Split (1)

train · 2.34k rows

title stringlengths 28 135	abstract stringlengths 0 12k	introduction stringlengths 0 12k
Sawdayee_OReX_Object_Reconstruction_From_Planar_Cross-Sections_Using_Neural_Fields_CVPR_2023	Abstract Reconstructing 3D shapes from planar cross-sections is a challenge inspired by downstream applications like med- ical imaging and geographic informatics. The input is an in/out indicator function fully defined on a sparse collection of planes in space, and the output is an interpolation of the indicator function to the entire volume. Previous works ad- dressing this sparse and ill-posed problem either produce low quality results, or rely on additional priors such as target topology, appearance information, or input normal direc- tions. In this paper, we present OReX, a method for 3D shape reconstruction from slices alone, featuring a Neural Field as the interpolation prior. A modest neural network is trained on the input planes to return an inside/outside estimate for a given 3D coordinate, yielding a powerful prior that induces smoothness and self-similarities. The main challenge for this approach is high-frequency details, as the neural prior is overly smoothing. To alleviate this, we offer an iterative estimation architecture and a hierarchical input sampling scheme that encourage coarse-to-fine training, allowing the training process to focus on high frequencies at later stages.In addition, we identify and analyze a ripple-like effect stem- ming from the mesh extraction step. We mitigate it by regular- izing the spatial gradients of the indicator function around input in/out boundaries during network training, tackling the problem at the root. Through extensive qualitative and quantitative experimentation, we demonstrate our method is robust, accurate, and scales well with the size of the in- put. We report state-of-the-art results compared to previous approaches and recent potential solutions, and demonstrate the benefit of our individual contributions through analysis and ablation studies.1	1. Introduction Reconstructing a 3D object from its cross-sections is a long-standing task. It persists in fields including medical imaging, topography mapping, and manufacturing. The typi- cal setting is where a sparse set of arbitrary planes is given, upon which the “inside” and “outside” regions of the de- 1Code and data available at https://github.com/haimsaw/ OReX 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. 20854 picted domain are labeled, and the entire shape in 3D is to be estimated (see Fig. 1). This is a challenging and ill-posed problem, especially due to the sparse and irregular nature of the data. Classical approaches first localize the problem by constructing an arrangement of the input planes, and then introduce a local regularizer that governs the interpo- lation of the input to within each cell. While sound, these approaches typically involve simplistic regularization func- tions, that only interpolate the volume within a cell bounded by the relevant cross-sections; as a consequence, they intro- duce over-smoothed solutions that do not respect features. In addition, finding a cellular arrangement of planes is a computationally-heavy procedure, adding considerable com- plexity to the problem and rendering it quickly infeasible for large inputs (see Sec. 4). As we demonstrate (Sec. 4), recent approaches that reconstruct a mesh from an input point cloud are not well suited to our setting, as they assume a rather dense sampling of the entire shape. In addition, these methods do not consider the information of an entire cross-sectional plane, but rather only on the shape boundary. In this paper, we introduce OReX —a reconstruction ap- proach based on neural networks that estimates an entire shape from its cross-sectional slices. Similar to recent ap- proaches, the neural network constitutes the prior that ex- trapolates the input to the entire volume. Neural networks in general have already been shown to inherently induce smoothness [17], and self-similarities [12], allowing natural recurrence of patterns. Specifically, we argue that Neural Fields [25] are a promising choice for the task at hand. Neu- ral Fields represent a 3D scene by estimating its density and other local geometric properties for every given 3D coordi- nate. They are typically trained on 2D planar images, and are required to complete the entire 3D scene according to multi- view renderings or photographs. This neural representation is differentiable by construction, and hence allows native geometric optimization of the scene, realized via training. We pose the reconstruction problem as a classification of space into “in” and “out” regions, which are known for the entire slice planes, and thus generate the indicator function which its decision boundary is the output shape. The main challenge with applying neural fields to this problem is high-frequency details. Directly applying es- tablished training schemes [19] shows strong spectral bias, yielding overly smoothed results and other artifacts (Fig. 12). Spectral bias is a well-known effect, indicating that higher frequency is effectively learned slower [22]. To facilitate effective high-frequency learning, avoiding the shadow cast by the low frequency, we introduce two alterations. First, we sample the planar data mostly around the inside/outside tran- sition regions, where the frequencies are higher. This sam- pling is further ordered from low to high-frequency regions (according to the distance from the inside/outside boundary), to encourage a low-to-high-frequency training