised image classification models, Soft Augmentation also boosts the performance of self- supervised models, demonstrating its generalizability. 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. 16241 -32 -24 -16 -8 0 816 24 32 tx -32-24-16-808162432ty Top-1 Error: 20.80Standard Hard Crop -32 -24 -16 -8 0 816 24 32 tx -32-24-16-808162432ty 22.99(+2.19)Aggressive Hard Crop -32 -24 -16 -8 0 816 24 32 tx -32-24-16-808162432ty 18.31(−2.49)Soft Augmentation 01 01 01 Target Confidence (p) original image 77% visible 38% visible 22% visibleFigure 1. Traditional augmentation encourages invariance by requiring augmented samples to produce the same target label; we visualize the translational offset range (tx, ty )of Standard Hard Crop augmentations for 32×32images from Cifar-100 on the left, reporting the top-1 error of a baseline ResNet-18. Naively increasing the augmentation range without reducing target confidence increases error (middle ), but softening the target label by reducing the target confidence for extreme augmentations reduces the error ( right ), allowing for training with even more aggressive augmentations that may even produce blank images . Our work also shows that soft augmentations produce models that are more robust to occlusions (since they encounter larger occlusions during training) and models that are better calibrated (since they are trained to be less-confident on such occluded examples).
Lin_Supervised_Masked_Knowledge_Distillation_for_Few-Shot_Transformers_CVPR_2023	Abstract Vision Transformers (ViTs) emerge to achieve impres- sive performance on many data-abundant computer vision tasks by capturing long-range dependencies among local features. However, under few-shot learning (FSL) settings on small datasets with only a few labeled data, ViT tends to overﬁt and suffers from severe performance degradation due to its absence of CNN-alike inductive bias. Previous works in FSL avoid such problem either through the help of self-supervised auxiliary losses, or through the dextile uses of label information under supervised settings. But the gap between self-supervised and supervised few-shot Transformers is still unﬁlled. Inspired by recent advances in self-supervised knowledge distillation and masked image modeling (MIM), we propose a novel Supervised Masked Knowledge Distillation model (SMKD) for few-shot Trans- formers which incorporates label information into self- distillation frameworks. Compared with previous self- supervised methods, we allow intra-class knowledge dis- tillation on both class and patch tokens, and introduce the challenging task of masked patch tokens reconstruc- tion across intra-class images. Experimental results on four few-shot classiﬁcation benchmark datasets show that our method with simple design outperforms previous meth- ods by a large margin and achieves a new start-of-the-art. Detailed ablation studies conﬁrm the effectiveness of each component of our model. Code for this paper is available here: https://github.com/HL-hanlin/SMKD.	1. Introduction Vision Transformers (ViTs) [16] have emerge as a competitive alternative to Convolutional Neural Networks (CNNs) [31] in recent years, and have achieved impressive performance in many vision tasks including image classi- ﬁcation [16, 38, 59, 66], object detection [3, 10, 26–28, 79], and object segmentation [50, 55]. Compared with CNNs, which introduce inductive bias through convolutional ker- nels with ﬁxed receptive ﬁelds [35], the attention layers in Equal contribution.yCorresponding author. (b) Few-Shot Transformers with Self-Supervision Auxiliary Regularization teacher student EMA same-class cross-image knowledge distillation (c) Few-Shot Transformers with Our Supervised Masked Knowledge Distillation image B mask same class label (a) Few-Shot Transformers with Soft Label Regularization image encoder supervised loss image A teacher student EMA same-image cross-view knowledge distillation image supervised loss … class labels view1 view2 [cls] token (global info) [patch] token (local info) ●Unified learning objective ●Maximizing similarity of both [cls] and corresponding [patch] tokens for intra-class images Our Idea Challenge ●How to establish correspondence for patch tokens across images ●Both CLS & patch tokens ●No negative samples ●Small batch size ●class prototypes Problems ●Hard to balance supervised & self-supervised learning objectives or patch-level soft labels Patch-level supervision from pre-trained teacher model [14] Latent attribute surrogates [70] Complex expert-designed targets …Figure 1. Comparison of the proposed idea and other exist- ing methods for few-shot Transformers. Our model mitigates the overﬁtting issue of few-shot Transformers, by extending the masked knowledge distillation framework into the supervised set- ting, and enforcing the alignment of [cls] and corresponding [patch] tokens for intra-class images. ViT allow it to model global token relations to capture long- range token dependencies. However, such ﬂexibility also comes at a cost: ViT is data-hungry and it needs to learn token dependencies purely from data. This property often makes it easy to overﬁt to datasets with small training set and suffer from severe generalization performance degra- dation [36, 37]. Therefore, we are motivated to study how to make ViTs generalize well on these small datasets, espe- cially under the few-shot learning (FSL) setting [19, 39, 62] which aims to recognize unseen new instances at test time just from only a few (e.g. one or ﬁve) labeled samples from each new categories. Most of the existing methods mitigate the overﬁtting is- sue of few-shot Transformers [14] using various regulariza- tions. For instance, some works utilize label information in a weaker [70], softer [42] way, or use label information efﬁciently through patch-level supervision [14]. However, these models usually design sophisticated learning targets. 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. 19649 On the other hand, self-distillation techniques [4, 8], and particularly, the recent masked self-distillation [29, 45, 77], which distills knowledge learned from an uncorrupted im- age to the knowledge predicted from a masked image, have lead an emerging trend in self-supervised Transformers in various ﬁelds [15, 68]. Inspired by such success, recent works in FSL attempt to incorporate self-supervised pretext tasks into the standard supervised learning through auxil- iary losses [37, 43, 46], or to adopt a self-supervised pre- training, supervised training two-stage framework to train few-shot Transformers [18, 32]. Compared with traditional supervised methods, self-supervision can learn less biased representations towards base class, which usually leads to better generalization ability for novel classes [40]. How- ever, the two learning objectives of self-supervision and supervision are conﬂicting and it is hard to balance them during training. Therefore, how to efﬁciently leverage the strengths of self-supervised learning to alleviate the overﬁt- ting issue of supervised training remains a challenge. In this work, we propose a novel supervised masked knowledge distillation framework (SMKD) for few-shot Transformers, which handles the aforementioned challenge through a natural extension of the self-supervised masked knowledge distillation framework into the supervised set- ting (shown in Fig. 1). Different from supervised con- trastive learning [33] which only utilizes global image fea- tures for training, we leverage multi-scale information from the images (both global [cls] token and local [patch] tokens) to formulate the learning objectives, which has been demonstrated to be effective in the recent self-supervised Transformer methods [29, 77]. For global [cls] tokens, we can simply maximize the similarity for intra-class im- ages. However, it is non-trivial and challenging to formulate the learning objectives for local [patch] tokens because we do not have ground-truth patch-level annotations. To address this problem, we propose to estimate the similarity between [patch] tokens across intra-class images using cross-attention, and enforce the alignment of correspond- ing[patch] tokens. Particularly, reconstructing masked [patch] tokens across intra-class images increases the difﬁculty of model learning, thus encouraging learning gen- eralizable few-shot Transformer models by jointly exploit- ing the holistic knowledge of the image and the similarity of intra-class images. As shown in Fig. 2, we compare our model with the existing self-supervised/supervised learning methods. Our model is a natural extension of the supervised contrastive learning method [33] and self-supervised knowledge dis- tillation methods [4, 77]. Thus our model inherits both the advantage of method [33] for effectively leveraging la- bel information, and the advantages of methods [4, 77] for not needing large batch size and negative samples. Mean- while, the newly-introduced challenging task of masked (momentum) encoder encoder EMA image similarity & dissimilarity (a) self-supervised contrastive (SimCLR, MOCO) (momentum) encoder encoder EMA image A similarity & dissimilarity (c) supervised contrastive (SupCon) image B momentum encoder encoder EMA image similarity (b) self-supervised knowledge distillation (DINO, iBOT) momentum encoder encoder EMA image A similarity (d) supervised masked knowledge distillation (ours) image B mask same class label same class label view1 view2 view1 view2 [cls] (+[patch]) [cls] [cls] [cls] + [patch] Figure 2. Comparison of other self-supervised/supervised frameworks. Our method (d) is a natural extension of (b) and (c), with the newly-introduced challenging task of masked [patch] tokens reconstruction across intra-class images. [patch] tokens reconstruction across intra-class images makes our method more powerful for learning generalizable few-shot Transformer models. Compared with contemporary works on few-shot Trans- formers [14, 32, 70], our framework enjoys several good properties from a practical point of view. (1) Our method does not introduce any additional learnable parameters be- sides the ViT backbone and projection head, which makes it easy to be combined with other methods [32,70,74]. (2) Our method is both effective and training-efﬁcient, with stronger performance and less training time on four few-shot classi- ﬁcation benchmarks, compared with [32, 70]. In a nutshell, our main contributions can be summarized as follows: We propose a new supervised knowledge distillation framework (SMKD) that incorporates class label in- formation into self-distillation, thus ﬁlling the gap be- tween self-supervised knowledge distillation and tra- ditional supervised learning. Within the proposed framework, we design two supervised-contrastive losses on both class and patch levels, and introduce the challenging task of masked patch tokens reconstruction across intra-class images. Given its simple design, we test our SMKD on four few-shot datasets, and show that it achieves a new SOTA on CIFAR-FS and FC100 by a large margin, as well as competitive performance on mini-ImageNet andtiered -ImageNet using the simple prototype clas- siﬁcation method for few-shot evaluation.
Liu_OSAN_A_One-Stage_Alignment_Network_To_Unify_Multimodal_Alignment_and_CVPR_2023	Abstract Extending from unimodal to multimodal is a critical challenge for unsupervised domain adaptation (UDA). Two major problems emerge in unsupervised multimodal domain adaptation: domain adaptation and modality alignment. An intuitive way to handle these two problems is to fulfill these tasks in two separate stages: aligning modalities followed by domain adaptation, or vice versa. However, domains and modalities are not associated in most existing two-stage studies, and the relationship between them is not lever- aged which can provide complementary information to each other. In this paper, we unify these two stages into one to align domains and modalities simultaneously. In our model, a tensor-based alignment module (TAL) is presented to ex- plore the relationship between domains and modalities. By this means, domains and modalities can interact sufficiently and guide them to utilize complementary information for better results. Furthermore, to establish a bridge between domains, a dynamic domain generator (DDG) module is proposed to build transitional samples by mixing the shared information of two domains in a self-supervised manner, which helps our model learn a domain-invariant common representation space. Extensive experiments prove that our method can achieve superior performance in two real-world applications. The code will be publicly available.	1. Introduction With explosively emerging multimedia data on the In- ternet, the field of multimodal analysis achieves more and more attention [10, 13, 18, 19, 43]. Compared to extensive unimodal models in NLP and CV , learning adequate knowl- edge from multimodal signals is still preliminary but very important. Abundant data plays a key role in different sce- narios of multimodal analysis, such as pre-training or down- stream multimedia tasks. However, it is prohibitively ex- pensive and time-consuming to obtain large amounts of la- beled data. To eliminate this issue, domain adaptation (DA) PPPPPPPPPPPsharedspacePPPPPPPPPPPPNNNNNNPPPPNNN PPPPPPPPNNNNNNPNNNsourcedomain targetdomainNNNNNNNNNNNNNNNNNNmodality alignmentone-stagealignment modality2positivePnegativeNmodality1modality3domain alignmentsourcetargetFigure 1. Conception of our one-stage model. is raised to learn a model from a labeled dataset (source domain) that can be generalized to other related tasks with- out sufficient labeled data (target domain) [3]. Classical do- main adaptation can be classified into different categories: unsupervised domain adaptation (UDA), fully supervised domain adaptation, and semi-supervised domain adaptation [31]. In this paper, we focus on UDA where no samples in target domain are annotated. With this technique, it is not necessary to prepare a customized training dataset for a specific task, but it can perform the task effectively and efficiently. There are two challenges when applying domain adap- tation to multimodal scenarios [17]: (1) how to align the source and target domains and remit domain discrepancy, and (2) how to align multiple modalities and leverage mul- timodal information. Most existing works address these two problems in two consecutive stages: multimodal align- ment followed by domain adaptation [34, 41], or vice versa [14, 44]. However, they solve these two issues separately without considering their relationship: domain and modal- ity can be treated as two views to portray the intrinsic char- acteristic of multimodal data [8], and the hidden underlying relationship in these two views can provide complemen- tary information to each other. Unimodal domain adapta- tion methods can not work well in multimodal tasks due to the inability to preserve the relations between modalities at the same time. Through our experiments and analysis, we observe that the two-stage model could not achieve ideal performance. Fig.2 shows the learning curve of two-stage 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. 3551 model during training phase by 800 iterations for the task of multimodal sentiment analysis. It can be found that the learning curve of two-stage model is oscillating and con- verges slowly, which indicates that two-stage model is prob- ably not a superior solution. To handle these challenges, the objective of multimodal domain adaptation can be defined as: (1) Exploring the relationship between domains and modalities; (2) Finding a common domain-invariant cross- modal representation space to align domains and modalities simultaneously. Therefore, in this paper, we design a One-Stage Alignment Network (OSAN) to unify multimodal align- ment and domain adaptation in one stage. Fig.1 shows the conception of our one-stage model. Our method benefits from: (1) The modality and domain are associated and in- teracted to capture the relationship between domains and modalities, which can provide rich complementary infor- mation to each other. (2) Multimodal alignment and domain adaptation are unified in one stage, which allows our model to perform domain adaptation and leverage multimodal in- formation at the same time. In Fig.2, we observe that the learning curve of our method is relatively stable and con- verges better, which indicates that exploring the relation be- tween modality and domain contributes to our task. In summary, our contributions are as follows: (1) To capture the relationship between domain and modality, we propose a one-stage alignment network, called OSAN, to associate domain and modality. In this way, a joint domain-invariant and cross-modal representation space is learned in one stage. (2) We design a TAL module to bring sufficient interac- tions between domains and modalities and guide them to utilize complementary information for each other. (3) To effectively bridge distinct domains, a DDG mod- ule is developed to dynamically construct multiple new do- mains by combining knowledge of source and target do- mains and exploring intrinsic structure of data distribution. (4) Extensive experiments on two totally different tasks demonstrate the effectiveness of our method compared to the supervised and strongly UDA methods.
Liu_Target-Referenced_Reactive_Grasping_for_Dynamic_Objects_CVPR_2023	Abstract Reactive grasping, which enables the robot to success- fully grasp dynamic moving objects, is of great interest in robotics. Current methods mainly focus on the temporal smoothness of the predicted grasp poses but few consider their semantic consistency. Consequently, the predicted grasps are not guaranteed to fall on the same part of the same object, especially in cluttered scenes. In this paper, we propose to solve reactive grasping in a target-referenced setting by tracking through generated grasp spaces. Given a targeted grasp pose on an object and detected grasp poses in a new observation, our method is composed of two stages: 1) discovering grasp pose correspondences through an attentional graph neural network and selecting the one with the highest similarity with respect to the target pose; 2) refining the selected grasp poses based on target and histor- ical information. We evaluate our method on a large-scale benchmark GraspNet-1Billion. We also collect 30 scenes of dynamic objects for testing. The results suggest that our method outperforms other representative methods. Further- more, our real robot experiments achieve an average suc- cess rate of over 80 percent. Code and demos are available at:https://graspnet.net/reactive .	1. Introduction Reactive grasping is in great demand in the industry. For instance, in places where human-robot collaboration is heavily required like factories, stress on laborers will be sig- nificantly relieved if robots can receive tools from humans and complete the harder work for laborers. Such a vision is based on reactive grasping. On the contrary to static environments, in reactive grasp- ing, dynamic task setting poses new challenges for algo- rithm design. Previous research in this area mainly focuses † Cewu Lu is the corresponding author, a member of Qing Yuan Re- search Institute and MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, Shanghai, China (a) (b) Figure 1. (a) Classic reactive grasping guarantees the smoothness of the grasp poses but cannot predicts grasps on the same part of the hammer. (b) Our target-referenced reactive grasping takes se- mantic consistency into consideration. The generated grasp poses across frames are illustrated with blue grippers. on planning temporally smooth grasps [22, 42] to avoid wavy and jerky robot motion. Few of them pay attention to its semantic consistency. In short, given a targeted grasp at the first frame, we want the robot to grasp the same part of the object in the following frames. Additionally, it is not guaranteed that grasp predictions made by classical meth- ods fall on the same object in cluttered scenes. Hence, most of their experiments are conducted on single-object scenes. Unlike previous works, this work is aimed at achieving tem- porally smooth and semantically consistent reactive grasp- ing in clutter given a targeted grasp. We refer to such a task setting as target-referenced reactive grasping as shown in Fig.1. Note that despite robot handover is a major applica- tion scenario of reactive grasping, this work focuses on a more general task setting - dynamic object grasping. 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. 8824 A naive idea to solve this task is to generate reference grasp poses for the initial scene and consecutively track the object’s 6D pose. As the object moves, the initial grasp pose can be projected to a new coming frame based on the object’s 6D pose. Although such an idea seems to be natu- ral and valid, some bottleneck greatly degrades its viabil- ity. First of all, the solution to reactive grasping should be able to handle objects’ motion in real-time, meaning that it requires fast inference speed and immediate response to continuous environmental changes. However, 6D pose tracking may be time-consuming due to commonly-used in- stance segmentation [11,40]. Second, 6D pose tracking usu- ally requires objects’ prior knowledge, such as CAD mod- els [4,41], which is not always available in the real world as well or achieves only category-level generalization [37]. Different from tracking objects, we propose to track grasps by a two-stage policy instead. We also comply with the restriction that no prior knowledge of the ob- jects is allowed. Given a target grasp on a partial-view point cloud, we first discover its corresponding grasp among future frame’s detected grasp poses as coarse estimation. These gasp poses can be given by an off-the-shelf grasp de- tector. Inspired by recent progress in local feature match- ing, which often uses image descriptors like SIFT [17] to describe interesting regions of images, we view grasp poses and their corresponding features as geometric descriptors on a partial-view point cloud. Based on such an assump- tion, we can simply estimate correspondences between two grasp sets from two different observation frames by match- ing the associated grasp features. Note that in opposition to classical local feature matching, features of the entire scene are also incorporated to help achieve global aware- ness. Furthermore, consecutive matching along an observa- tion sequence may lead to the accumulation of error, on top of the coarse estimation through correspondence matching, we further use a memory-augmented coarse-to-fine module which uses both target grasp features and historical grasp features to refine the grasp tracking results for better tem- poral smoothness and semantic consistency. We conduct extensive experiments on two benchmarks to evaluate our method and demonstrate its effectiveness. The results show that our method outperforms two repre- sentative baseline methods. We also conduct real robot ex- periments on both single-object scenes and cluttered scenes. We report success rates of 81.25% for single-object scenes and81.67% for multi-object scenes.
Kennerley_2PCNet_Two-Phase_Consistency_Training_for_Day-to-Night_Unsupervised_Domain_Adaptive_Object_CVPR_2023	Abstract Object detection at night is a challenging problem due to the absence of night image annotations. Despite several domain adaptation methods, achieving high-precision re- sults remains an issue. False-positive error propagation is still observed in methods using the well-established student- teacher framework, particularly for small-scale and low- light objects. This paper proposes a two-phase consistency unsupervised domain adaptation network, 2PCNet, to ad- dress these issues. The network employs high-confidence bounding-box predictions from the teacher in the first phase and appends them to the student’s region proposals for the teacher to re-evaluate in the second phase, resulting in a combination of high and low confidence pseudo-labels. The night images and pseudo-labels are scaled-down before be- ing used as input to the student, providing stronger small- scale pseudo-labels. To address errors that arise from low- light regions and other night-related attributes in images, we propose a night-specific augmentation pipeline called NightAug. This pipeline involves applying random aug- mentations, such as glare, blur, and noise, to daytime im- ages. Experiments on publicly available datasets demon- strate that our method achieves superior results to state-of- the-art methods by 20%, and to supervised models trained directly on the target data.1	1. Introduction Nighttime object detection is critical in many applica- tions. However, the requirement of annotated data by su- pervised methods is impractical, since night data with anno- tations is few, and supervised methods are generally prone to overfitting to the training data. Among other reasons, this scarcity is due to poor lighting conditions which makes nighttime images hard to annotate. Hence, methods that 1www.github.com/mecarill/2pcnet Figure 1. Qualitative results of state-of-the-art DA methods, DA Faster-RCNN [3], UMT [7], Adaptive Teacher (AT) [15] and our method 2PCNet on the BDD100K [36] dataset. Unlike the SOTA methods, our method is able to detect dark and small scale objects with minimal additional false positive predictions. do not assume the availability of the annotations are more advantageous. Domain adaptation (DA) is an efficient solu- tion to this problem by allowing the use of readily available annotated source daytime datasets. A few domain adaptation methods have been proposed, e.g., adversarial learning which uses image and instance level classifiers [3] and similar concepts [22, 32]. However, these methods isolate the domain adaptation task purely to- wards the feature extractor, and suppress features of the target data for the sake of domain invariance. Recent un- supervised domain adaptation methods exploit the student- teacher framework (e.g. [1,7,11,15]). Since the student ini- tially learns from the supervised loss, there is a bias towards the source data. Augmentation [7,11] and adversarial learn- ing [15] have been proposed to address this problem. Un- fortunately, particularly for day-to-night unsupervised do- main adaptation, these methods suffer from a large num- 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. 11484 ber of inaccurate pseudo-labels produced by the teacher. In our investigation, the problem is notably due to insufficient knowledge of small scale features in the nighttime domain, which are then propagated through the learning process be- tween the teacher and student, resulting in poor object de- tection performance. To address the problem, in this paper, we present 2PC- Net, a two-phase consistency unsupervised domain adapta- tion network for nighttime object detection. Our 2PCNet merges the bounding-boxes of highly-confident pseudo- labels, which are predicted in phase one, together with re- gions proposed by the student’s region proposal network (RPN). The merged proposals are then used by the teacher to generate a new set of pseudo-labels in phase two. This provides a combination of high and low confidence pseudo- labels. These pseudo-labels are then matched with pre- dictions generated by the student. We can then utilise a weighted consistency loss to ensure that a higher weightage of our unsupervised loss is based on stronger pseudo-labels, yet allow for weaker pseudo-labels to influence the training. Equipped with this two-phase strategy, we address the problem of errors from small-scale objects. We devise a student-scaling technique, where night images and their pseudo-labels for the student are deliberately scaled down. In order to generate accurate pseudo-labels, images to the teacher remain at their full scale. This results in the pseudo- labels of larger objects, which are easier to predict, to be scaled down to smaller objects, allowing for an increase in small scale performance of the student. Nighttime images suffer from multiple complications not found in daytime scenes such as dark regions, glare, promi- nent noise, prominent blur, imbalanced lighting, etc. All these cause a problem, since the student, which was trained on daytime images, is much more biased towards the day- time domain’s characteristics. To mitigate this problem, we propose NightAug, a set of random nighttime specific augmentations. NightAug includes adding artificial glare, noise, blur, etc. that mimic the night conditions to day- time images. With NightAug we are able to reduce the bias of the student network towards the source data without re- sulting to adversarial learning or compute-intensive trans- lations. Overall, using 2PCNet, we can see the qualitative improvements of our result in Figure 1. In summary, the contributions of this paper are as follows: • We present 2PCNet, a two-phase consistency approach for student-teacher learning. 2PCNet takes advantage of highly confident teacher labels augmented with less confident regions, which are proposed by the scaled student. This strategy produces a sharp reduction of the error propagation in the learning process. • To address the bias of the student towards the source domain, we propose NightAug, a random night spe-cific augmentation pipeline to shift the characteristics of daytime images toward nighttime. • The effectiveness of our approach has been verified by comparing it with the state-of-the-art domain adapta- tion approaches. An improvement of +7.9AP(+20%) and +10.2AP(26%) over the SOTA on BDD100K and SHIFT has been achieved, respectively.
Ling_ShadowNeuS_Neural_SDF_Reconstruction_by_Shadow_Ray_Supervision_CVPR_2023	Abstract By supervising camera rays between a scene and multi- view image planes, NeRF reconstructs a neural scene rep- resentation for the task of novel view synthesis. On the other hand, shadow rays between the light source and the scene have yet to be considered. Therefore, we propose a novel shadow ray supervision scheme that optimizes both the samples along the ray and the ray location. By su- pervising shadow rays, we successfully reconstruct a neu- ral SDF of the scene from single-view images under mul- tiple lighting conditions. Given single-view binary shad- ows, we train a neural network to reconstruct a complete scene not limited by the camera’s line of sight. By further modeling the correlation between the image colors and the shadow rays, our technique can also be effectively extended to RGB inputs. We compare our method with previous works on challenging tasks of shape reconstruction from single- view binary shadow or RGB images and observe signif- icant improvements. The code and data are available at https://github.com/gerwang/ShadowNeuS .	1. Introduction Neural ﬁeld [ 43] has been used for 3D scene representa- tion in recent years. It achieves remarkable quality because of the ability to continuously parameterize a scene with a compact neural network. The neural network nature makes it amenable to various optimization tasks in 3D vision, in- cluding long-standing problems like image-based [ 28,51] and point cloud-based [ 26,31] 3D reconstruction. So more and more works are using neural ﬁelds as the 3D scene rep- resentation for various related tasks. Among these works, NeRF [ 27] is a representative method that incorporates a part of physically based light transport [ 38] into the neural ﬁeld. The light transport de- scribes light travels from the light source to the scene and then from the scene to the camera. NeRF considers the latter part to model the interaction between the scene and the cam- eras along the camera rays (rays from the camera through *Corresponding author Single-view inputs Reconstruction viewed at novel views Figure 1. Our method can reconstruct neural scenes from single- view images captured under multiple lightings by effectively lever- aging a novel shadow ray supervision scheme. the scene). By supervising these camera rays of different viewpoints with the corresponding recorded images, NeRF optimizes a neural ﬁeld to represent the scene. Then NeRF casts camera rays from novel viewpoints through the opti- mized neural ﬁeld to generate novel-view images. However, NeRF does not model the rays from the scene to the light source, which motivates us to consider: can we optimize a neural ﬁeld by supervising these rays? These rays are often called shadow rays as the light emitted from the light source can be absorbed by scene particles along the rays, resulting in varying light visibility (a.k.a. shadows) at the scene surface. By recording the incoming radiance at the surface, we should be able to supervise the shadow rays to infer the scene geometry. Given this observation, we derive a novel problem of su- pervising the shadow rays to optimize a neural ﬁeld rep- resenting the scene, analogizing to NeRF that models the camera rays. Like multiple viewpoints in NeRF, we illumi- nate the scene multiple times using different light directions to obtain sufﬁcient observations. For each illumination, we use a ﬁxed camera to record the light visibility at the scene 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. 175 surface as supervision labels for the shadow rays. As rays connecting the scene and the light source march through the 3D space, we can reconstruct a complete 3D shape not con- strained by the camera’s line of sight. We solve several challenges when supervising the shadow rays using camera inputs. In NeRF, each ray’s position can be uniquely determined by the known cam- era center, but shadow rays need to be determined by the scene surface, which is not given and has yet to be recon- structed. We solve this using an iterative updating strategy, where we sample shadow rays starting at the current sur- face estimation. More importantly, we make the sampled locations differentiable to the geometry representation, thus can optimize the starting positions of shadow rays. How- ever, this technique is insufﬁcient to derive correct gradi- ents at surface boundaries with abrupt depth changes, which coincides with recent ﬁndings in differentiable rendering [2,20,23,40,54]. Thus, we compute surface boundaries by aggregating shadow rays starting at multiple depth can- didates. It remains efﬁcient as boundaries only occupy a small amount of surface, while it signiﬁcantly improves the surface reconstruction quality. In addition, RGB val- ues recorded by the camera encode the outgoing radiance at the surface instead of the incoming radiance. The outgoing radiance is a coupling effect of light, material, and surface orientation. We propose to model the material and surface orientation to decompose the incoming radiance from RGB inputs to achieve reconstruction without needing shadow segmentation (Row 1 and 2 in Fig. 1). As material modeling is optional, our framework can also take binary shadow im- ages [ 18] to achieve shape reconstruction (Row 3 in Fig. 1). We compare our method with previous single-view re- construction methods (including shadow-only and RGB- based) and observe signiﬁcant improvements in shape re- construction. Theoretically, our method handles a dual problem of NeRF. So, comparing the corresponding parts of the two techniques can inspire readers to get a deeper un- derstanding of the essence of neural scene representation to a certain extent, as well as the relationship between them. Our contributions are: • A framework that exploits light visibility to reconstruct neural SDF from shadow or RGB images under multi- ple light conditions. • A shadow ray supervision scheme that embraces dif- ferentiable light visibility by simulating physical inter- actions along shadow rays, with efﬁcient handling of surface boundaries. • Comparisons with previous works on either RGB or binary shadow inputs to verify the accuracy and com- pleteness of the reconstructed scene representation.
Lee_Decomposed_Cross-Modal_Distillation_for_RGB-Based_Temporal_Action_Detection_CVPR_2023	Abstract Temporal action detection aims to predict the time inter- vals and the classes of action instances in the video. Despite the promising performance, existing two-stream models ex- hibit slow inference speed due to their reliance on compu- tationally expensive optical ﬂow. In this paper, we intro- duce a decomposed cross-modal distillation framework to build a strong RGB-based detector by transferring knowl- edge of the motion modality. Speciﬁcally, instead of direct distillation, we propose to separately learn RGB and motion representations, which are in turn combined to perform ac- tion localization. The dual-branch design and the asymmet- ric training objectives enable effective motion knowledge transfer while preserving RGB information intact. In addi- tion, we introduce a local attentive fusion to better exploit the multimodal complementarity. It is designed to preserve the local discriminability of the features that is important for action localization. Extensive experiments on the bench- marks verify the effectiveness of the proposed method in en- hancing RGB-based action detectors. Notably, our frame- work is agnostic to backbones and detection heads, bring- ing consistent gains across different model combinations.	1. Introduction With the popularization of mobile devices, a signiﬁcant number of videos are generated, uploaded, and shared ev- ery single day through various online platforms such as YouTube and TikTok. Accordingly, there arises the impor- tance of automatically analyzing untrimmed videos. As one of the major tasks, temporal action detection (or localiza- tion) [ 56] has attracted much attention, whose goal is to ﬁnd the time intervals of action instances in the given video. In recent years, a lot of efforts have been devoted to improving the action detection performance [ 28–30,36,37,74,79,84]. Most existing action detectors take as input two-stream data consisting of RGB frames and motion cues, e.g., opti- cal ﬂow [ 21,66,78]. Indeed, it is widely known that differ- *Corresponding author Distillation DetectorDetector (b) Decomposed distillation (Ours)Detector RGB framesDetector DetectorDistillation (a) Conventional distillation Optical flow RGB framesOptical flowFigure 1. Comparison between conventional distillation and ours. Framework MethodAverage mAP (%) RGB+OF RGB ∆ Anchor-based G-TAD [ 74] 41.5 26.9 −14.6 Anchor-freeAFSD [ 34] 52.4 43.3 −9.1 Actionformer [ 80] 62.2 55.5 −6.7 DETR-like TadTR [ 42] 56.7 46.0 −10.7 Proposal-free TAGS [ 47] 52.8 47.9 −4.9 Table 1. Impact of motion modality. We measure the average mAP under the IoU thresholds of [0.3:0.7:0.1] on THUMOS’14. ent modalities provide complementary information [ 6,24, 58,69]. To examine how much two-stream action detec- tors rely on the motion modality, we conduct an ablative study using a set of representative models1. As shown in Table 1, regardless of the framework types, all the mod- els experience sharp performance drops when the motion modality is absent, probably due to the static bias of video models [ 11,27,31,32]. This indicates that explicit motion cues are essential for accurate action detection. However, two-stream action detectors impose a cycle of dilemmas for real-world applications due to the heavy com- putational cost of motion modality. For instance, the most popular form of motion cues for action detection, TV- L1 optical ﬂow [ 66], is not real-time, taking 1.8 minutes to process a 1-min 224×224 video of 30 fps on a single GPU [ 58]. Although cheaper motion clues such as temporal gradient [ 63,70,85] can be alternatives, two-stream models still exhibit inefﬁciency at inference by doubling the net- 1Each model is reproduced by its ofﬁcial codebase. 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. 2373 work forwarding process. Therefore, it would be desirable to build strong RGB-based action detectors that can bridge the performance gap with two-stream methods. To this end, we focus on cross-modal knowledge distilla- tion [ 12,16], where the helpful knowledge of motion modal- ity is transferred to an RGB-based action detector during training in order to improve its performance. In contrast to conventional knowledge distillation [ 19,20,46,52] where the superior teacher guides the weak student, cross-modal distillation requires exploiting the complementarity of the teacher and student. However, existing cross-modal distil- lation approaches [ 12,13] fail to consider the difference and directly transfer the motion knowledge to the RGB model (Fig. 1a), as conventional distillation does. By design, the RGB and motion information are entangled with each other, making it difﬁcult to balance between them. As a result, they often achieve limited gains without careful tuning. To tackle the issue, we introduce a novel framework, named decomposed cross-modal distillation (Fig. 1b). In detail, our model adopts the split-and-merge paradigm, where the high-level features are decomposed into appear- ance and motion components within a dual-branch design. Then only the motion branch receives the distillation signal, while the other branch remains intact to learn appearance in- formation. For explicit decomposition, we adopt the shared detection head and the asymmetric objective functions for the branches. Moreover, we design a novel attentive fusion to effectively combine the multimodal information provided by the two branches. In contrast to existing attention meth- ods, the proposed fusion preserves local sensitivity which is important for accurate action detection. With these key components, we build a strong action detector that produces precise action predictions given only RGB frames. We conduct extensive experiments on the popular bench- marks, THUMOS’14 [ 22] and ActivityNet1.3 [ 4]. Experi- mental results show that the proposed framework enables effective cross-modal distillation by separating the RGB and motion features. Consequently, our model largely im- proves the performance of RGB-based action detectors, ex- hibiting its superiority over conventional distillation. The resulting RGB-based action detectors effectively bridge the gap with two-stream models. Moreover, we validate our ap- proach by utilizing another motion clue, i.e., temporal gra- dient, which has been underexplored for action detection. To summarize, our contributions are three-fold: 1) We propose a decomposed cross-modal distillation framework, where motion knowledge is transferred in a separate way such that appearance information is not harmed. 2) We de- sign a novel attentive fusion method that is able to exploit the complementarity of two modalities while sustaining the local discriminability of features. 3) Our method is gener- alizable to various backbones and detection heads, showing consistent improvements.