progression.In addition, we follow recent literature and allow the net- work to iteratively infer the result, where later iterations are responsible for finer, higher-frequency corrections [1, 23]. Finally, we consider another high-frequency artifact, also found in other neural-field-based works [9]. The desired density (or indicator) function dictates a sharp drop in value at the shape boundary. This is contradictory to the induced neural prior, causing sampling-related artifacts in the down- stream task of mesh extraction (Sec. 3.4). To alleviate this, we penalize strong spatial gradients around the boundary contours. This enforces a smoother transition between the in and out regions, allowing higher-quality mesh extraction. As we demonstrate (see Fig. 1), our method yields state- of-the-art reconstructions from planar data, both for woman- made and organic shapes. The careful loss and training schemes are validated and analyzed through quantitative and qualitative experimentation. Our method is arrangement- free, and thus both interpolates all data globally, avoiding local artifacts, and scales well to a large number of slices (see Sec. 4).
Ryoo_Token_Turing_Machines_CVPR_2023	Abstract We propose Token Turing Machines (TTM), a sequen- tial, autoregressive Transformer model with memory for real-world sequential visual understanding. Our model is inspired by the seminal Neural Turing Machine, and has an external memory consisting of a set of tokens which sum- marise the previous history (i.e., frames). This memory is efﬁciently addressed, read and written using a Transformer as the processing unit/controller at each step. The model’s memory module ensures that a new observation will only be processed with the contents of the memory (and not the entire history), meaning that it can efﬁciently process long sequences with a bounded computational cost at each step. We show that TTM outperforms other alternatives, such as other Transformer models designed for long sequences and recurrent neural networks, on two real-world sequential vi- sual understanding tasks: online temporal activity detection from videos and vision-based robot action policy learning. Code is publicly available at: https://github.com/google- research/scenic/tree/main/scenic/projects/token turing.	1. Introduction Processing long, sequential visual inputs in a causal man- ner is a problem central to numerous applications in robotics and vision. For instance, human activity recognition mod- els for monitoring patients and elders are required to make real-time inference on ongoing activities from streaming videos. As the observations grow continuously, these models require an efﬁcient way of summarizing and maintaining information in their past image frames with limited com- pute. Similarly, robots learning their action policies from training videos, need to abstract history of past observations and leverage it when making its action decisions in real-time. This is even more important if the robot is required to learn complicated tasks with longer temporal horizons. A traditional way of handling online observations of variable sequence lengths is to use recurrent neural net- works (RNNs), which are sequential, auto-regressive mod- els [13, 22, 35]. As Transformers [64] have become the defacto model architecture for a range of perception tasks, sev- eral works have proposed variants which can handle longer sequence lengths [19, 61, 67]. However, in streaming, or sequential inference problems, efﬁcient attention operations for handling longer sequence lengths themselves are often not sufﬁcient since we do not want to run our entire trans- former model for each time step when a new observation (e.g., a new frame) is provided. This necessitates developing models with explicit memories, enabling a model to fuse relevant past history with current observation to make a pre- diction at current time step. Another desideratum for such models, to scale to long sequence lengths, is that the compu- tational cost at each time step should be constant, regardless of the length of the previous history. In this paper, we propose Token Turing Machines (TTMs), a sequential, auto-regressive model with external memory and constant computational time complexity at each step. Our model is inspired by Neural Turing Machines [30] (NTM), an inﬂuential paper that was among the ﬁrst to propose an explicit memory and differentiable addressing mechanism. The original NTM was notorious for being a complex architecture that was difﬁcult to train, and it has therefore been largely forgotten as other modelling advances have been made in the ﬁeld. However, we show how we can formulate an external memory as well as a processing unit that reads and writes to this memory using Transformers (plus other operations which are common in modern Trans- former architectures). Another key component of TTMs is a token summarization module, which provides an inductive bias that intuitively encourages the memory to specialise to different parts of its history during the reading and writing operations. Moreover, this design choice ensures that the computational cost of our network is constant irrespective of the sequence length, enabling scalable, real-time, online inference. In contrast to the original NTM, our Transformer-based modernisation is simple to implement and train. We demon- strate its capabilities by achieving substantial improvements over strong baselines in two diverse and challenging tasks: (1) online temporal action detection (i.e., localisation) from videos and (2) vision-based robot action policy learning. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19070