Liu_Class_Adaptive_Network_Calibration_CVPR_2023	Abstract Recent studies have revealed that, beyond conventional accuracy, calibration should also be considered for train- ing modern deep neural networks. To address miscalibra- tion during learning, some methods have explored different penalty functions as part of the learning objective, along- side a standard classiﬁcation loss, with a hyper-parameter controlling the relative contribution of each term. Never- theless, these methods share two major drawbacks: 1) the scalar balancing weight is the same for all classes, hinder- ing the ability to address different intrinsic difﬁculties or imbalance among classes; and 2) the balancing weight is usually ﬁxed without an adaptive strategy, which may pre- vent from reaching the best compromise between accuracy and calibration, and requires hyper-parameter search for each application. We propose Class Adaptive Label Smooth- ing (CALS) for calibrating deep networks, which allows to learn class-wise multipliers during training, yielding a powerful alternative to common label smoothing penalties. Our method builds on a general Augmented Lagrangian approach, a well-established technique in constrained opti- mization, but we introduce several modiﬁcations to tailor it for large-scale, class-adaptive training. Comprehensive eval- uation and multiple comparisons on a variety of benchmarks, including standard and long-tailed image classiﬁcation, se- mantic segmentation, and text classiﬁcation, demonstrate the superiority of the proposed method. The code is available at https://github.com/by-liu/CALS .	1. Introduction Deep Neural Networks (DNNs) have become the prevail- ing model in machine learning, particularly for computer vision [ 13] and natural language processing applications [ 44]. Increasingly powerful architectures [ 3,13,24], learning meth- ods [ 4,12] and a large body of other techniques [ 15,27] are constantly introduced. Nonetheless, recent studies [ 11,31] have shown that regardless of their superior discriminative *Equal Contributions. Correspondence to: { liubingyuan1988@ gmail.com ,[email protected] }performance, high-capacity modern DNNs are poorly cali- brated, i.e. failing to produce reliable predictive conﬁdences. Speciﬁcally, they tend to yield over-conﬁdent predictions, where the probability associated with the predicted class overestimates the actual likelihood. Since this is a critical is- sue in safety-sensitive applications like autonomous driving or computational medical diagnosis, the problem of DNN calibration has been attracting increasing attention in recent years [ 11,31,38]. Current calibration methods can be categorized into two main families. The ﬁrst family involves techniques that per- form an additional post-processing parameterized operation on the output logits (or pre-softmax activations) [ 11], with the calibration parameters of that operation obtained from a validation set by either learning or grid-search. Despite the simplicity and low computational cost, these methods have empirically proven to be highly effective [ 8,11]. However, their main drawback is that the choice of the optimal cali- bration parameters is highly sensitive to the trained model instance and validation set [ 22,31]. The second family of methods attempts to simultaneously optimize for accuracy and calibration during network train- ing. This is achieved by introducing, explicitly or implicitly, a secondary optimization goal involving the model’s predic- tive uncertainty, alongside the main training objective. As a result, a scalar balancing hyper-parameter is required to tune the relative contribution of each term in the overall loss function. Some examples of this type of approaches include: Explicit Conﬁdence Penalty (ECP) [ 38], Label Smoothing (LS) [ 32], Focal Loss (FL) [ 21] and its variant, Sample- Dependent Focal Loss (FLSD) [ 31]. It has been recently demonstrated in [ 22] that all these methods can be formu- lated as different penalty terms that enforce the same equal- ity constraint on the logits of the DNN: driving the logit distances towards zero. Here, logit distances refers to the vector of L1 distances between the highest logit value and the rest. Observing the non-informative nature of this equality constraint, [ 22] proposed to use a generalized inequality con- straint, only penalizing those logits for which the distance is larger than a pre-deﬁned margin, achieving state-of-the-art calibration performance on many different benchmarks. Although learning based methods achieve greater calibra- 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. 16070 Figure 1. Many techniques have been proposed for jointly improving accuracy and calibration during training [ 11,31], but they fail to consider uneven learning scenarios like high class imbalance or long-tail distributions. We show a comparison of the proposed CALS-ALM method and different learning approaches in terms of Calibration Error (ECE) vs Accuracy on the (a) ImageNet and (b) ImageNet-LT (long-tailed ImageNet) datasets. A lower ECE indicates better calibration: a better model should attain high ACC and low ECE . Among all the considered methods, CALS-ALM shows superior performance when considering both discriminative power and well-balanced probabilistic predictions, achieving best accuracy and calibration on ImageNet, and best calibration and second best accuracy on ImageNet-LT. tion performance [ 22,31], they have two major limitations: 1) The scalar balancing weight is equal for all classes. This hinders the network performance when some classes are harder to learn or less represented than others, such as in datasets with a large number of categories (ImageNet) or considerable class imbalance (ImageNet-LT). 2) The balanc- ing weight is usually ﬁxed before network optimization, with no learning or adaptive strategy throughout training. This can prevent the model from reaching the best compromise between accuracy and calibration. To address the above is- sues, we introduce Class Adaptive Label Smoothing method based on an Augmented Lagrangian Multiplier algorithm, which we refer to as CALS-ALM. Our Contributions can be summarized as follows: •We propose Class Adaptive Label Smoothing (CALS) for network calibration. Adaptive class-wise multipliers are introduced instead of the widely used single balancing weight, which addresses the above two issues: 1) CALS can handle a high number of classes with different intrinsic difﬁculties, e.g. ImageNet; 2) CALS can effectively learn from data suffering from class imbalance or a long-tailed distribution, e.g. ImageNet-LT. •Different from previous penalty based methods, we solve the resulting constrained optimization problem by imple- menting a modiﬁed Augmented Lagrangian Multiplier (ALM) algorithm, which yields adaptive and optimal weights for the constraints. We make some critical de- sign decisions in order to adapt ALM to the nature of modern learning techniques: 1) The inner convergence cri- terion in ALM is relaxed to a ﬁxed number of iterations in each inner stage, which is amenable to mini-batch stochas-tic gradient optimization in deep learning. 2) Popular techniques, such as data augmentation, batch normaliza- tion [ 15] and dropout [ 10], rule out the possibility of track- ing original samples and applying sample-wise multipliers. To overcome this complication, we introduce class-wise multipliers, instead of sample-wise multipliers in the stan- dard ALM. 3) The outer-step update for estimating optimal ALM multipliers is performed on the validation set, which is meaningful for training on large-scale training set and avoids potential overﬁtting. •Comprehensive experiments over a variety of applications and benchmarks, including standard image classiﬁcation (Tiny-ImageNet and ImageNet), long-tailed image clas- siﬁcation (ImageNet-LT), semantic segmentation (PAS- CAL VOC 2012), and text classiﬁcation (20 Newsgroups), demonstrate the effectiveness of our CALS-ALM method. As shown in Figure 1 , CALS-ALM yields superior per- formance over baselines and state-of-the-art calibration losses when considering both accuracy and calibration, es- pecially for more realistic large-scaled datasets with large number of classes or class imbalance.
Kim_Grounding_Counterfactual_Explanation_of_Image_Classifiers_to_Textual_Concept_Space_CVPR_2023	Abstract Concept-based explanation aims to provide concise and human-understandable explanations of an image classifier. However, existing concept-based explanation methods typ- ically require a significant amount of manually collected concept-annotated images. This is costly and runs the risk of human biases being involved in the explanation. In this paper, we propose Counterfactual explanation with text-driven concepts (CounTEX), where the concepts are defined only from text by leveraging a pre-trained multi- modal joint embedding space without additional concept- annotated datasets. A conceptual counterfactual explana- tion is generated with text-driven concepts. To utilize the text-driven concepts defined in the joint embedding space to interpret target classifier outcome, we present a novel pro- jection scheme for mapping the two spaces with a simple yet effective implementation. We show that CounTEX generates faithful explanations that provide a semantic understanding of model decision rationale robust to human bias.	1. Introduction Explainable artificial intelligence (XAI) aims to unveil the reasoning process of a black-box deep neural network. In the vision field, heatmap-style explanation has been ex- tensively studied to interpret image classifiers [20, 21, 24]. However, simply highlighting the pixels that significantly contribute to model outcome does not answer intuitive and actionable questions such as “What aspect of the region is important? Is it color? Or pattern?”. On the other hand, drawing human-understandable rationale from the highlighted pixels requires domain expert’s intervention and can thus be impacted by the human subjectivity [11]. In contrast, concept-based explanation can provide a more human-understandable and high-level semantic expla- †Work done during the internship at Amazon Alexa AI BlackBox DNN (a) Conventional concept activation vector (b) Proposed method (CounTEX)CLIP (text) CLIP latent space BlackBox latent spaceNegative dataset ( ) ... Positive dataset ( ) ...Concept activation vector (CAV) Concept direction“object” ( ) “striped object” ( )Figure 1. (a) Conventional concept-based explanation derives a CA V with the target model’s embedding of manually collected concept-annotated images. (b) CounTEX derives the concept di- rection directly from texts in CLIP latent space. nation [3, 4, 7, 9, 11]. Concept fundamentally indicates an abstract idea, and it is generally equated as a word such as “stripe” or “red”. The earliest approach to interpret how a specific concept affects the outcome of the target image classifier is concept activation vector, or CA V [11]. A CA V represents the direction of a concept within the target clas- sifier embedding space and has been widely adopted to sub- sequent concept-based explanations [15, 19, 25]. However, the CA Vs acquisition requires collections of human annotations. The CA V of a concept is typically pre- computed via two steps as depicted in Figure 1 (a); 1) col- lecting a number of positive and negative images that best represent a concept (e.g., images with and without stripes), 2) training a linear classifier (commonly support vector ma- chine) with the images. The vector normal to the deci- sion boundary serves as a CA V . Collecting positive/negative datasets in step 1 is not only costly but also poses the risk of admitting human biases in two aspects; diverging CA Vs for the same concept and unintended entanglement of multiple concepts. We will demonstrate in Section 2 that this may threaten credibility of explanation. 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. 10942 To tackle such challenges, we propose Counterfactual explanation with text-driven concepts (CounTEX), which derives the concept direction only from a text by leveraging the text-image joint embedding space, CLIP [16] (Figure 1 (
Liu_VLPD_Context-Aware_Pedestrian_Detection_via_Vision-Language_Semantic_Self-Supervision_CVPR_2023	Abstract Detecting pedestrians accurately in urban scenes is sig- nificant for realistic applications like autonomous driving or video surveillance. However, confusing human-like ob- jects often lead to wrong detections, and small scale or heavily occluded pedestrians are easily missed due to their unusual appearances. To address these challenges, only object regions are inadequate, thus how to fully utilize more explicit and semantic contexts becomes a key problem. Meanwhile, previous context-aware pedestrian detectors ei- ther only learn latent contexts with visual clues, or need laborious annotations to obtain explicit and semantic con- texts. Therefore, we propose in this paper a novel approach via Vision-Language semantic self-supervision for context- aware Pedestrian Detection (VLPD) to model explicitly se- mantic contexts without any extra annotations. Firstly, we propose a self-supervised Vision-Language Semantic (VLS) segmentation method, which learns both fully-supervised pedestrian detection and contextual segmentation via self- generated explicit labels of semantic classes by vision- language models. Furthermore, a self-supervised Prototyp- ical Semantic Contrastive (PSC) learning method is pro- posed to better discriminate pedestrians and other classes, based on more explicit and semantic contexts obtained from VLS. Extensive experiments on popular benchmarks show that our proposed VLPD achieves superior performances over the previous state-of-the-arts, particularly under chal- lenging circumstances like small scale and heavy occlusion. Code is available at https://github.com/lmy98129/VLPD.	1. Introduction With the recent advances of pedestrian detection, enor- mous applications benefit from such a fundamental per- ∗Equal contribution. †Corresponding author. (a) (c)HumanCarTruck… Self-SupervisedPretrainedLinguisticVectorsMappingtoVision(b) Task1:“LearntoRecognizetheContextsbyCross-ModalMapping”Task2:“LearntoDiscriminatePedestriansandContexts”…NegativePrototypes PedestrianBoundingBoxesPositivePrototypeandQueriesPrototypicalSemanticContrastiveLearning(Self-Supervised)(d)SemanticClasses PedestrianDetectorPretrainedVL-ModelMappingtoVision PedestrianDetectorAggregatePixelFeaturesSegmentPredictionInitializeloss Figure 1. Illustration of the problems by previous works (top) and our proposed method to tackle them (bottom). (a) and (b) are pre- dicted by [27]. Green boxes are correct, red ones are human-like traffic signs, and dashed blue ones are missing heavily occluded or small scale pedestrians. (c) and (d): We propose self-supervisions to recognize the contexts and discriminate them from pedestrians. ception technique, including person re-identification, video surveillance and autonomous driving. In the meantime, various challenges from the urban contexts, i.e., pedestri- ans and non-human objects, still hinder the better perfor- mances of detection. For example, confusing appearances of human-like objects often mislead the detector, as shown in Figure 1(a). Moreover, heavily occluded or small scale pedestrians have unusual appearances and cause missing detections as Figure 1(a) and (b). Apart from the object re- gions, the contexts are crucial to address these challenges. Nevertheless, previous methods still make inadequate in- vestigations on the contexts in urban scenarios. For 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. 6662 PretrainedVisualEncoderPretrainedVisualEncoder “Apictureof[CLS].” PretrainedTextEncoder PedestrianDetectionHead PedestrianBoundingBoxesHumanCarTruck…PretrainedLinguisticVectors *Cross-ModalMappingPixel-wiseAggregation Frozen Training𝓛𝑫𝒆𝒕𝓛𝑽𝑳𝑺VisualFeaturesVisualFeatures …NegProtos𝓛𝑷𝑺𝑪 …PosProtos&Queries(a) (b) Figure 2. The overall architecture of our proposed VLPD approach. (a) Vision-Language Semantic (VLS) segmentation obtains pseudo labels via Cross-Modal Mapping, then the Pretrained Visual Encoder learns fully-supervised detection ( LDet) and self-supervised segmen- tation to recognize semantic classes for explicit contexts without any annotations. (b) Prototypical Semantic Contrastive (PSC) learning lets the pixel-wise pedestrian features as queries closer to positive prototypes and further to negative ones based on Pixel-wise Aggregation. stance, manual contextual annotations from CityScapes [6] boost SMPD [15] on the pedestrian benchmark CityPersons [43], because they share homologous image data. Besides, a semi-supervised model yields pseudo labels for the Cal- tech dataset [9]. However, both these two solutions require expensive fine-grained annotations, especially for training the semi-supervised model. Moreover, other methods learn regional latent contexts merely from limited visual neigh- borhood [47], or non-human local proposals as negative samples for contrastive learning [23]. Without an explicit awareness of semantic classes in the contexts, these meth- ods thus still suffer from unsatisfactory performance. Besides, some pedestrian detection methods also indi- rectly handle the contexts. For the occlusion problems, many part-aware methods [4, 13, 19, 20, 28, 33, 41, 44, 45] adopt visible annotations for the occluded pedestrians, which indicate the occlusion by other pedestrians or non- human objects in the contexts. Whereas, these labels still need heavy labors of human annotators. For scale varia- tion [2,8,21,39,46], crowd occlusion [14,25,38,40,48,50] or generic hard pedestrians [1, 24, 26, 27, 34], most previ- ous works are intra-class, e.g., small pedestrians or crowded scenes, and thus irrelevant to context modeling problems. Inspired by the vision-language models, we notice a more explicit context modeling without any annotations via cross-modal mapping. For instance, DenseCLIP [32] is ini- tialized with vision-language pretrained CLIP model [31] to learn cross-modal mapping from pixel-wise features to linguistic vectors of human-annotated classes. Meanwhile, MaskCLIP [49] generates pseudo labels via cross-modal mapping and train another visual model. Hence, comple- menting the initialized mapping and pseudo labeling, we propose to recognize the semantic classes for explicit con-texts via self-supervised Vision-Language Semantic (VLS) segmentation, as shown in Figure 1(c) and 2(a). Furthermore, we consider that only pixel-wise scores are ambiguous to discriminate pedestrians and contexts. Due to the coarse-grained pseudo labels, some parts of pedestrians might have higher scores of other classes. Different from the regional contrastive learning [23], we introduce the con- cept of prototype [35, 51] for a global discrimination. Each pixel of pedestrian features is pulled closer to pixel-wise aggregated positive prototypes and pushed away from the negative ones of other classes based on the explicit contexts obtained from VLS. As illustrated in Figure 1(d) and 2(b), a novel contrastive self-supervision for pedestrian detection is proposed to better discriminate pedestrians and contexts. In conclusion, we have observed a dilemma between the heavy burden of manual annotation for explicit contexts and local implicit context modeling. Hence, we propose a novel approach to tackle these problems via Vision- Language se- mantic self-supervision for Pedestrian Detection ( VLPD ). The main contributions of this paper are as follows: • Firstly, the Vision-Language Semantic (VLS) segmen- tation method is proposed to model explicit seman- tic contexts by vision-language models. With pseudo labels via cross-modal mapping, the visual encoder learns fully-supervised detection and self-supervised segmentation to recognize the semantic classes for ex- plicit contexts. To our best knowledge, this is the first work to propose such a vision-language extra- annotation-free method for pedestrian detection . • Secondly, we further propose the Prototypical Seman- tic Contrastive (PSC) learning method to better dis- criminate pedestrians and contexts. The negative and 6663 positive prototypes are aggregated via the score maps of contextual semantic classes obtained from VLS and pedestrian bounding boxes, respectively. Each pixel of pedestrian features is pulled close to positive proto- types and pushed away from the negative ones, in order to strengthen the discrimination power of the detector. • Finally, by the integration of VLS and PSC, our pro- posed approach VLPD achieves superior performances over the previous state-of-the-art methods on popu- lar Caltech and CityPersons benchmarks, especially on the challenging small scale and occlusion subsets.
Lin_Memory-Friendly_Scalable_Super-Resolution_via_Rewinding_Lottery_Ticket_Hypothesis_CVPR_2023	Abstract Scalable deep Super-Resolution (SR) models are in- creasingly in demand, whose memory can be customized and tuned to the computational recourse of the platform.The existing dynamic scalable SR methods are not memory- friendly enough because multi-scale models have to be saved with a ﬁxed size for each model. Inspired by the suc-cess of Lottery Tickets Hypothesis (LTH) on image classi-ﬁcation, we explore the existence of unstructured scalable SR deep models, that is, we ﬁnd gradual shrinkage sub-networks of extreme sparsity named winning tickets. In thispaper , we propose a Memory-friendly Scalable SR frame-work (MSSR). The advantage is that only a single scalablemodel covers multiple SR models with different sizes, in- stead of reloading SR models of different sizes. Concretely, MSSR consists of the forward and backward stages, the for- mer for model compression and the latter for model expan-sion. In the forward st age, we take advant age of LT H with rewinding weights to progressively shrink the SR model and the pruning-out masks that form nested sets. Moreover , stochastic self-distillation (SSD) is conducted to boost theperformance of sub-networks. By stochastically selecting multiple depths, the current model inputs the selected fea- tures into the corresponding parts in the larger model andimproves the performance of the current model based onthe feedback results of the larger model. In the backward stage, the smaller SR model could be expanded by recov- ering and ﬁne-tuning the pruned parameters according tothe pruning-out masks obtained in the forward. Extensive experiments show the effectiveness of MMSR. The smallest- scale sub-network could achieve the sparsity of 94% andoutperforms the compared lightweight SR methods. Equal contribution †Corresponding authorsModel Scalability 20% sparsest sub-model 40% sparse sub-model 80% sparse sub-mode l 100% whole modelModel Deployment KVRU_ LM H :<20% :40%-80% MM :80%- 100% H LL :S_n :M_n :Cur_n :Free_n :S_w :M_w :Cur_w Figure 1. The ﬂowchart of scalable SR network. S n and S w denote the neurons and the neural connections (weights) of the simplest subnetwork; M n and M w denote the intermediate neu- rons and neural connections; Cur n and Cur w denote the speciﬁc neurons and neural connections belonging to the current subnet- work. Free n denotes the pruning-out neurons. The ﬁnal model (with 100% recovered parameters) reaches the original size. Thescalable model is adjustable to the memory resource allocation.	1. Introduction Single image super-resolution (SISR) aims to reconstruct a high-resolution (HR) image from the corresponding low- resolution (LR) one. With the rising of deep learning, deep SR methods have made incredible progress. However, theexisting SR models mostly require computational and mem- ory resources, so they do not favor resource-limited devices such as mobile phones, robotics, and some edge devices. The lightweight SR methods are attracting more at- tention for better application to resource-limited devices. The existing lightweight SR methods mainly focus on de- signing compact architectures [ 17,20] with a ﬁxed size, such as multi-scale pyramid [ 20], multiple-level receptive ﬁelds [ 17,18], and recursive learning [ 19]. However, most lightweight SR models with ﬁxed sizes are not ﬂexible inapplications. If one model does not match the resources 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. 14398 the platform, it has to be retrained by compression methods to match the resources and then reloaded onto the devices. The urgent demand to customize models based on de- ployment resources is increasing. Dynamic neural networksfor SR [ 14,22] are proposed to adjust the network architec- ture according to different computational resources. Theexisting dynamic deep SR models often explore dynamic depth or width [ 22,26], but they either require large mem- ory resources or are not convenient for users to wait for re- training another SR model. The former leads to saving themulti-scale SR models of different sizes and the latter leadsto retraining the model before being reused again. The lim-itation lies in that they are not memory-friendly. In manyedge-device SR applications, the devices may be scalable,that is, their memories may be small in the beginning andbe expanded later. Thus, we discuss two issues in this pa-per: 1) how to make a scalable lightweight model for themulti-scale computational recourse. 2) how to make the lightweight model expand to a larger-size model for better performance if the computational recourse is increased. As for the ﬁrst issue, inspired by the success of Lottery Ticket Hypothesis [ 10] which points out that there could ex- ist a properly pruned sub-network named winning tickets to achieve comparable performance against the original densenetwork in model compression of classiﬁcation, it is usedto ﬁnd the sub-network for SR. We are the ﬁrst to study the existence of scalable winning tickets for SR. Iterative prun-ing and rewinding weights in LTH are beneﬁcial to the scal- able lightweight SR model. Iterative pruning may compress the SR model according to an arbitrary size. It is observed in [24] that the winning tickets are related to an insufﬁcient DNN, and rewinding LTH outperforms the original LTH. That is, the initial weights in LTH are replaced with the T-iteration weights during pruning and ﬁne-tuning. The scal- able deep SR model is shown in Fig. 1. As for the second issue, the scalable SR model can cus- tomize parameters to adapt to different memory resources rather than load or ofﬂoad different models for different de- vices. In other words, during real applications, there will beonly one simple model to be employed for inference whose size is decided by the computational resource. In this paper, we propose a memory-friendly scalable deep SR model (MSSR) via rewinding LTH. We use the rewinding LTH [ 10] to generate our unstructured scalable mask. MSSR is backtracking and contains forward and backward stages. The former focuses on model compres-sion by rewinding LTH with iterative pruning and ﬁne- tuning, and the latter focuses on iterative model expansionuntil it goes back to its original size. Multi-scale winning tickets together with the pruning-out masks are obtained by rewinding LTH in the forward stage with the decrease in the number of parameters. The pruning-out masks are nested.In order to make the compressed SR model not degrade sig-niﬁcantly, stochastic self-distillation (SSD) is used to im- prove the representation of the small-scale SR model, andknowledge is transferred from the last-scale model to thecurrent scale model. In the backward stage, the smallest model is expanded gradually to the model with the originalsize with the expanded mask. The main contributions of this work are three-fold: • A memory-friendly scalable dynamic SR lightweight model via rewinding LTH is proposed. MSSR isre-conﬁgurable and switchable to sub-networks withdifferent sizes according to on-device resource con- straints on the ﬂy. • MSSR is backtracking, which contains forward and backward stages. Multi-scale winning tickets form nested masks for the multi-scale models. SSD is con- ducted by replacing the features in randomly selectedlayers between Teacher and Student to improve the performance of the scalable SR lightweight models. • Extensive experiments demonstrate that MSSR can generalize to different SR models as well as state-of-the-art attention-based models, ENLCN [ 1].
Kim_Relational_Context_Learning_for_Human-Object_Interaction_Detection_CVPR_2023	Abstract Recent state-of-the-art methods for HOI detection typ- ically build on transformer architectures with two decoder branches, one for human-object pair detection and the other for interaction classification. Such disentangled transform- ers, however, may suffer from insufficient context exchange between the branches and lead to a lack of context informa- tion for relational reasoning, which is critical in discover- ing HOI instances. In this work, we propose the multiplex relation network (MUREN) that performs rich context ex- change between three decoder branches using unary, pair- wise, and ternary relations of human, object, and interac- tion tokens. The proposed method learns comprehensive re- lational contexts for discovering HOI instances, achieving state-of-the-art performance on two standard benchmarks for HOI detection, HICO-DET and V-COCO.	1. Introduction The task of Human-Object Interaction (HOI) detection is to discover the instances of ⟨human, object, interaction ⟩ from a given image, which reveal semantic structures of hu- man activities in the image. The results can be useful for a wide range of computer vision problems such as human action recognition [1,25,42], image retrieval [9,33,37], and image captioning [12,34,36] where a comprehensive visual understanding of the relationships between humans and ob- jects is required for high-level reasoning. With the recent success of transformer networks [31] in object detection [2, 45], transformer-based HOI detection methods [4, 15, 16, 29, 38, 44, 46] have been actively devel- oped to become a dominant base architecture for the task. Existing transformer-based methods for HOI detection can be roughly divided into two types: single-branch and two- branch. The single-branch methods [16, 29, 46] update a token set through a single transformer decoder and detect HOI instances using the subsequent FFNs directly. As a sin- gle transformer decoder is responsible for all sub-tasks ( i.e., < human,bicycle, riding> Ternary Unary Pairwise Figure 1. The illustration of relation context information in an HOI instance. We define three types of relation context information in an HOI instance: unary, pairwise, and ternary relation contexts. Each relation context provides useful information for detecting an HOI instance. For example, in our method, the unary context about an interaction (green) helps to infer that a human (yellow) and an object (red) are associated with the interaction, and vice versa. Our method utilizes the multiplex relation context consisting of the three relation contexts to perform context exchange for relational reasoning. human detection, object detection, and interaction classifi- cation), they are limited in adapting to the different sub- tasks with multi-task learning, simultaneously [38]. To re- solve the issue, the two-branch methods [4, 15, 38, 40, 44] adopt two separated transformer decoder branches where one detects human-object pairs from a human-object to- ken set while the other classifies interaction classes between human-object pairs from an interaction token set. However, the insufficient context exchange between the branches pre- vents the two-branch methods [15,38,40] from learning re- lational contexts, which plays a crucial role in identifying HOI instances. Although some methods [4, 44] tackle this issue with additional context exchange, they are limited to propagating human-object context to interaction context. To address the problem, we introduce the MUtiplex 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. 2925 RElation Network (MUREN) that performs rich context ex- change using unary, pairwise, and ternary relations of hu- man, object, and interaction tokens for relational reasoning. As illustrated in Figure 1, we define three types of relation context information in an HOI instance: unary, pairwise, and ternary, each of which provides useful information to discover HOI instances. The ternary relation context gives holistic information about the HOI instance while the unary and pairwise relation contexts provide more fine-grained in- formation about the HOI instance. For example, as shown in Figure 1, the unary context about an interaction ( e.g., ‘rid- ing’) helps to infer which pair of a human and an object is associated with the interaction in a given image, and the pairwise context between a human and an interaction ( e.g., ‘human’ and ‘riding’) helps to detect an object ( e.g., ‘bicy- cle’). Motivated by this, our multiplex relation embedding module constructs the context information that consists of the three relation contexts, thus effectively exploiting their benefits for relational reasoning. Since each sub-task re- quires different context information for relational reason- ing, our attentive fusion module selects requisite context in- formation for each sub-task from multiplex relation context and propagates the selected context information for con- text exchange between the branches. Unlike previous meth- ods [4, 15, 38, 44], we adopt three decoder branches which are responsible for human detection, object detection, and interaction classification, respectively. Therefore, the pro- posed method learns discriminative representation for each sub-task. We evaluate MUREN on two public benchmarks, HICO- DET [3] and V-COCO [10], showing that MUREN achieves state-of-the-art performance on two benchmarks. The abla- tion study demonstrates the effectiveness of the multiplex relation embedding module and the attentive fusion mod- ule. Our contribution can be summarized as follows: • We propose multiplex relation embedding module for HOI detection, which generates context information using unary, pairwise, and ternary relations in an HOI instance. • We propose the attentive fusion module that effectively propagates requisite context information for context exchange. • We design a three-branch architecture to learn more discriminative features for sub-tasks, i.e., human de- tection, object detection, and interaction classification. • Our proposed method, dubbed MUREN, outperforms state-of-the-art methods on HICO-DET and V-COCO benchmarks.
Lee_TTA-COPE_Test-Time_Adaptation_for_Category-Level_Object_Pose_Estimation_CVPR_2023	Abstract Test-time adaptation methods have been gaining atten- tion recently as a practical solution for addressing source- to-target domain gaps by gradually updating the model without requiring labels on the target data. In this paper, we propose a method of test-time adaptation for category-level object pose estimation called TTA-COPE. We design a pose ensemble approach with a self-training loss using pose- aware confidence. Unlike previous unsupervised domain adaptation methods for category-level object pose estima- tion, our approach processes the test data in a sequential, online manner, and it does not require access to the source domain at runtime. Extensive experimental results demon- strate that the proposed pose ensemble and the self-training loss improve category-level object pose performance dur- ing test time under both semi-supervised and unsupervised settings.	1. Introduction Object pose estimation is a crucial problem in com- puter vision and robotics. Advanced methods that fo- cus on diverse variations of object 6D pose estimation have been introduced, such as known 3D objects (instance- level) [28, 38], category-level [18, 36, 43], few-shot [52], and zero-shot pose estimation [13, 47]. These techniques are useful for downstream applications requiring an on- line operation, such as robotic manipulation [6, 25, 48] and augmented reality [23, 24, 32]. Our paper focuses on the category-level object pose estimation problem since it is more broadly applicable than the instance-level problem. Many works on category-level object pose estimation [2, 3, 17, 18, 36, 43, 44] have been proposed recently. These approaches estimate multiple classes of object pose more efficiently in a single network compared to the instance- level object pose estimation methods [27, 38, 41, 49–51], which depend on known 3D shape knowledge and the size of the objects. Notably, Wang et al . [43] introduced a novel representation called Normalized Object Coordinate Space (NOCS) to align various object instances within each Figure 1. We propose a Test-Time Adaptation for Category-level Object Pose Estimation framework (TTA-COPE) that automati- cally improves the network in an online manner without labeled target data. As new image frames are processed, our method fine- tunes the network using the unlabeled data and simultaneously ap- plies the network to perform pose estimation via inference. This approach successfully handles domain shifts compared with no adaptation, as seen here. category in a canonical 3D space. The strengths of the NOCS representation have led to its adoption by follow-up work [3, 17, 36]. In order to obtain accurate category-level object pose methods in unseen real-world scenarios, it is desirable to fine-tune the models in the new environment with labeled target data. The model that is not fine-tuned on the tar- get domain distribution will almost certainly exhibit lower performance than the fine-tuned model [37]. However, an- notating 6D poses of objects in the target environment is an expensive process [1, 39, 43, 45] that we seek to avoid. 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. 21285 Table 1. Comparison with prior unsupervised works for category-level object pose estimation. Our unsupervised method trains models without 2D or 3D labels of target data, similar to Self-DPDN [16]. Unlike previous methods, our proposed ap- proach updates the model online without offline batch processing. Moreover, we do not use the source data during test time (source- free) because it is impractical to train on a large amount of source data every iteration. There also may be privacy or legal constraints to access source data [21]. MethodUnsupervised Test-time Adaptation Target 3D Target 2D Source-FreeOnline Adaptation Supervised ✗ ✗ ✗ ✗ SSC-6D [29] ✓ ✗ ✗ ✗ RePoNet [5] ✓ ✗ ✗ ✗ UDA-COPE [15] ✓ ✗ ✓ ✗ Self-DPDN [16] ✓ ✓ ✗ ✗ Ours ✓ ✓ ✓ ✓ Compared to annotating in 2D space, labeling in 3D space requires specific knowledge about geometry [7] from the annotator and is much more laborious, time-consuming, and error-prone due to the complex nature of SE(3)space. Therefore, it is usually challenging to annotate real-world data with 3D annotations for fine-tuning. In order to solve the aforementioned problem of anno- tating object pose data in the real world, several recent methods [5, 15, 16] propose unsupervised domain adapta- tion (UDA) that aims to train the network without utilizing the ground truth of target pose labels. Although they show promising results using UDA techniques, these approaches still do not meet some of the requirements for online ap- plications. For example, when a robot encounters a new environment, it is desirable to adapt the scene online man- ner while estimating object poses rather than waiting for enough data to be collected in the novel scene to train the model offline. This problem definition of online fine-tuning is more practical for real applications, where we desire to update the model instantly when new data becomes available for fast domain adaptation. This setting is known as test-time adaptation (TTA) [42]. For TTA, the requirements are as follows: 1) labeled source data should not be accessed at test time, 2) adaptation should be online (rather than of- fline batch processing), and 3) the method should be fully unsupervised, without using 2D or 3D target labels during online fine-tuning. Since we do not have access to labeled source data (source-free) at test time this problem is more challenging than existing unsupervised category-level ob- ject pose methods [5,15,16,29]. Table 1 summarizes the dif- ference between our problem definition and existing meth- ods, showing that test-time adaptation for category-level ob- ject pose estimation remains an open problem. In this paper, we propose Test-time Adaptation for Category-level Object Pose Estimation ( TTA-COPE ) tohandle domain shifts without any target domain annota- tions (see Fig. 1). Prior works on general test-time adapta- tion [42,46] propose self-training to minimize entropy loss. TENT [42] has shown improvement in 2D classification and segmentation tasks. We show, however, that simply extend- ing TENT for the category-level object pose estimation is not effective. Another self-training strategy is the teacher- student framework [35] with pseudo labels. However, since pseudo labels are created without any noise filtering, naive pseudo labels may be unreliable and cause convergence to a suboptimal model. To tackle this problem, we design a novel pose ensemble method to perform test-time adaptation for category-level object pose estimation by extending the pose-aware filter- ing of UDA-COPE [15]. The proposed method uses an en- semble of teacher-student predictions based on pose-aware confidence, which is used both for generating pseudo labels and inference. Also, the pose ensemble helps to train mod- els with additional self-training loss to reduce the domain shift for category-level pose estimation by using pose-aware confidence. We demonstrate the advantages of our proposed pose ensemble and self-training loss with extensive stud- ies in both semi-supervised and unsupervised settings. We show that our TTA-COPE framework achieves state-of-the- art performance compared to strong TTA baselines. In summary, the main contributions of our work are as follows: • We propose Test-Time Adaptation for Category-level Object Pose Estimation (TTA-COPE), which handles domain shifts without labeling target data and without accessing source data during test time. • We introduce a pose ensemble with self-training loss that utilizes the teacher-student predictions to generate robust pseudo labels and estimates accurate poses for inference. • We evaluate our framework with experimental com- parisons against strong test-time baselines and state- of-the-art methods under both semi-supervised and un- supervised settings.
Koryakovskiy_One-Shot_Model_for_Mixed-Precision_Quantization_CVPR_2023	Abstract Neural network quantization is a popular approach for model compression. Modern hardware supports quanti- zation in mixed-precision mode, which allows for greater compression rates but adds the challenging task of search- ing for the optimal bit width. The majority of existing searchers find a single mixed-precision architecture. To select an architecture that is suitable in terms of perfor- mance and resource consumption, one has to restart search- ing multiple times. We focus on a specific class of methods that find tensor bit width using gradient-based optimization. First, we theoretically derive several methods that were em- pirically proposed earlier. Second, we present a novel One- Shot method that finds a diverse set of Pareto-front architec- tures in O(1) time. For large models, the proposed method is 5 times more efficient than existing methods. We verify the method on two classification and super-resolution mod- els and show above 0.93 correlation score between the pre- dicted and actual model performance. The Pareto-front ar- chitecture selection is straightforward and takes only 20 to 40 supernet evaluations, which is the new state-of-the-art result to the best of our knowledge.	1. Introduction In recent years, neural network quantization [31] has be- come a popular hardware-friendly compression technique. It is common to quantize linear and convolutional layer operands while leaving vector operands unchanged. Mod- ern algorithms achieve lossless quantization into fixed 8-bit integer values in many applications [45, 15, 40, 35, 49, 25, 5]. At higher compression rates, mixed-precision is often needed [22, 44]. For example, models often require 8-bit precision for the first and last layers, while the middle lay- ers can tolerate lower precision [15, 40]. In addition, the selected precision may depend on a quantized operation [7] or a hardware at hand [41]. This motivates many vendors to010Time (m)ESPCN 050SRResNet 0200ResNet18 0250MobileNet-v2 EdMIPS DNAS GMPQ One-Shot MPS Figure 1. The searching time taken by each algorithm to dis- cover a single bit width architecture belonging to a Pareto front. EdMIPS [7], DNAS [44], GMPQ [42], and One-Shot MPS (our) use a proxy dataset for ResNet-18 and MobileNet-v2. Bayesian Bits [38] and HAQ [41] roughly take 100 times more searching time compared to our method. support mixed-precision models in hardware. To attain the best mixed-precision performance, it is cru- cial to find an optimal precision for each matrix multipli- cation factor. Unfortunately, all possible bit width combi- nations cannot be examined since the search space scales exponentially with the number of multiplications. The in- ability to predict the influence of individual loss coefficients on the resulting compression rate furthermore exacerbates the difficulty of searching. Existing methods [44, 7, 38, 9] require multiple restarts of the searching process until a sat- isfactory bit width allocation is found. This results in O(N) searching time, where Nis the number of restarts. The authors of EdMIPS [7] mention that “sometimes, it is mysterious why and how an architecture is found by Neu- ral Architecture Search (NAS)”. We answer this question in the context of EdMIPS and DNAS [44] methods. To do so, we simplify and generalize Bayesian Bits [38], where a variational inference (VI) approach is used to derive the loss function for a hierarchical supernet. Then, we demonstrate how the EdMIPS and DNAS loss functions can be derived. Next, using our derivation, we propose a novel One- Shot Mixed-Precision Search (One-Shot MPS) method that finds a diverse set of Pareto-front architectures in O(1) time. We extend the commonly used supernet transforma- 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. 7939 tion of a floating-point model with a set of trainable func- tions that predict the bit width probability depending on a hardware regularization parameter. For ImageNet [11], we observe at least 0.96correlation score between child mod- els sampled from a One-Shot model and standalone fine- tuned models, cf. previous result attains a maximum score of0.55[16]. Such a high score allows one to plot a Pareto front of performance versus hardware resources using a lin- ear sweep over the regularization parameter, and select the most promising precision given hardware constraints be- fore fine-tuning . The Pareto-front architecture selection is straightforward and takes only 20 to 40 supernet evaluations while existing One-Shot methods require at least 1000 eval- uations [16, 10]. To sum up, our contribution is twofold. First, we pro- vide a theoretical derivation of the earlier empirically-found state-of-the-art searching methods. Second, we propose to augment a supernet with a bit width prediction model that allows searching for Pareto-front bit width combinations corresponding to different compression rates in a constant time. We validate the benefits of the approach on sev- eral widely-used models including mobile-friendly archi- tectures. To the best of our knowledge, the proposed pre- dictor is not described in the existing literature.