Shi_Learning_3D-Aware_Image_Synthesis_With_Unknown_Pose_Distribution_CVPR_2023	Abstract Existing methods for 3D-aware image synthesis largely depend on the 3D pose distribution pre-estimated on the training set. An inaccurate estimation may mislead the model into learning faulty geometry. This work proposes PoF3D that frees generative radiance fields from the re- quirements of 3D pose priors. We first equip the generator with an efficient pose learner, which is able to infer a pose from a latent code, to approximate the underlying true pose distribution automatically. We then assign the discriminator a task to learn pose distribution under the supervision of the generator and to differentiate real and synthesized images with the predicted pose as the condition. The pose-free generator and the pose-aware discriminator are jointly trained in an adversarial manner. Extensive results on a couple of datasets confirm that the performance of our approach, regarding both image quality and geometry quality, is on par with state of the art. To our best knowledge, PoF3D demonstrates the feasibility of learning high-quality 3D-aware image synthesis without using 3D pose priors for the first time. Project page can be found here.	1. Introduction 3D-aware image generation has recently received grow- ing attention due to its potential applications [3, 4, 9, 21, 30, 34, 47]. Compared with 2D synthesis, 3D-aware image synthesis requires the understanding of the geometry underlying 2D images, which is commonly achieved by incorporating 3D representations, such as neural radiance fields (NeRF) [2, 16, 17, 24, 25, 50], into generative models like generative adversarial networks (GANs) [8]. Such a formulation allows explicit camera control over the synthe- sized results, which fits better with our 3D world. To enable 3D-aware image synthesis from 2D image collections, existing attempts usually rely on adequate camera pose priors [3, 4, 9, 20, 21, 30, 47] for training. †indicates equal contribution. * This work was done during an internship at Ant Group. (a) 𝜋-GAN: good pose distribution(b) 𝜋-GAN: bad pose distribution (c) CAMPARI: good pose initialization(d) CAMPARI: bad pose initialization Figure 1. Sensitivity to pose priors in existing 3D-aware image synthesis approaches. π-GAN [4] works well given an adequate pose distribution in (a), but fails to learn decent geometries given aslightly changed distribution in (b). CAMPARI [20] delivers reasonable results relying on a good initial pose distribution in (c), but suffers from a wrongly estimated initialization in (d). The priors are either estimated by conducting multiple pre- experiments [4, 9, 47], or obtained by borrowing external pose predictors or annotations [3]. Besides using a fixed distribution for pose sampling, some studies also propose to tune the pose priors together with the learning of image gen- eration [20], but they are still dependent on the initial pose distribution. Consequently, although previous methods can produce satisfying images and geometry, their performance is highly sensitive to the given pose prior. For example, on the Cats dataset [51], π-GAN [4] works well with a uniform pose distribution [-0.5, 0.5] (Fig. 1a), but fails to generate a decent underlying geometry when changing the distribution to [-0.3, 0.3] (Fig. 1b). Similarly, on the CelebA dataset [15], CAMPARI [20] learns 3D rotation when using Gaussian distribution N(0,0.24)as the initial prior (Fig. 1c), but loses the canonical space ( e.g., the middle image of Fig. 1d should be under the frontal view) when changing the initialization to N(0,0.12)(Fig. 1d). Such a reliance on pose priors causes unexpected instability of 3D-aware image synthesis, which may burden this task with heavy experimental cost. 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. 13062 In this work, we present a new approach for 3D-aware image synthesis, which removes the requirements for pose priors. Typically, a latent code is bound to the 3D content alone, where the camera pose is independently sampled from a manually designed distribution. Our method, how- ever, maps a latent code to an image, which is implicitly composed of a 3D representation ( i.e., neural radiance field) and a camera pose that can render that image. In this way, the camera pose is directly inferred from the latent code and jointly learned with the content, simplifying the input requirement. To facilitate the pose-free generator better capturing the underlying pose distribution, we re-design the discriminator to make it pose-aware. Concretely, we tailor the discriminator with a pose branch which is required to predict a camera pose from a given image. Then, the estimated pose is further treated as the conditional pseudo label when performing real/fake discrimination. The pose branch in the discriminator learns from the synthesized data and its corresponding pose that is encoded in the latent code, and in turn use the discrimination score to update the pose branch in the generator. With such a loop-back optimization process, two pose branches can align the fake data with the distribution of the dataset. We evaluate our pose-free method, which we call PoF3D for short, on various datasets, including FFHQ [12], Cats [51], and Shapenet Cars [5]. Both qualitative and quantitative results demonstrate that PoF3D frees 3D- aware image synthesis from hyper-parameter tuning on pose distributions and dataset labeling, and achieves on par performance with state-of-the-art in terms of image quality and geometry quality.
Seth_DeAR_Debiasing_Vision-Language_Models_With_Additive_Residuals_CVPR_2023	Abstract Large pre-trained vision-language models (VLMs) re- duce the time for developing predictive models for various vision-grounded language downstream tasks by providing rich, adaptable image and text representations. However, these models suffer from societal biases owing to the skewed distribution of various identity groups in the training data. These biases manifest as the skewed similarity between the representations for speciﬁc text concepts and images of peo- ple of different identity groups and, therefore, limit the use- fulness of such models in real-world high-stakes applica- tions. In this work, we present DEAR(Debiasing with Additive Residuals), a novel debiasing method that learns additive residual image representations to offset the orig- inal representations, ensuring fair output representations. In doing so, it reduces the ability of the representations to distinguish between the different identity groups. Further, we observe that the current fairness tests are performed *Equal Contributionon limited face image datasets that fail to indicate why a speciﬁc text concept should/should not apply to them. To bridge this gap and better evaluate DEAR, we introduce thePROTECTED ATTRIBUTE TAGASSOCIATION (PATA) dataset – a new context-based bias benchmarking dataset for evaluating the fairness of large pre-trained VLMs. Ad- ditionally, PATAprovides visual context for a diverse human population in different scenarios with both positive and neg- ative connotations. Experimental results for fairness and zero-shot performance preservation using multiple datasets demonstrate the efﬁcacy of our framework. The dataset is released here.	1. Introduction Deep learning-based vision-language models (VLMs) [ 45] unify text and visual data into a common representation and reduce the computing cost of training for speciﬁc computer vision [ 19] and visually-grounded linguistic tasks [ 12,47]. VLMs are trained with a large amount of data with 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. 6820 aim of matching image and text representations for image- caption pairs to capture diverse visual and linguistic con- cepts. However, VLMs exhibit societal biases manifesting as the skew in the similarity between their representation of certain textual concepts and kinds of images [ 2,4,31]. These biases arise from the underlying imbalances in train- ing data [ 2,3] and ﬂawed training practices [ 55]. In this work, we present a method to signiﬁcantly reduce the bias in VLMs by modifying the visual features of the models. These societal biases in VLMs show up as the selective association (or dissociation) of their representations of hu- man images with speciﬁc physical characteristics and their text representations for describing social labels [ 52]. For in- stance, a higher degree of similarity between the representa- tion of the text “doctor” and images of men than that of the women can have trust consequences for models using the representations from these VLMs. To estimate the degree of such biases in VLMs, we compute the cosine similarity between the representations of a set of human images and speciﬁc key text phrases and compare their distribution over some associated protected attributes . These attributes rep- resent visually discernible characteristics like gender ,race, andagecommon to certain collective identity groups. In this work, we introduce D EAR, an additive residual-based de-biasing technique that can be augmented with any pre- trained visual-language model with separate visual and lan- guage encoders [ 33,45,49] to improve their fairness. Our empirical analysis indicates that the protected at- tribute (PA) information from an image-text pair can be disentangled using their representation by simply learning a linear transformation of the visual representations pro- duced by a VLM and subtracting (adding a negative resid- ual) them from the original representation. We propose to train our framework using two objectives: i) learn a resid- ual representation that when added to the original repre- sentation, renders it incapable of predicting different pro- tected attributes, and ii) ensure that this modiﬁed represen- tation is as close to the original as possible. We demon- strate that learning additive residual enables the de-biasing of pre-trained VLMs using quantitative skew computations on multiple datasets and qualitative evaluations. We also show that the resulting representation retains much of its predictive properties by means of zero-shot evaluation on different downstream