Lee_Shape-Aware_Text-Driven_Layered_Video_Editing_CVPR_2023	Abstract Temporal consistency is essential for video editing appli- cations. Existing work on layered representation of videos allows propagating edits consistently to each frame. These methods, however, can only edit object appearance rather than object shape changes due to the limitation of using a fixed UV mapping field for texture atlas. We present a shape-aware, text-driven video editing method to tackle this challenge. To handle shape changes in video editing, we first propagate the deformation field between the in- put and edited keyframe to all frames. We then leverage a pre-trained text-conditioned diffusion model as guidance for refining shape distortion and completing unseen regions. The experimental results demonstrate that our method can achieve shape-aware consistent video editing and compare favorably with the state-of-the-art.	1. Introduction Image editing. Recently, image editing [19, 20, 24, 34, 40, 44] has made tremendous progress, especially those using diffusion models [19, 20, 40, 44]. With free-form text prompts, users can obtain photo-realistic edited images without artistic skills or labor-intensive editing. However, unlike image editing, video editing is more challenging due to the requirement of temporal consistency. Independently editing individual frames leads to undesired inconsistent frames, as shown in Fig. 2a. A na ¨ıve way to deal with tem- poral consistency in video editing is to edit a single frame and then propagate the change to all the other frames. Nev- ertheless, artifacts are presented when there are unseen pix- els from the edited frame in the other frames, as shown in Fig. 2b. Video editing and their limitations. For consistent video editing, Neural Layered Atlas (NLA) [18] decomposes a video into unified appearance layers atlas . The layered de- composition helps consistently propagate the user edit 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. 14317 (a) Multi-frame editing with frame interpolation [42] (b) Single-frame editing with frame propagation [17] (c) Text2LIVE [2] with prompt “ sports car ” Figure 2. Limitation of existing work. Compare these results from baseline methods with our “ sports car ” result in Fig. 1. (a) Multiple frames are edited independently and interpolated by frame interpolation method [42]. Such an approach shows realistic per-frame results but suffers from temporal flickering. (b) Extracting a single keyframe for image editing, the edits are propagated to each frame via [17]. The propagated edits are temporally stable. However, it yields visible distortions due to the unseen pixels from the keyframe. (c) The SOTA Text2LIVE [2] results demonstrate temporally-consistent appearance editing but remain the source shape “Jeep ” instead of the target prompt “ sports car ” by using the fixed UV mapping of NLA. individual frames with per-frame UV sampling association. Based on NLA, Text2LIVE [2] performs text-driven editing on atlases with the guidance of the Vision-Language model, CLIP [39]. Although Text2LIVE [2] makes video editing easier with a text prompt, it can only achieve appearance manipulation due to the use of fixed-shape associated UV sampling. Since per-frame UV sampling gathers informa- tion on motion and shape transformation in each frame to learn the pixel mapping from the atlas, shape editing is not feasible, as shown in Fig. 2c. Our work. In this paper, we propose a shape-aware text- guided video editing approach. The core idea in our work lies in a novel UV map deformation formulation. With a se- lected keyframe and target text prompt, we first generate an edited frame by image-based editing tool ( e.g., Stable Diffu- sion [44]). We then perform pixel-wise alignment between the input and edited keyframe pair through a semantic cor- respondence method [51]. The correspondence specifies the deformation between the input-edited pair at the keyframe. According to the correspondence, the shape and appearance change can then be mapped back to the atlas space. We can thus obtain per-frame deformation by sampling the defor- mation from the atlas to the original UV maps. While this method helps with shape-aware editing, it is insufficient due to unseen pixels in the edited keyframe. We tackle this byfurther optimizing the atlas texture and the deformation us- ing a pretrained diffusion model by adopting the gradient update procedure described in DreamFusion [38]. Through the atlas optimization, we achieve consistent shape andap- pearance editing, even in challenging cases where the mov- ing object undergoes 3D transformation (Fig. 1). Our contributions. • We extend the capability of existing video editing methods to enable shape-aware editing. • We present a deformation formulation for frame- dependent shape deformation to handle target shape edits. • We demonstrate the use of a pre-trained diffusion model for guiding atlas completion in layered video representation.
Lin_Being_Comes_From_Not-Being_Open-Vocabulary_Text-to-Motion_Generation_With_Wordless_Training_CVPR_2023	Abstract Text-to-motion generation is an emerging and challeng- ing problem, which aims to synthesize motion with the same semantics as the input text. However, due to the lack of di- verse labeled training data, most approaches either limit to specific types of text annotations or require online optimiza- tions to cater to the texts during inference at the cost of ef- ficiency and stability. In this paper, we investigate offline open-vocabulary text-to-motion generation in a zero-shot learning manner that neither requires paired training data nor extra online optimization to adapt for unseen texts. In- spired by the prompt learning in NLP , we pretrain a motion generator that learns to reconstruct the full motion from the masked motion. During inference, instead of chang- ing the motion generator, our method reformulates the in- put text into a masked motion as the prompt for the motion generator to “reconstruct” the motion. In constructing the prompt, the unmasked poses of the prompt are synthesized by a text-to-pose generator. To supervise the optimization of the text-to-pose generator, we propose the first text-pose alignment model for measuring the alignment between texts and 3D poses. And to prevent the pose generator from over- fitting to limited training texts, we further propose a novel wordless training mechanism that optimizes the text-to-pose generator without any training texts. The comprehensive experimental results show that our method obtains a signif- icant improvement against the baseline methods. The code is available at https://github.com/junfanlin/ oohmg .	1. Introduction Motion generation has attracted increasing attention due to its practical value in the fields of virtual reality, video games, and movies. Especially for text-conditional motion generation, it can largely improve the user experience if the virtual avatars can react to the communication texts in real time. However, most current text-to-motion approaches *Corresponding author: [email protected] Figure 1. Demonstrations of our OOHMG. Given an unseen open- vocabulary text (e.g., an object name “a basketball”, or a simile description “fly like a bird”, or a usual text “he walks”), OOHMG translates the text into the text-consistent pose, which is used to prompt the motion generator for synthesizing the motion. are trained on paired text-motion data with limited types of annotations, and thus could not well-generalize to unseen open-vocabulary texts. To handle the open-vocabulary texts, recent works lever- age the powerful zero-shot text-image alignment ability of the pretrained model, i.e., CLIP [35], to facilitate the text- to-motion generation. Some works like MotionCLIP [42] use the CLIP text encoder to extract text features and learn a motion decoder to decode the features into motions. How- ever, they require paired text-motion training data and still could not handle texts that are dissimilar to the training texts. Instead of learning an offline motion generator with paired data, some works like AvatarCLIP [13] generate mo- tions for the given textual descriptions via online matching and optimization. Nevertheless, matching cannot generate new poses to fit diverse texts and online optimization is usu- ally time-consuming and unstable. In this paper, we investigate filling the blank of offline open-vocabulary text-to-motion generation in a zero-shot learning manner. For convenience, we term our method asOOHMG which stands for Offline Open-vocabulary Human Motion Generation. The main philosophy 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. 23222 Figure 2. The sketch of OOHMG. A text is fed to the text-to-pose generator to obtain a text-consistent pose. Then, the pose is used to construct the motion prompt for the pretrained motion model to generate a motion. OOHMG is inspired by prompt learning [5,19,41,48,50,52] in the field of natural language processing (NLP). Specifi- cally, instead of changing the pretrained motion generator to cater to the given texts online, OOHMG reformulates the texts into a familiar input format to prompt the pre- trained motion generator for synthesizing motions in the manner of “reconstruction”. As for prompt construction, OOHMG learns a text-to-pose generator using the novel wordless training mechanism so that the pose generator can generalize to unseen texts during inference. After train- ing, OOHMG uses the text-to-pose generator to translate texts into poses to construct the prompt. The overall sketch and demonstrations of OOHMG are illustrated in Fig. 2 and Fig. 1, respectively. In this sense, the two key ingredients of OOHMG include the motion generator pretraining and the prompt construction for open-vocabulary texts. In the fol- lowing, we further elaborate on each of these ingredients. As for the motion generator, we learn a motion genera- tor by mask-reconstruction self-supervised learning. Par- ticularly, our method adopts a bidirectional transformer- based [44] architecture for the motion generator. During training, the motion generator takes the randomly-masked motions as inputs and is optimized to reconstruct the orig- inal motions. To predict and reconstruct the masked poses from the unmasked, the motion generator is required to fo- cus on learning motion dynamics which is the general need for diverse motion generation tasks. By this means, unlike previous methods that design different models for different tasks [1,2,11,18], our motion model can be directly applied to diverse downstream tasks by unifying the input of these tasks into masked motions to prompt the generator for mo- tion generation. Moreover, our generator can flexibly con- trol the generated content, such as the number, the order, and the positions of different poses of the generated motion by editing the masked motions, resulting in a controllable and flexible motion generation. In constructing the motion prompt for open-vocabulary motion generation, OOHMG learns a text-to-pose generator and uses it to generate the unmasked poses of the masked motions, as shown in Fig. 2. There are two major diffi-culties in learning the text-to-pose generator: 1) what can associate diverse texts and poses to supervise the pose gen- erator, and 2) how to obtain diverse texts as the training in- puts. For difficulty 1, we build the first large-scale text-pose alignment model based on CLIP, namely TPA, that can effi- ciently measure the alignment between texts and 3D SMPL poses [27, 31] in the feature space. With TPA, the text-to- pose generator learns to generate poses for texts by maxi- mizing the text-pose alignments via gradient descent. As for difficulty 2, instead of collecting massive texts laboriously for training, we consider an extreme training paradigm, termed wordless training. Just as its name implies, word- less training only samples random training inputs from the latent space of texts. And we found that the optimized pose generator can well-generalize to real-world texts. Overall, the contributions of OOHMG are as follows. 1) We propose an offline open-vocabulary text-to-motion gen- eration framework, inspired by prompt learning, and 2) to supervise the training process of the text-to-pose generator, we propose the first text-pose alignment model, i.e., TPA, and 3) to endow the text-to-pose generator with the ability to handle open-vocabulary texts, we train the generator with the novel wordless training mechanism. 4) Extensive exper- iment results show that OOHMG is able to generate motions for open-vocabulary texts efficiently and effectively, and ob- tain clear improvement over the advanced baseline methods qualitatively and quantitatively.
Liu_MixTeacher_Mining_Promising_Labels_With_Mixed_Scale_Teacher_for_Semi-Supervised_CVPR_2023	Abstract Scale variation across object instances remains a key challenge in object detection task. Despite the remarkable progress made by modern detection models, this challenge is particularly evident in the semi-supervised case. While existing semi-supervised object detection methods rely on strict conditions to filter high-quality pseudo labels from network predictions, we observe that objects with extreme scale tend to have low confidence, resulting in a lack of positive supervision for these objects. In this paper, we propose a novel framework that addresses the scale vari- ation problem by introducing a mixed scale teacher to im- prove pseudo label generation and scale-invariant learn- ing. Additionally, we propose mining pseudo labels using score promotion of predictions across scales, which bene- fits from better predictions from mixed scale features. Our extensive experiments on MS COCO and PASCAL VOC benchmarks under various semi-supervised settings demon- strate that our method achieves new state-of-the-art per- formance. The code and models are available at https: //github.com/lliuz/MixTeacher .	1. Introduction The remarkable performance of deep learning on various tasks can largely be attributed to large-scale datasets with accurate annotations. However, collecting a large amount of high-quality annotations is infeasible as it is labor-intensive and time-consuming, especially for tasks with complicated annotations such as object detection [23, 30] and segmen- tation [5, 6]. To reduce reliance on manual labeling, semi- supervised learning (SSL) has gained much attention. SSL aims to train models on a small amount of labeled im- ages and a large amount of easily accessible unlabeled data. †Corresponding Authors. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 [email protected]_1x Recall_0.5x 0.30.40.50.60.70.80.91.0 [email protected]_1x Precision_0.5x 0.5 0.6 0.7 0.8 0.9 [email protected]_1x Recall_0.5x 0.650.700.750.800.850.900.95 [email protected]_1x Precision_0.5x (a) Precision and recall for all objects (b) Precision and recall for large objects (c) Detection results with different input scalesFigure 1. Detection results with input of regular 1 ×scale and 0.5×down-sampled scale images. We plot the precision and recall under different score thresholds for (a) all objects and (b) large objects in COCO val2017 with the same model but different input scales. Two examples of unlabeled images are given in (c). Large scale inputs have clear advantages in overall metrics, but down-sampled images are more suitable for large objects. Following extensive pioneering studies on semi-supervised image classification [2, 14, 32], several methods on semi- supervised object detection have emerged. Most early studies on semi-supervised object detec- tion [13, 24, 33] can be considered as a direct extension of SSL methods designed for image classification, using a teacher-student training paradigm [2,32,35]. In these meth- ods, a teacher model generates pseudo bounding boxes and corresponding class predictions on unlabeled images, and the pseudo labels are used to train a student model. Despite the performance improvement from using a large amount of unlabeled data, these methods overlooked the characteris- tics of object detection to some extent, resulting in a huge 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. 7370 gap from the fully supervised counterpart. Compared to image classification, object instances in de- tection tasks can vary in a wider range of scales. To address this challenge of detecting and localizing multiple objects across scales and locations, numerous works in object de- tection have been proposed, such as FPN [21], Trident [20], and SNIP [31]. However, the large scale variation brings new challenges in the semi-supervised context. In order to guarantee high precision, most existing semi-supervised ob- ject detection methods adopt strict conditions ( e.g. score > 0.9) to filter out highly confident pseudo labels. Although this ensures the quality of pseudo labels, many objects with low confidence are wrongly assigned as background, es- pecially for those with extreme scales. As shown in Fig- ure 1 (c), inappropriate scales will lead to false negatives, which can mislead the network in semi-supervised learn- ing. We further observe the influence of the test scale of the images. Consistent with common sense, large-scale inputs have clear advantages in overall metrics, as shown in Fig- ure 1 (a). However, down-sampled images show a superior- ity for large objects, as shown in Figure 1 (b). This provides a new view to handle the scale variation issue. It is worth mentioning that recent works have paid at- tention to the scale variation issue in semi-supervised ob- ject detection. As shown in Figure 2 (a) and (b), previ- ous methods have introduced an additional down-sampled view to encourage the model to make scale-invariant pre- dictions. Specifically, SED [10] proposes to distill predic- tions of class probability from the regular scale to the down- sampled scale and constrain consistent predictions of local- ization for all proposals in two scales. PseCo [17] adopts the same pseudo labels generated from the regular scale for both scales. However, these methods mainly focus on the consistency of predictions across scales, which indirectly improves the models with regularization. Moreover, they highly rely on the pseudo labels generated from the regular scale in the teacher network. The false negatives caused by inappropriate scales still remain in these methods. Based on the above methods, which are equipped with an additional down-sampled view of unlabeled images, we propose to explicitly improve the quality of pseudo la- bels to handle the scale variation of objects. As shown in Figure 2 (c), we introduce a mixed-scale feature pyra- mid, which is built from the large-scale feature pyramid in the regular view and the small-scale feature pyramid in the down-sampled view. The mixed-scale feature pyramid is supposed to be capable of adaptively fusing features across scales, thus making better predictions in the teacher net- work. Furthermore, to avoid object instances missing in the pseudo labels due to low confidence scores, we propose to leverage the improvement of score as an indicator for min- ing pseudo labels from low confidence predictions. In sum- mary, the main contributions are as follows:• We propose a semi-supervised object detection frame- work MixTeacher, in which high-quality pseudo labels are generated from a mixed scale feature pyramid. • We propose a method for pseudo labels mining, which leverages the improvement of predictions as the indi- cator to mining the promising pseudo labels. • Our method achieves state-of-the-art performance on MS COCO and Pascal VOC benchmarks under various semi-supervised settings.