tasks. Recent efforts to mitigate these biases from VLMs, such as by Berg et al. [ 2], use unimodal datasets like the FairFace [ 28] and UTK [ 63] datasets. These face-image datasets lack the context necessary to infer the situations in which benign text associations can turn offensive when applied selectively. For instance, if the image of a person drinking water from a bottle is misconstrued as partaking of alcohol and if this association (in terms of image-text similarity) is selective to a speciﬁc group of people withsome speciﬁc visual characteristics, the association may be deemed offensive. To this end, we introduce the P RO- TECTED ATTRIBUTE TAGASSOCIATION (PATA) dataset to test different associations for VLMs. P ATAcomprises im- ages of persons in different contexts and their respective text captions with positive and negative connotations. We present our de-biasing results on both the FairFace and the PATAdatasets. In summary, the paper makes the following contributions: •We present the D EAR framework – a simple, computationally-efﬁcient, and effective de-biasing method for VLMs that only adapts the image encoder of the VLM by adding a learned residual representa- tion to it. •We introduce the P ROTECTED ATTRIBUTE TAGAS- SOCIATION dataset – a novel context-based bias eval- uation dataset that enables nuanced reporting of biases in VLMs for race, age, and gender protected attributes.
Rahman_Ambiguous_Medical_Image_Segmentation_Using_Diffusion_Models_CVPR_2023	Abstract Collective insights from a group of experts have al- ways proven to outperform an individual’s best diagnostic for clinical tasks. For the task of medical image segmen- tation, existing research on AI-based alternatives focuses more on developing models that can imitate the best indi- vidual rather than harnessing the power of expert groups. In this paper, we introduce a single diffusion model-based approach that produces multiple plausible outputs by learn- ing a distribution over group insights. Our proposed model generates a distribution of segmentation masks by leverag- ing the inherent stochastic sampling process of diffusion us- ing only minimal additional learning. We demonstrate on three different medical image modalities- CT, ultrasound, and MRI that our model is capable of producing several possible variants while capturing the frequencies of their occurrences. Comprehensive results show that our pro- posed approach outperforms existing state-of-the-art am- biguous segmentation networks in terms of accuracy while preserving naturally occurring variation. We also propose a new metric to evaluate the diversity as well as the ac- curacy of segmentation predictions that aligns with the in- terest of clinical practice of collective insights. Implemen- tation code: https://github.com/aimansnigdha/Ambiguous- Medical-Image-Segmentation-using-Diffusion-Models.	1. Introduction Diagnosis is the central part of medicine, which heavily relies on the individual practitioner assessment strategy. Re- cent studies suggest that misdiagnosis with potential mor- tality and morbidity is widespread for even the most com- mon health conditions [32, 49]. Hence, reducing the fre- quency of misdiagnosis is a crucial step towards improving healthcare. Medical image segmentation, which is a cen- tral part of diagnosis, plays a crucial role in clinical out- comes. Deep learning-based networks for segmentation are now getting traction for assisting in clinical settings, how- ever, most of the leading segmentation networks in the lit- erature are deterministic [17, 23, 34, 36, 41, 42, 44], mean- ing they predict a single segmentation mask for each in- put image. Unlike natural images, ground truths are not PriorNet μPrior 𝜎Prior Xb,TXb,tXb,t-1Xb,0b Sampling Network Network z1, z2, z3 …a)Deterministic b) V AE-based c)Diffusion-based (Ours) Input Prediction Predictions Predictions Input Input Annotations from different radiologists Input Figure 1. a) Deterministic networks produce a single output for an input image. b) c-V AE-based methods encode prior informa- tion about the input image in a separate network and sample latent variables from there and inject it into the deterministic segmenta- tion network to produce stochastic segmentation masks. c) In our method the diffusion model learns the latent structure of the seg- mentation as well as the ambiguity of the dataset by modeling the way input images are diffused through the latent space. Hence our method does not need an additional prior encoder to provide latent variables for multiple plausible annotations. deterministic in medical images as different diagnosticians