Li_GLIGEN_Open-Set_Grounded_Text-to-Image_Generation_CVPR_2023	Abstract Large-scale text-to-image diffusion models have made amazing advances. However, the status quo is to use text input alone, which can impede controllability. In this work, we propose GLIGEN ,Grounded- Language-to- Image Generation, a novel approach that builds upon and extends the functionality of existing pre-trained text-to-image dif- fusion models by enabling them to also be conditioned on grounding inputs. To preserve the vast concept knowledge ofthe pre-trained model, we freeze all of its weights and inject the grounding information into new trainable layers via a gated mechanism. Our model achieves open-world grounded text2img generation with caption and bounding box condi- tion inputs, and the grounding ability generalizes well to novel spatial configurations and concepts. GLIGEN ’s zero- shot performance on COCO and LVIS outperforms existing supervised layout-to-image baselines by a large margin. §Part of the work performed at Microsoft; ¶Co-senior authors 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. 22511	1. Introduction Image generation research has witnessed huge advances in recent years. Over the past couple of years, GANs [13] were the state-of-the-art, with their latent space and con- ditional inputs being well-studied for controllable manipu- lation [42, 54] and generation [25, 27, 41, 75]. Text condi- tional autoregressive [46, 67] and diffusion [45, 50] models have demonstrated astonishing image quality and concept coverage, due to their more stable learning objectives and large-scale training on web image-text paired data. These models have gained attention even among the general public due to their practical use cases ( e.g., art design and creation). Despite exciting progress, existing large-scale text-to- image generation models cannot be conditioned on other input modalities apart from text, and thus lack the ability to precisely localize concepts, use reference images, or other conditional inputs to control the generation process. The cur- rent input, i.e., natural language alone, restricts the way that information can be expressed. For example, it is difficult to describe the precise location of an object using text, whereas bounding boxes / keypoints can easily achieve this, as shown in Figure 1. While conditional diffusion models [9, 47, 49] and GANs [24, 33, 42, 64] that take in input modalities other than text for inpainting, layout2img generation, etc., do exist, they rarely combine those inputs for controllable text2img generation. Moreover, prior generative models—regardless of the generative model family—are usually independently trained on each task-specific dataset. In contrast, in the recognition field, the long-standing paradigm has been to build recogni- tion models [29, 37, 76] by starting from a foundation model pretrained on large-scale image data [4,15,16] or image-text pairs [30, 44, 68]. Since diffusion models have been trained on billions of image-text pairs [47], a natural question is: Can we build upon existing pretrained diffusion models and endow them with new conditional input modalities? In this way, analogous to the recognition literature, we may be able to achieve better performance on other generation tasks due to the vast concept knowledge that the pretrained models have, while acquiring more controllability over existing text- to-image generation models. With the above aims, we propose a method for providing new grounding conditional inputs to pretrained text-to-image diffusion models. As shown in Figure 1, we still retain the text caption as input, but also enable other input modalities such as bounding boxes for grounding concepts, grounding reference images, grounding part keypoints, etc. The key challenge is preserving the original vast concept knowledge in the pretrained model while learning to inject the new grounding information. To prevent knowledge forgetting, we propose to freeze the original model weights and add new trainable gated Transformer layers [61] that take in the new grounding input ( e.g., bounding box). During training,we gradually fuse the new grounding information into the pretrained model using a gated mechanism [1]. This design enables flexibility in the sampling process during generation for improved quality and controllability; for example, we show that using the full model (all layers) in the first half of the sampling steps and only using the original layers (without the gated Transformer layers) in the latter half can lead to generation results that accurately reflect the grounding conditions while also having high image quality. In our experiments, we primarily study grounded text2img generation with bounding boxes, inspired by the recent scaling success of learning grounded language-image understanding models with boxes in GLIP [31]. To en- able our model to ground open-world vocabulary con- cepts [29,31,69,72], we use the same pre-trained text encoder (for encoding the caption) to encode each phrase associated with each grounded entity ( i.e., one phrase per bounding box) and feed the encoded tokens into the newly inserted layers with their encoded location information. Due to the shared text space, we find that our model can generalize to unseen objects even when only trained on the COCO [36] dataset. Its generalization on LVIS [14] outperforms a strong fully-supervised baseline by a large margin. To further im- prove our model’s grounding ability, we unify the object detection and grounding data formats for training, following GLIP [31]. With larger training data, our model’s general- ization is consistently improved. Contributions. 1) We propose a new text2img genera- tion method that endows new grounding controllability over existing text2img diffusion models. 2) By preserving the pre- trained weights and learning to gradually integrate the new localization layers, our model achieves open-world grounded text2img generation with bounding box inputs, i.e., synthesis of novel localized concepts unobserved in training. 3) Our model’s zero-shot performance on layout2img tasks signifi- cantly outperforms the prior state-of-the-art, demonstrating the power of building upon large pretrained generative mod- els for downstream tasks.
Li_AShapeFormer_Semantics-Guided_Object-Level_Active_Shape_Encoding_for_3D_Object_Detection_CVPR_2023	Abstract 3D object detection techniques commonly follow a pipeline that aggregates predicted object central point fea- tures to compute candidate points. However, these can- didate points contain only positional information, largely ignoring the object-level shape information. This eventu- ally leads to sub-optimal 3D object detection. In this work, we propose AShapeFormer, a semantics-guided object-level shape encoding module for 3D object detection. This is a plug-n-play module that leverages multi-head attention to encode object shape information. We also propose shape tokens and object-scene positional encoding to ensure that the shape information is fully exploited. Moreover, we in- troduce a semantic guidance sub-module to sample more foreground points and suppress the influence of background points for a better object shape perception. We demon- strate a straightforward enhancement of multiple existing methods with our AShapeFormer. Through extensive exper- iments on the popular SUN RGB-D and ScanNetV2 dataset, we show that our enhanced models are able to outperform the baselines by a considerable absolute margin of up to 8.1%. Code will be available at https://github. com/ZechuanLi/AShapeFormer	1. Introduction As an important scene understanding task, 3D object de- tection [13,20,47] aims to detect 3D bounding boxes and se- mantic categories in 3D point cloud scenes. It plays an im- portant role in many downstream tasks, such as augmented reality [2, 3], mobile robots [18, 41, 52], and autonomous navigation [1,36,37,39]. Object detection has made signifi- cant progress in the 2D domain [15,22,33]. However, owing to the sparse and irregular nature of the point cloud data, 2D detection techniques are generally not readily applicable to the 3D object detection task. *Corresponding author Figure 1. ( Top) V oteNet [29] seed points contain many back- ground points, leading to sub-optimal candidate points, which are also intrinsically weak as they fail to account for object shape and contour features. ( Bottom ) The proposed AShapeFormer se- lects more relevant seed points, leading to more appropriate can- didates that additionally encode the object shape information ac- tively. This results in high quality 3D object detection. Inspired by their 2D counterparts, early attempts in 3D object detection, e.g., [16, 39], mapped irregular point clouds to regular 3D voxels, thereafter using 3DCNNs for feature extraction and object detection. However, voxeliza- tion inevitably loses fine-grained information of the point clouds, which adversely affects the detection performance. With the advances that allow direct processing of the point clouds with deep neural models, e.g., [30, 31], recent meth- ods aim at directly predicting the 3D bounding boxes from the original unordered point clouds. Among these tech- niques, V oteNet [29] and its variants [7, 12, 28, 45, 46, 50] have achieved remarkable performance. These point-wise methods follow a common underlying pipeline which includes first aggregating certain predicted point features into candidate points. The candidate points are later used to estimate the 3D bounding box information, e.g., center, size, and orientation, along with the associated semantic labels. As illustrated in Fig. 1, despite their ex- cellent performance, these methods still face a few major 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. 1012 challenges. (1) The final prediction relies strongly on the quality of the candidate points. However, these points fail to encode important object-level features such as contours and the shape of the 3D objects. (2) The methods must regress over the candidate points, and these points are often influenced by the background points. This propagates the error to cause offsets in the eventual predictions. Current attempts to mitigate these issues rely on generating new fea- tures [7, 45] or sampling more points [42]. However, these are resource intensive solutions, which must still rely on the vote point regression quality. To address the problems, we introduce a novel plug-n- play neural module, named AShapeFormer. It can be easily assembled with many existing 3D object detection methods to provide a considerable performance boost. Our key driv- ing insight is that by utilizing implicit object-level shape features, a detector can be made aware of the object shape distribution. Specifically, our module utilizes multi-head at- tention to encode the object shape information. We aggre- gate the object shape features using a self-attention mecha- nism. Inspired by ViT [10] and BERT [19], we introduce a shape token as the output of the final shape feature to avoid information loss caused by simplistic operations, e.g., pool- ing. Additionally, we devise a semantic guidance mecha- nism to sample more foreground points and assign different weights to their features, which improves the shape feature generation. Semantic segmentation scores are also utilized during the aggregation of vote points to reduce the influence of irrelevant vote points and obtain better candidates. We provide successful demonstration of boosting both point-based [28, 29] and Transformer [23] baselines with our method, achieving strong performance gains. Our ex- perimental results (§ 4.1) show that AShapeFormer boosts the multi-class mean average precision (mAP) up to 3.5% on the challenging SUN RGB-D dataset [38] and 8.1% on the ScanNet V2 dataset [9]. Highlights of our contributions include the following. • We propose a plug-and-play active shape encoding module named AShapeFormer, which can be com- bined with many existing 3D object detection networks to achieve a considerable performance boost. • To the best of our knowledge, our method is the first to combine multi-head attention and semantic guidance to encode strong object shape features for robust clas- sification and accurate bounding box regression. • We demonstrate a considerable mAP boost on SUN RGB-D ([email protected]) and ScanNet V2 datasets by en- hancing the state-of-the-art methods with our module.
Lee_Revisiting_Self-Similarity_Structural_Embedding_for_Image_Retrieval_CVPR_2023	Abstract Despite advances in global image representation, exist- ing image retrieval approaches rarely consider geometric structure during the global retrieval stage. In this work, we revisit the conventional self-similarity descriptor from a convolutional perspective, to encode both the visual and structural cues of the image to global image representation. Our proposed network, named Structural Embedding Net- work (SENet), captures the internal structure of the images and gradually compresses them into dense self-similarity descriptors while learning diverse structures from various images. These self-similarity descriptors and original im- age features are fused and then pooled into global embed- ding, so that global embedding can represent both geomet- ric and visual cues of the image. Along with this novel structural embedding, our proposed network sets new state- of-the-art performances on several image retrieval bench- marks, convincing its robustness to look-alike distractors. The code and models are available: https://github. com/sungonce/SENet .	1. Introduction Content-based image retrieval is the task of searching for images with the same content present in the query im- age in the large-scale database. What across images rep- resents the same content are two things: the visual prop- erties and the geometrical structure, so comparing them well is the key to the image retrieval task. To achieve this goal, two image representation types have been exten- sively explored in many image retrieval solutions. The first one is local features [1–3, 5, 19–22, 24, 47] that comprise visual descriptors and spatial information about local re- gions of the image, and the other one is a global descriptor [3,11–13,18,24–26,28,33,43], also known as global embed- ding, that summarizes the local features of the entire image. In a general sense, the global descriptor loses spatial infor- mation of local features during the summarization process. Thus, many image retrieval solutions [3, 18, 24, 35, 36, 42] first retrieve coarse candidates with similar visual proper- *Corresponding author. Negative Image Query Image Visual Property Self- Similarity Positive Image Figure 1. Images of the same content share both similar image properties and internal self-similarities. Our proposed networks leverage both visual features and self-similarity features and en- code them to global embedding in an end-to-end manner. ties for the query using global embeddings (typically re- ferred to as global retrieval) and further verify that coarse candidates have geometrically similar shapes to the query using local features (typically referred to as local feature re- ranking). This separation of tasks may sound reasonable at first glance. However, in fact, they miss the opportunities to perform robust retrieval by comparing structural informa- tion also in the global retrieval stage. In computer vision, a self-similarity descriptor [31] has long been used as a regional descriptor for matching images based on the aggregation of local internal structures. This work has shown its effectiveness in challenging matching problems even in situations where the visual properties of images are not shared at all ( e.g. matching between draw- ing and photo domains). However, their self-similarity en- coding process is neither learnable nor differentiable. And it also completely ignores visual properties, making it dif- ficult to use directly for image retrieval tasks where visual properties are also valuable cues. In this paper, we revisit the self-similarity descriptor in a convolutional manner and propose a novel global em- bedding network named Structural Embedding Network (SENet ). The proposed network captures the internal struc- tures of the images and encodes them to self-similarity de- 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. 23412 scriptors while learning diverse structures from various im- ages. These self-similarity descriptors and original image features are fused and then pooled into global embedding, so that global embedding can represent both valuable geo- metric and visual cues of the image. All proposed modules of our networks are comprised of point-wise operations, en- abling efficient descriptor encoding. Our proposed network sets state-of-the-art performance on several image retrieval benchmarks, convincing its robustness to look-alike distrac- tors.
Lin_Magic3D_High-Resolution_Text-to-3D_Content_Creation_CVPR_2023	Abstract DreamFusion [ 31] has recently demonstrated the utility of a pre-trained text-to-image diffusion model to optimize Neural Radiance Fields (NeRF) [ 23], achieving remarkable text-to-3D synthesis results. However, the method has two in- herent limitations: (a) extremely slow optimization of NeRF and (b) low-resolution image space supervision on NeRF , leading to low-quality 3D models with a long processing time. In this paper, we address these limitations by utilizing a two-stage optimization framework. First, we obtain a coarse model using a low-resolution diffusion prior and accelerate with a sparse 3D hash grid structure. Using the coarse repre- sentation as the initialization, we further optimize a textured 3D mesh model with an efficient differentiable renderer in- teracting with a high-resolution latent diffusion model. Our method, dubbed Magic3D, can create high quality 3D mesh models in 40 minutes, which is 2×faster than DreamFu- sion (reportedly taking 1.5 hours on average), while also achieving higher resolution. User studies show 61.7% raters to prefer our approach over DreamFusion. Together with the image-conditioned generation capabilities, we provide users with new ways to control 3D synthesis, opening up new avenues to various creative applications.	1. Introduction 3D digital content has been in high demand for a variety of applications, including gaming, entertainment, architec- ture, and robotics simulation. It is slowly finding its way into virtually every possible domain: retail, online conferencing, virtual social presence, education, etc. However, creating professional 3D content is not for anyone Ð it requires immense artistic and aesthetic training with 3D modeling ex- pertise. Developing these skill sets takes a significant amount of time and effort. Augmenting 3D content creation with natural language could considerably help democratize 3D content creation for novices and turbocharge expert artists. *†: equal contribution.Image content creation from text prompts [ 2,28,33,36] has seen significant progress with the advances of diffusion models [ 13,41,42] for generative modeling of images. The key enablers are large-scale datasets comprising billions of samples (images with text) scrapped from the Internet and massive amounts of compute. In contrast, 3D content generation has progressed at a much slower pace. Existing 3D object generation models [ 4,9,47] are mostly categorical. A trained model can only be used to synthesize objects for a single class, with early signs of scaling to multiple classes shown recently by Zeng et al. [47]. Therefore, what a user can do with these models is extremely limited and not yet ready for artistic creation. This limitation is largely due to the lack of diverse large-scale 3D datasets Ð compared to image and video content, 3D content is much less accessible on the Internet. This naturally raises the question of whether 3D generation capability can be achieved by leveraging powerful text-to-image generative models. Recently, DreamFusion [ 31] demonstrated its remarkable ability for text-conditioned 3D content generation by uti- lizing a pre-trained text-to-image diffusion model [ 36] that generates images as a strong image prior. The diffusion model acts as a critic to optimize the underlying 3D repre- sentation. The optimization process ensures that rendered images from a 3D model, represented by Neural Radiance Fields (NeRF) [ 23], match the distribution of photorealis- tic images across different viewpoints, given the input text prompt. Since the supervision signal in DreamFusion oper- ates on very low-resolution images ( 64×64), DreamFusion cannot synthesize high-frequency 3D geometric and texture details. Due to the use of inefficient MLP architectures for the NeRF representation, practical high-resolution synthesis may not even be possible as the required memory footprint and the computation budget grows quickly with the resolu- tion. Even at a resolution of 64×64, optimization times are in hours (1.5 hours per prompt on average using TPUv4). In this paper, we present a method that can synthesize highly detailed 3D models from text prompts within a re- duced computation time. Specifically, we propose a coarse- 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. 300 asilver platter piled high with fruits ablue poison -dart frog sitting on a water lilyan imperial state crown of england neuschwanstein castle, aerial view astuffed grey rabbit holding a pretend carrot an iguana holding a balloon abeautiful dress made out of garbage bagsmichelangelo style statue of an astronaut ametal bunny sitting on top of a stack of broccoliasphinx sitting on top of a stack of chocolate cookieLow resolution bunny before editing ametal bunny sitting on top of a stack of chocolate cookiea baby bunny sitting on top of a stack of pancakes Figure 1. Results and applications of Magic3D. Top: high-resolution text-to-3D generation . Magic3D can generate high-quality and high-resolution 3D models from text prompts. Bottom: high-resolution prompt-based editing . Magic3D can edit 3D models by fine-tuning with the diffusion prior using a different prompt. Taking the low-resolution 3D model as the input (left), Magic3D can modify different parts of the 3D model corresponding to different input text prompts. Together with various creative controls on the generated 3D models, Magic3D is a convenient tool for augmenting 3D content creation. to-fine optimization approach that uses multiple diffusion priors at different resolutions to optimize the 3D representa- tion, enabling the generation of both view-consistent geome- try as well as high-resolution details. In the first stage, we optimize a coarse neural field representation akin to Dream- Fusion, but with a memory- and compute-efficient scene representation based on a hash grid [ 25]. In the second stage,we switch to optimizing mesh representations, a critical step that allows us to utilize diffusion priors at resolutions as high as512×512. As 3D meshes are amenable to fast graphics renderers that can render high-resolution images in real-time, we leverage an efficient differentiable rasterizer [ 9,26] and make use of camera close-ups to recover high-frequency details in geometry and texture. As a result, our approach 301 produces high-fidelity 3D content (see Fig. 1) that can con- veniently be imported and visualized in standard graphics software and does so at 2 ×the speed of DreamFusion. Fur- thermore, we showcase various creative controls over the 3D synthesis process by leveraging the advancements developed for text-to-image editing applications [ 2,35]. Our approach, dubbed Magic3D, endows users with unprecedented control in crafting their desired 3D objects with text prompts and reference images, bringing this technology one step closer to democratizing 3D content creation. In summary, our work makes the following contributions: •We propose Magic3D, a framework for high-quality 3D content synthesis using text prompts by improving several major design choices made in DreamFusion. It consists of a coarse-to-fine strategy that leverages both low- and high- resolution diffusion priors for learning the 3D representa- tion of the target content. Magic3D, which synthesizes 3D content with an 8 ×higher resolution supervision, is also 2×faster than DreamFusion. 3D content synthesized by our approach is significantly preferable by users (61.7%). •We extend various image editing techniques developed for text-to-image models to 3D object editing and show their applications in the proposed framework.
Liu_FAC_3D_Representation_Learning_via_Foreground_Aware_Feature_Contrast_CVPR_2023	Abstract Contrastive learning has recently demonstrated great potential for unsupervised pre-training in 3D scene un- derstanding tasks. However, most existing work ran- domly selects point features as anchors while building con- trast, leading to a clear bias toward background points that often dominate in 3D scenes. Also, object aware- ness and foreground-to-background discrimination are ne- glected, making contrastive learning less effective. To tackle these issues, we propose a general foreground-aware feature contrast (FAC) framework to learn more effective point cloud representations in pre-training. FAC consists of two novel contrast designs to construct more effective and informative contrast pairs. The first is building positive pairs within the same foreground segment where points tend to have the same semantics. The second is that we prevent over-discrimination between 3D segments/objects and en- courage foreground-to-background distinctions at the seg- ment level with adaptive feature learning in a Siamese cor- respondence network, which adaptively learns feature cor- relations within and across point cloud views effectively. Visualization with point activation maps shows that our contrast pairs capture clear correspondences among fore- ground regions during pre-training. Quantitative exper- iments also show that FAC achieves superior knowledge transfer and data efficiency in various downstream 3D se- mantic segmentation and object detection tasks. All codes, data, and models are available.	1. Introduction 3D scene understanding is crucial to many tasks such as robot grasping and autonomous navigation [12, 21, 30]. However, most existing work is fully supervised which re- lies heavily on large-scale annotated 3D data that is often laborious to collect. Self-supervised learning (SSL), which allows learning rich and meaningful representations from large-scale unannotated data, has recently demonstrated great potential to mitigate the annotation constraint [1, 5]. It learns with auxiliary supervision signals derived from unannotated data, which are usually much easier to col- †Corresponding Authors. Figure 1. Constructing informative contrast pairs matters in con- trastive learning: Conventional contrast requires strict point-level correspondence. The proposed method FAC takes both fore- ground grouping and foreground-background distinction cues into account, thus forming better contrast pairs to learn more informa- tive and discriminative 3D feature representations. lect. In particular, contrastive learning as one prevalent SSL approach has achieved great success in various visual 2D recognition tasks [6, 29]. Contrastive learning has also been explored for point cloud representation learning in various downstream tasks such as semantic segmentation [7, 18, 22, 42], instance seg- mentation [19, 20], and object detection [26, 44]. However, many successful 2D contrastive learning methods [6,14,46] do not work well for 3D point clouds, largely because point clouds often capture wide-view scenes which con- sist of complex points of many irregularly distributed fore- ground objects as well as a large number of background points. Several studies attempt to design specific contrast to cater to the geometry and distribution of point clouds. For example, [22] employ max-pooled features of two aug- mented scenes to form the contrast, but they tend to over- emphasize holistic information and overlook informative 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. 9476 features about foreground objects. [19, 26, 42] directly use registered point/voxel features as positive pairs and treat all non-registered as negative pairs, causing many false con- trast pairs in semantics. We propose exploiting scene foreground evidence and foreground-background distinction to construct more fore- ground grouping aware and foreground-background distinc- tion aware contrast for learning discriminative 3D represen- tations. For foreground grouping aware contrast, we first obtain regional correspondences with over-segmentation and then build positive pairs with points of the same re- gion across views, leading to semantic coherent repre- sentations. In addition, we design a sampling strategy to sample more foreground point features while building contrast, because the background point features are often less-informative and have repetitive or homogeneous pat- terns. For foreground-background contrast, we first en- hance foreground-background point feature distinction, and then design a Siamese correspondence network that se- lects correlated features by adaptively learning affinities among feature pairs within and across views in both fore- ground and background to avoid over-discrimination be- tween parts/objects. Visualizations show that foreground- enhanced contrast guides the learning toward foreground re- gions while foreground-background contrast enhances dis- tinctions among foreground and background features effec- tively in a complementary manner, the two collaborating to learn more informative and discriminative representation as illustrated in Fig. 1. The contributions of this work can be summarized in three aspects. First , we propose FAC, a foreground-aware feature contrast framework for large-scale 3D pre-training. Second , we construct region-level contrast to enhance the local coherence and better foreground awareness in the learned representations. Third , on top of that, we de- sign a Siamese correspondence framework that can lo- cate well-matched keys to adaptively enhance the intra- and inter-view feature correlations, as well as enhance the foreground-background distinction. Lastly , extensive ex- periments over multiple public benchmarks show that FAC achieves superior self-supervised learning when compared with the state-of-the-art. FAC is compatible with the preva- lent 3D segmentation backbone network SparseConv [15] and 3D detection backbone networks including PV-RCNN, PointPillars [25], and PointRCNN [36]. It is also applica- ble to both indoor dense RGB-D and outdoor sparse LiDAR point clouds.
Liu_Learning_Customized_Visual_Models_With_Retrieval-Augmented_Knowledge_CVPR_2023	Abstract Image-text contrastive learning models such as CLIP have demonstrated strong task transfer ability. The high generality and usability of these visual models is achieved via a web-scale data collection process to ensure broad con- cept coverage, followed by expensive pre-training to feed all the knowledge into model weights. Alternatively, we propose REACT ,REtrieval- Augmented CusTomization, a framework to acquire the relevant web knowledge to build customized visual models for target domains. We retrieve the most relevant image-text pairs ( ∼3% of CLIP pre-training data) from the web-scale database as external knowledge and propose to customize the model by only training new modularized blocks while freezing all the original weights. The effectiveness of REACT is demonstrated via extensive experiments on classification, retrieval, detection and seg- mentation tasks, including zero, few, and full-shot settings. Particularly, on the zero-shot classification task, compared with CLIP , it achieves up to 5.4% improvement on ImageNet and 3.7% on the ELEVATER benchmark (20 datasets).	1. Introduction It has been a fundamental research problem in computer vision (CV) to build a transferable visual system that can easily adapt to a wide range of downstream tasks. With remarkable advances in deep learning, a de facto solution to achieve this is to train deep neural networks on a large amount of data to pursue the so-called generic visual repre- sentations. This dates back to the standard supervised train- ing on ImageNet [10], whose superb representation power is further demonstrated in BiT [23]/ViT [12] by scaling up the training to JFT300M [50]. Along the way, recent efforts have been applied to the popular image self-supervised learn- ing [6, 16, 17] to reduce the demand for labeled data. The ♠core contribution; ¶equal advising; §work initiated during an internship at Microsoft.third approach is image-text contrastive learning trained on billion-scale web-crawled image-text pairs. Such models, like CLIP [43] and ALIGN [20], are able to achieve great performance on different downstream domains, without the need of any human labels. Excellent empirical performance has been achieved with the above three pre-training methods, by following the well established two-stage pre-training then adaptation pipeline: model pre-training from scratch on large data, then model adaptation directly on downstream tasks. Specifically, the pre-trained models are adapted to downstream tasks by con- sidering the available task-specific samples only: either eval- uated in a zero-shot task transfer manner, or updated using linear probing (LP) [43], finetuning (FT) [27], or prompt tun- ing [44,71]. Following this two-stage pipeline, most research has reverted to the faith that building transferable visual sys- tems is equivalent to developing more generic visual models by feeding all knowledge in the model pre-training stage. Therefore, the community has been witnessing a trend in exploring scaling success of pre-training model and data size with less care on the target domain, hoping that the model can adapt to any downstream scenario. In this paper, we argue that the conventional two-stage pipeline above is over-simplified and less efficient, in achiev- ing the goal of building a transferable visual system in real- world settings. Instead, we propose a customization stage in between the pre-training and adaptation, where customiza- tion is implemented by systematically leveraging retrieved external knowledge. The inspiration comes from how hu- mans are specialized in society for better generalization: instead of trying to memorize all concepts, humans are trained/prepared in a relevant subject to master a certain skill, while maintaining the basic skills in pre-training. To this end, we explore a systematic approach to acquire and learn with external knowledge sources from a large image-text corpus for model customization. The process of collecting external image-text knowledge is fully automatic without extra human annotation. The acquired knowledge typically contains richer information about the concept: rel- evant images that never appear in the downstream training 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. 15148 400 410 2000 # Images Observed in Model Lifecycle (M)707880 ImageNet Zero-Shot (%) CLIP-B32CLIP-L14 OpenCLIP-L14 REACT-B32-LTREACT-B32REACT-L14 OpenCLIP-B32OpenCLIP-L14OpenCLIP-HOpenCLIP-GREACT-G 0100200 400 800 # Parameters (Millions)65707580ImageNet-1K T op-1 Accuracy REACT CLIP Semi-ViT SimCLR-v2 60657075808590Classification / Retrieval Performance+2.8+1.1+0.4 +3.8+3.5 +5.1+2.0+1.4+3.0 +10 Classification ImageNetZero-Shot 1% 10% Classification ELEVATER BenchmarkZero-Shot Few-ShotLP FT Full-ShotLP FT Retrieval Flickr30KI2T T2I Detection MSCOCOZero-Shot Open-Voc Segmentation MSCOCOZero-Shot Anno-FreeREACT CLIP 1520253035404550 Dense Prediction Performance+1.5+3.4 +3.6+2.6Figure 1. REACT achieves the best zero-shot ImageNet performance among public checkpoints (Left), achieves new SoTA on semi-supervised ImageNet classification in the 1% labeled data setting (Middle), and consistently transfer better than CLIP on across a variety of tasks, including ImageNet classification, zero/few/full-shot classification on 20 datasets in ELEV ATER benchmark, image-text retrieval, object detection and segmentation (Right). Please see the detailed numbers and settings in the experimental section. For the left figure, circle size indicates model size. and evaluation set, and richer text descriptions about concept semantics. Such multi-modal knowledge sources are gen- erally available on the web, and further open-sourced like LAION [45, 46]. They cover a variety of domains, making it possible to develop customized visual models for task-level transfer. Similar retrieval-augmented intuitions have been exploited in computer vision for class-level transfer [32], but not yet for task-level transfer (similar to that of CLIP). Our main findings/contributions can be summarized as follows. We propose to explore the potential of the web-scale image-text corpus as external knowledge to significantly improve task-level transfer performance on the target do- main at an affordable cost. A simple and effective strategy is proposed. To begin with, we build a large-scale multi- modal indexing system to retrieve the relevant image-text pairs using CLIP features and approximate nearest neigh- bor search. For a CV problem, the task instruction is often sufficiently specified with text such as class names, which allows us to utilize them as queries to retrieve the relevant image-text pair knowledge from the indexing system. No images from the CV problem are needed. To efficiently build the customized visual model, we propose a novel modular- ized learning strategy: only updating the additional trainable weights on the retrieved knowledge, and freezing the origi- nal model weights. Hence, the model masters the new skill without forgetting basic skills. The generality and effectiveness of the proposed cus- tomization strategy is demonstrated on four CV problems . We instantiate it with CLIP, and develop the customized visual models for image classification on ImageNet and 20 datasets in ELEVATER [27], image-text retrieval on COCO [30]/Flickr [41], as well as object detection and se- mantic segmentation on COCO [30]. The knowledge bases are considered as LAION [46] and larger web-crawled multi- modal data. The retrieval-augmented knowledge ( ∼3% image-text pairs compared with the original training data) significantly improves the model’s zero-shot performancewithout the need of accessing any images on downstream tasks. See Figure 1 for highlighted results. For example, our ViT-L/14 checkpoint achieves 78.5% zero-shot accuracy on ImageNet [10], surpassing all public checkpoints from CLIP [43] and OpenCLIP [18], including those with larger model size and trained on a much larger LAION-2B [45]. The new customized models demonstrate higher few/full- shot performance than the generic model counterparts. Our retrieval system, codebase, and pre-trained models are publicly available . To make this line of research more accessible, our retrieved subsets for both ELEVATER and ImageNet will also be made available, with an easy-to-use toolkit to download the subsets without storing the whole dataset locally. It poses a feasible direction for leveraging the ever-increasing data from the Internet for customized visual recognition, especially for the low-resource regimes.
Kong_vMAP_Vectorised_Object_Mapping_for_Neural_Field_SLAM_CVPR_2023	Abstract We present vMAP, an object-level dense SLAM system using neural field representations. Each object is repre- sented by a small MLP , enabling efficient, watertight object modelling without the need for 3D priors. As an RGB-D camera browses a scene with no prior in- formation, vMAP detects object instances on-the-fly, and dynamically adds them to its map. Specifically, thanks to the power of vectorised training, vMAP can optimise as many as 50 individual objects in a single scene, with an extremely efficient training speed of 5Hz map update. We experimentally demonstrate significantly improved scene- level and object-level reconstruction quality compared to prior neural field SLAM systems. Project page: https: //kxhit.github.io/vMAP .	1. Introduction For robotics and other interactive vision applications, an object-level model is arguably semantically optimal, with scene entities represented in a separated, composable way, but also efficiently focusing resources on what is important in an environment. The key question in building an object-level mapping system is what level of prior information is known about the objects in a scene in order to segment, classify and re-construct them. If no 3D object priors are available, then usually only the directly observed parts of objects can be reconstructed, leading to holes and missing parts [4, 46]. Prior object information such as CAD models or category- level shape space models enable full object shape estima- tion from partial views, but only for the subset of objects in a scene for which these models are available. In this paper, we present a new approach which applies to the case where no 3D priors are available but still often en- ables watertight object reconstruction in realistic real-time scene scanning. Our system, vMAP, builds on the attractive properties shown by neural fields as a real-time scene repre- sentation [31], with efficient and complete representation of shape, but now reconstructs a separate tiny MLP model of each object. The key technical contribution of our work is to show that a large number of separate MLP object models can be simultaneously and efficiently optimised on a single GPU during live operation via vectorised training. We show that we can achieve much more accurate and complete scene reconstruction by separately modelling ob- jects, compared with using a similar number of weights in a single neural field model of the whole scene. Our real- time system is highly efficient in terms of both computation and memory, and we show that scenes with up to 50 objects can be mapped with 40KB per object of learned parameters across the multiple, independent object networks. 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. 952 We also demonstrate the flexibility of our disentangled object representation to enable recomposition of scenes with new object configurations. Extensive experiments have been conducted on both simulated and real-world datasets, showing state-of-the-art scene-level and object-level recon- struction performance.
Lee_Learning_Rotation-Equivariant_Features_for_Visual_Correspondence_CVPR_2023	AbstractExtracting discriminative local features that are invari-ant to imaging variations is an integral part of establish-ing correspondences between images. In this work, weintroduce a self-supervised learning framework to extractdiscriminative rotation-invariant descriptors using group-equivariant CNNs. Thanks to employing group-equivariantCNNs, our method effectively learns to obtain rotation-equivariant features and their orientations explicitly, with-out having to perform sophisticated data augmentations.The resultant features and their orientations are further pro-cessed by group aligning, a novel invariant mapping tech-nique that shifts the group-equivariant features by their ori-entations along the group dimension. Our group align-ing technique achieves rotation-invariance without any col-lapse of the group dimension and thus eschews loss of dis-criminability. The proposed method is trained end-to-endin a self-supervised manner, where we use an orientationalignment loss for the orientation estimation and a con-trastive descriptor loss for robust local descriptors to ge-ometric/photometric variations. Our method demonstratesstate-of-the-art matching accuracy among existing rotation-invariant descriptors under varying rotation and also showscompetitive results when transferred to the task of keypointmatching and camera pose estimation.	1. IntroductionExtracting local descriptors is an essential step for vi-sual correspondence across images, which is used for awide range of computer vision problems such as visual lo-calization [29,47,48], simultaneous localization and map-ping [7,8,39], and 3D reconstruction [1,16,17,49,66]. Toestablish reliable visual correspondences, the properties ofinvariance and discriminativeness are required for local de-scriptors; the descriptors need to be invariant to geomet-ric/photometric variations of images while being discrimi-native enough to distinguish true matches from false ones.Since the remarkable success of deep learning for visualrecognition, deep neural networks have also been adoptedto learn local descriptors, showing enhanced performanceson visual correspondence [44,45,64]. Learningrotation-invariant local descriptors, however, remains challenging;the classical techiniques [11,27,46] for rotation-invariantdescriptors, which are used for shallow gradient-based fea-ture maps, cannot be applied to feature maps from stan-dard deep neural networks, in which rotation of input in-duces unpredictable feature variations. Achieving rotationinvariance without sacriﬁcing disriminativeness is particu-larly important for local descriptors as rotation is one of themost frequent imaging variations in reality.In this work, we propose a self-supervised approach toobtain rotation-invariant and discriminative local descrip-tors by leveraging rotation-equivariant CNNs. First, weuse group-equivariant CNNs [60] to jointly extract rotation-equivariant local features and their orientations from an im-age. To extract reliable orientations, we use an orientationalignment loss [21,23,63], which trains the network to pre-dict the dominant orientation robustly against other imag-ing variations, including illumination or viewpoint changes.Using group-equivariant CNNs enables the local featuresto be empowered with explicitly encoded rotation equiv-ariance without having to perform rigorous data augmen-tations [58,60]. Second, to obtain discriminative rotation-invariant descriptors from rotation-equivariant features, wepropose group-aligning thatshiftsthe group-equivariantfeatures by their dominant orientation along their groupdimension. Conventional methods to yield invariant fea-tures from group-equivariant features collapse the group di-mension by group-pooling,e.g.,max-pooling or bilinear-pooling [26], resulting in a drop in feature discriminabil-ity and quality. In contrast, our group-aligning preservesthe group dimension, achieving rotation-invariance whileeschewing loss of discriminability. Furthermore, by pre-serving the group dimension, we can obtain multiple de-scriptors by performing group-aligning using multiple ori-entation candidates, which improves the matching perfor-mance by compensating for potential errors in dominantorientation prediction. Finally, we evaluate our rotation-invariant descriptors against existing local descriptors, 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. 21887 our group-aligning scheme against group-pooling methodson various image matching benchmarks to demonstrate theefﬁcacy of our method.The contribution of our paper is fourfold:•We propose to extract discriminative rotation-invariantlocal descriptors to tackle the task of visual correspon-dence by utilizing rotation-equivariant CNNs.•We propose group-aligning, a method to shift a group-equivariant descriptor in the group dimension by itsdominant orientation to obtain a rotation-invariant de-scriptor without having to collapse the group informa-tion to preserve feature discriminability.•We use self-supervisory losses of orientation align-ment loss for orientation estimation, and a contrastivedescriptor loss for robust local descriptor extraction.•We demonstrate state-of-the-art performances undervarying rotations on the Roto-360 dataset and showcompetitive transferability on the HPatches dataset [2]and the MVS dataset [53].
Kuang_PaletteNeRF_Palette-Based_Appearance_Editing_of_Neural_Radiance_Fields_CVPR_2023	Abstract Recent advances in neural radiance fields have enabled the high-fidelity 3D reconstruction of complex scenes for novel view synthesis. However, it remains underexplored how the appearance of such representations can be effi- ciently edited while maintaining photorealism. In this work, we present PaletteNeRF , a novel method for photorealis- tic appearance editing of neural radiance fields (NeRF) based on 3D color decomposition. Our method decom- poses the appearance of each 3D point into a linear com- bination of palette-based bases (i.e., 3D segmentations de- fined by a group of NeRF-type functions) that are shared across the scene. While our palette-based bases are view- independent, we also predict a view-dependent function to capture the color residual (e.g., specular shading). Dur- ing training, we jointly optimize the basis functions and the color palettes, and we also introduce novel regulariz- ers to encourage the spatial coherence of the decomposi- *Parts of this work were done when Zhengfei Kuang was an intern at Adobe Research.tion. Our method allows users to efficiently edit the appear- ance of the 3D scene by modifying the color palettes. We also extend our framework with compressed semantic fea- tures for semantic-aware appearance editing. We demon- strate that our technique is superior to baseline methods both quantitatively and qualitatively for appearance edit- ing of complex real-world scenes. Our project page is https://palettenerf.github.io.	1. Introduction Neural Radiance Fields (NeRF) [23] and its variants [8, 25, 27, 39] have received increasing attention in recent years for their ability to robustly reconstruct real-world 3D scenes from 2D images and enable high-quality, photoreal- istic novel view synthesis. However, such volumetric repre- sentations are challenging to edit due to the fact that scene appearance is implicitly encoded in neural features and net- work weights that do not support local manipulation or in- tuitive modification. Multiple approaches have been proposed to support edit- 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. 20691 ing of NeRF. One category of methods [4, 18,41,45] re- cover the material properties of the scene so that they can re-render them under novel lighting conditions or ad- just the material properties such as surface roughness. Such methods rely on accurate estimation of the scene reflectance, which is typically challenging for real-world complex scenes captured under unconstrained environment. Another category of methods [21, 35] learns a latent code on which NeRF can be conditioned to produce the desired ap- pearance. However, these methods often suffer from limited capacity and flexibility and do not support fine-grained edit- ing. In addition, some other methods [40] learn to transfer the appearance of NeRF to match a given style image, but sometimes fail to maintain the same level of photorealism in the original scene. In this paper, we propose PaletteNeRF, a novel method to support flexible and intuitive editing of NeRF. Our method is inspired by previous image-editing methods based on color palettes [7, 31], where a small set of col- ors are used to represent the full range of colors in the im- age. We model the radiance of each point using a combi- nation of specular and diffuse components, and we further decompose the diffuse component into a linear combina- tion of view-independent color bases that are shared across the scene. During training, we jointly optimize the per- point specular component, the global color bases and the per-point linear weights to minimize the difference between the rendered images and the ground truth images. We also introduce novel regularizers on the weights to encourage the sparseness and spatially coherence of the decomposition and achieve more meaningful grouping. With the proposed framework, we can intuitively edit the appearance of NeRF by freely modifying the learned color bases (Fig. 1). We further show that our framework can be combined with se- mantic features to support semantic-aware editing. Unlike previous palette-based image [1, 31] or video [10] editing methods, our method produces more globally coherent and 3D consistent recoloring results of the scene across arbi- trary views. We demonstrate that our method can enable more fine-grained local color editing while faithfully main- taining the photorealism of the 3D scene, and achieves bet- ter performance than baseline methods both quantitatively and qualitatively. In summary, our contributions include: • We propose a novel framework to facilitate the edit- ing of NeRF by decomposing the radiance field into a weighted combination of learned color bases. • We introduced a robust optimization scheme with novel regularizers to achieve intuitive decompositions. • Our approach enables practical palette-based appear- ance editing, making it possible for novice users to in- teractively edit NeRF in an intuitive and controllable manner on commodity hardware.
Liu_PD-Quant_Post-Training_Quantization_Based_on_Prediction_Difference_Metric_CVPR_2023	Abstract Post-training quantization (PTQ) is a neural network compression technique that converts a full-precision model into a quantized model using lower-precision data types. Although it can help reduce the size and computational cost of deep neural networks, it can also introduce quantiza- tion noise and reduce prediction accuracy, especially in ex- tremely low-bit settings. How to determine the appropriate quantization parameters ( e.g., scaling factors and round- ing of weights) is the main problem facing now. Existing methods attempt to determine these parameters by minimize the distance between features before and after quantization, but such an approach only considers local information and may not result in the most optimal quantization parameters. We analyze this issue and propose PD-Quant, a method that addresses this limitation by considering global infor- mation. It determines the quantization parameters by us- ing the information of differences between network predic- tion before and after quantization. In addition, PD-Quant can alleviate the overfitting problem in PTQ caused by the small number of calibration sets by adjusting the distribu- tion of activations. Experiments show that PD-Quant leads to better quantization parameters and improves the predic- tion accuracy of quantized models, especially in low-bit settings. For example, PD-Quant pushes the accuracy of ResNet-18 up to 53.14% and RegNetX-600MF up to 40.67% in weight 2-bit activation 2-bit. The code is released at https://github.com/hustvl/PD-Quant .	1. Introduction Various neural networks have been used in many real- world applications with high prediction accuracy. When de- ployed on resource-limited devices, networks’ vast memory and computation costs become significant challenges. Re- ⋆Equal contribution.⋄This work was done when Jiawei Liu and Lin Niu were interns at Houmo AI.†Corresponding authors.ducing overhead while maintaining the model accuracy has received considerable attention. Network quantization is an effective technique that can compress the neural networks by converting the format of values from floating-point to low-bit [10, 12, 27]. There are two types of quantization: post-training quantization (PTQ) [32] and quantization- aware training (QAT) [18]. QAT requires retraining a model on the labeled training dataset, which is time-consuming and computationally expensive. While PTQ only requires a small number of unlabeled calibration samples to quantize the pre-trained models without retraining, which is suitable for quick deployment. Existing PTQ methods can achieve good prediction accuracy with 8-bit or 4-bit quantization by selecting appropriate quantization parameters. [22, 23, 30]. Local metrics (such as MSE [7] or cosine distance [45] of the activation before and after quantization in layers) are commonly used to search for quantization scaling factors. These factors are chosen layer by layer by minimizing the local metric with a small number of calibration samples. In this paper, we observe that there is a gap between the se- lected scaling factors and the optimal scaling factors . Since the noise from quantization will be more severe at low-bit, the prediction accuracy of the quantized model significantly decreases at 2-bit. Recently, some meth- ods [24, 25, 44] have added a new class of quantization pa- rameters, weight rounding value, to adjust the rounding of weights. They optimize both quantization scaling factors and rounding values by reconstructing features layer-wisely or block-wisely. Besides, the quantized model by PTQ re- construction is more likely to be overfitting to the calibra- tion samples because adjusting the rounding of weights will significantly increase the PTQ’s degree of freedom. We propose an effective PTQ method, PD-Quant, to ad- dress the above-mentioned issues. In this paper, we fo- cus on improving the performance of PTQ on extremely low bit-width. PD-Quant uses the metric that considers the We define the optimal quantization scaling factors as the factors that make the quantized model have the lowest task loss (cross-entropy loss calculated by real label) on the validation set. 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. 24427 global information from the prediction difference between the quantized model and the full-precision (FP) model. We show that the quantization parameters optimized by predic- tion difference are more accurate in modeling the quan- tization noise. Besides, PD-Quant adjusts the activations for calibration in PTQ to mitigate the overfitting problem. The distribution of the activations is adjusted to meet the mean and variance saved in batch normalization layers. Ex- periments show that PD-Quant leads to better quantization parameters and improves the prediction accuracy of quan- tized models, especially in low-bit settings. Our PD-Quant achieves state-of-the-art performance in PTQ. For example, PD-Quant pushes the accuracy of weight 2-bit activation 2-bit ResNet-18 up to 53.14% and RegNetX-600MF up to 40.67%. Our contributions are summarized as follows: 1. We analyze the influence of different metrics and indi- cate that the widely used local metric can be improved further. 2. We propose to use the information of the prediction difference in PTQ, which improves the performance of the quantized model. 3. We propose Distribution Correction (DC) to adjust the activation distribution to approximate the mean and variance stored in the batch normalization layer, which mitigates the overfitting problem.
Li_Lite_DETR_An_Interleaved_Multi-Scale_Encoder_for_Efficient_DETR_CVPR_2023	Abstract Recent DEtection TRansformer-based (DETR) models have obtained remarkable performance. Its success cannot be achieved without the re-introduction of multi-scale feature fusion in the encoder. However, the excessively increased to- kens in multi-scale features, especially for about 75% of low- level features, are quite computationally inefficient, which hinders real applications of DETR models. In this paper, we present Lite DETR, a simple yet efficient end-to-end object detection framework that can effectively reduce the GFLOPs of the detection head by 60% while keeping 99% of the origi- nal performance. Specifically, we design an efficient encoder block to update high-level features (corresponding to small- resolution feature maps) and low-level features (correspond- ing to large-resolution feature maps) in an interleaved way. In addition, to better fuse cross-scale features, we develop a key-aware deformable attention to predict more reliable attention weights. Comprehensive experiments validate the effectiveness and efficiency of the proposed Lite DETR, and the efficient encoder strategy can generalize well across existing DETR-based models. The code will be available inhttps://github.com/IDEA-Research/Lite- DETR .	1. Introduction Object detection aims to detect objects of interest in im- ages by localizing their bounding boxes and predicting the corresponding classification scores. In the past decade, re- markable progress has been made by many classical de- tection models [23, 24] based on convolutional networks. *This work was done when Feng Li was an intern at IDEA. †Corresponding author. Figure 1. Average precision (Y axis) versus GFLOPs (X axis) for different detection models on COCO without extra training data. All models except EfficientDet [29] and YOLO series [12, 30] use ResNet-50 and Swin-Tiny as backbones. Specifically, two markers on the same line use ResNet-50 and Swin-Tiny, respectively. In- dividual markers only use ResNet-50. Each dashed line connects algorithm variants before and after adding our algorithm. The size of the listed models vary from 32M to 82M. Recently, DEtection TRansformer [1] (DETR) introduces Transformers into object detection, and DETR-like models have achieved promising performance on many fundamental vision tasks, such as object detection [13, 36, 37], instance segmentation [5, 6, 14], and pose estimation [26, 28]. Conceptually, DETR [1] is composed of three parts: a backbone, a Transformer encoder, and a Transformer de- coder. Many research works have been improving the back- bone and decoder parts. For example, the backbone in DETR is normally inherited and can largely benefit from a pre- trained classification model [10, 20]. The decoder part in DETR is the major research focus, with many research works trying to introduce proper structure to DETR query and im- prove its training efficiency [11, 13, 18, 21, 36, 37]. By 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. 18558 trast, much less work has been done to improve the encoder part. The encoder in vanilla DETR includes six Transformer encoder layers, stacked on top of a backbone to improve its feature representation. Compared with classical detection models, it lacks multi-scale features, which are of vital im- portance for object detection, especially for detecting small objects [9, 16, 19, 22, 29]. Simply applying Transformer en- coder layers on multi-scale features is not practical due to the prohibitive computational cost that is quadratic to the number of feature tokens. For example, DETR uses the C5 feature map, which is 1/32 of the input image resolution, to apply the Transformer encoder. If a C3 feature (1/8 scale) is included in the multi-scale features, the number of tokens from this scale alone will be 16 times of the tokens from the C5 feature map. The computational cost of self-attention in Transformer will be 256 times high. To address this problem, Deformable DETR [37] devel- ops a deformable attention algorithm to reduce the self- attention complexity from quadratic to linear by compar- ing each query token with only a fixed number of sampling points. Based on this efficient self-attention computation, Deformable DETR introduces multi-scale features to DETR, and the deformable encoder has been widely adopted in subsequent DETR-like models [11, 13, 18, 36]. However, due to a large number of query tokens intro- duced from multi-scale features, the deformable encoder still suffers from a high computational cost. To reveal this problem, we conduct some analytic experiments as shown in Table 1 and 2 using a DETR-based model DINO [36] to analyze the performance bottleneck of multi-scale features. Some interesting results can be observed. First, the low-level (high-resolution map) features account for more than 75% of all tokens. Second, direct dropping some low-level features (DINO-3scale) mainly affects the detection performance for small objects (AP_S) by a 10% drop but has little impact on large objects (AP_L). Inspired by the above observations, we are keen to address a question: can we use fewer feature scales but maintain important local details? Taking advantage of the structured multi-scale features, we present an efficient DETR frame- work, named Lite DETR . Specifically, we design a simple yet effective encoder block including several deformable self-attention layers, which can be plug-and-play in any multi-scale DETR-base models to reduce 62%∼78% en- coder GFLOPs and maintain competitive performance. The encoder block splits the multi-scale features into high-level features (e.g., C6, C5, C4) and low-level features (e.g., C3). High-level and low-level features will be updated in an in- terleaved way to improve the multi-scale feature pyramid. That is, in the first few layers, we let the high-level features query all feature maps and improve their representations, but keep low-level tokens intact. Such a strategy can effectively reduce the number of query tokens to 5%∼25% of theoriginal tokens and save a great amount of computational cost. At the end of the encoder block, we let low-level to- kens query all feature maps to update their representations, thus maintaining multi-scale features. In this interleaved way, we update high-level and low-level features in different frequencies for efficient computation. Moreover, to enhance the lagged low-level feature up- date, we propose a key-aware deformable attention (KDA) approach to replacing all attention layers. When performing deformable attention, for each query, it samples both keys and values from the same sampling locations in a feature map. Then, it can compute more reliable attention weights by comparing the query with the sampled keys. Such an approach can also be regarded as an extended deformable attention or a sparse version of dense attention. We have found KDA very effective in bringing the performance back with our proposed efficient encoder block. To summarize, our contributions are as follows. •We propose an efficient encoder block to update high- level and low-level features in an interleaved way, which can significantly reduce the feature tokens for efficient detection. This encoder can be easily plugged into existing DETR-based models. •To enhance the lagged feature update, we introduce a key-aware deformable attention for more reliable atten- tion weights prediction. •Comprehensive experiments show that Lite DETR can reduce the detection head GFLOPs by 60% and main- tain99% detection performance. Specifically, our Lite- DINO-SwinT achieves 53.9AP with 159GFLOPs.
Kolek_Explaining_Image_Classifiers_With_Multiscale_Directional_Image_Representation_CVPR_2023	Abstract Image classiﬁers are known to be difﬁcult to interpret and therefore require explanation methods to understand their decisions. We present ShearletX, a novel mask ex- planation method for image classiﬁers based on the shear- let transform – a multiscale directional image representa- tion. Current mask explanation methods are regularized by smoothness constraints that protect against undesirable ﬁne-grained explanation artifacts. However, the smooth- ness of a mask limits its ability to separate ﬁne-detail pat- terns, that are relevant for the classiﬁer, from nearby nui- sance patterns, that do not affect the classiﬁer. ShearletX solves this problem by avoiding smoothness regularization all together, replacing it by shearlet sparsity constraints. The resulting explanations consist of a few edges, textures, and smooth parts of the original image, that are the most relevant for the decision of the classiﬁer. To support our method, we propose a mathematical deﬁnition for explana- tion artifacts and an information theoretic score to evaluate the quality of mask explanations. We demonstrate the supe- riority of ShearletX over previous mask based explanation methods using these new metrics, and present exemplary situations where separating ﬁne-detail patterns allows ex- plaining phenomena that were not explainable before.	1. Introduction Modern image classiﬁers are known to be difﬁcult to explain. Saliency maps comprise a well-established ex- plainability tool that highlights important image regions for the classiﬁer and helps interpret classiﬁcation deci- sions. An important saliency approach frames saliency map computation as an optimization problem over masks [8,10,13,14,18,24,29]. The explanation mask is opti- mized to keep only parts of the image that sufﬁce to retain the classiﬁcation decision. However, Fong and Vedaldi [ 14] showed that an unregularized explanation mask is very sus- ceptible to explanation artifacts and is hence unreliable. Figure 1. Left column: ImageNet samples with prediction. Mid- dle column: Smooth pixel mask explanation from Fong et al. [ 13]. Right column: ShearletX (ours). Retained probability is computed as class probability after masking divided by class probability be- fore masking. ShearletX is the ﬁrst mask explanation method that can separate ﬁne-detail patterns, that are relevant for the classiﬁer, from nearby patterns that are irrelevant, without producing arti- facts. Therefore, current practice [ 8,13,14] heavily regularizes the explanation masks to be smooth. The smooth explanation masks can communicate useful explanatory information by roughly localizing the relevant image region. However, the pattern that is relevant for the classiﬁer is often overlaid on patterns that do not affect the classiﬁer. In such a situa- tion the mask cannot effectively separate the relevant pat- tern from the nuisance pattern, due to the smoothness con- 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. 18600 straints. As a result, many details that are irrelevant to the classiﬁer, such as background elements, textures, and other spatially localized patterns, appear in the explanation. An ideal mask explanation method should be resistant to explanation artifacts and capable of highlighting only rele- vant patterns. We present such a method, called ShearletX , that is able to separate different patterns that occupy nearby spatial locations by optimizing a mask in the shearlet rep- resentation of an image [ 25]. Due to the ability of shear- lets to efﬁciently encode directional features in images, we can separate relevant ﬁne-grained image parts, like edges, smooth areas, and textures, extremely well. We show both theoretically and experimentally that deﬁning the mask in the shearlet domain circumvents explanation artifacts. The masked image is optimized so that the classiﬁer retains its prediction as much as possible and to have small spatial sup- port (but not high spatial smoothness), while regularizing the mask to be sparse in the shearlet domain. This regular- ization assures that ShearletX retains only relevant parts, a fact that we support by a new information theoretic score for the quality of mask explanations. Figure 1gives examples demonstrating that ShearletX can separate relevant details from nuisance patterns, which smooth pixel masks cannot. Our contributions are summarized as follows: 1.ShearletX: The ﬁrst mask explanation method that can effectively separate ﬁne-detail patterns, that are rele- vant for the classiﬁer, from nearby nuisance patterns, that do not affect the classiﬁer. 2.Artifact Analysis: Our explanation method is based on low-level vision for maximal interpretability and belongs to the family of methods that produce out- of-distribution explanations. To validate that the re- sulting out-of-distribution explanations are meaning- ful, we develop a theory to analyze and quantify expla- nation artifacts, and prove that ShearletX is resilient to such artifacts. 3.Hallucination Score: a new metric for mask explana- tions that quantiﬁes explanation artifacts by measuring the amount of edges in the explanation that do not ap- pear in the original image. 4.Concisesness-Preciseness Score: A new information theoretic metric for mask explanations that gives a high score for explanations that extract the least amount of information from the image to retain the classiﬁcation decision as accurately as possible. 5.Experimental Results: We demonstrate that ShearletX performs better than previous mask explanations using our new metrics and give examples where ShearletX allows to explain phenomena that were not explainable with previous saliency methods. The source code for the experiments is publicly available1. 1https://github.com/skmda37/ShearletX

Subsets and Splits

No community queries yet

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