can have different opinions on the type and extent of an anomaly [1,15,37,39]. Due to this, the diagnosis from med- ical images is quite challenging and often results in a low inter-rater agreement [22,24,56]. Depending on only pixel- wise probabilities and ignoring co-variances between the pixels might lead to misdiagnosis. In clinical practice, ag- gregating interpretations of multiple experts have shown to improve diagnosis and generate fewer false negatives [57]. In fact, utilizing the aptitude of multiple medical ex- perts has been a part of long-standing clinical traditions such as case conferences, specialist consultations, and tu- mor boards. By harnessing the power of collective intelli- gence, team-based decision-making provides safer health- care through improved diagnosis [32, 40]. Although collec- tive insight is gaining traction in healthcare for its potential 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. 11536 in enhancing diagnostic accuracy, the method and its im- plication remain poorly characterized in automated medical vision literature. It has been suggested that the use of artifi- cial intelligence can optimize these processes while consid- ering physician workflows in clinical settings [40]. In recent times, there has been an outstanding improve- ment in specialized deterministic models for different med- ical image segmentation tasks [13, 44, 52–55, 61]. Deter- ministic models are notorious for choosing the most likely hypothesis even if there is uncertainty which might lead to sub-optimal segmentation. To overcome this, some models incorporate pixel-wise uncertainty for segmentation tasks, however, they produce inconsistent outputs [25,26]. Condi- tional variational autoencoders (c-V AE) [48], a conditional generative model, can be fused with deterministic segmen- tation networks to produce unlimited numbers of predic- tions by sampling from the latent space conditioned on the input image. Probabilistic U-net and its variants use this technique during the inference process. Here, the latent spaces are sampled from a prior network which has been trained to be similar to c-V AE [8, 29, 30]. This dependency on a prior network as well as injecting stochasticity only at the highest resolution of the segmentation network pro- duces less diverse and blurry segmentation predictions [46]. To overcome this problem, we introduce a single inherently probabilistic model without any additional prior network that represents the collective intelligence of several experts to leverage multiple plausible hypotheses in the diagnosis pipeline (visualized in Figure 1). Diffusion probabilistic models are a class of generative models consisting of Markov chains trained using varia- tional inference [21]. The model learns the latent structure of the dataset by modeling the diffusion process through la- tent space. A neural network is trained to denoise noisy image blurred using Gaussian noise by learning the reverse diffusion process [50]. Recently, diffusion models have been found to be widely successful for various tasks such as image generation [14], and inpainting [35]. Certain ap- proaches have also been proposed to perform semantic seg- mentation using diffusion models [6,59]. Here, the stochas- tic element in each sampling step of the diffusion model using the same pre-trained model paves the way for gen- erating multiple segmentation masks from a single input image. However, there is still no exploration of using dif- fusion models for ambiguous medical image segmentation despite its high potential. In this paper, we propose the CIMD ( Collectivly Intelligent Medical Diffusion), which addresses ambiguous segmentation tasks of medical imag- ing. First, we introduce a novel diffusion-based probabilis- tic framework that can generate multiple realistic segmen- tation masks from a single input image. This is motivated by our argument that the stochastic sampling process of the diffusion model can be harnessed to sample multiple plausi-ble annotations. The stochastic sampling process also elim- inates the need for a separate ‘prior’ distribution during the inference stage, which is critical for c-V AE-based segmen- tation models to sample the latent distribution for ambigu- ous segmentation. The hierarchical structure of our model also makes it possible to control the diversity at each time step hence making the segmentation masks more realistic as well as heterogeneous. Lastly, in order to assess ambigu- ous medical image segmentation models, one of the most commonly used metrics is known as GED (Generalized En- ergy Distance), which matches the ground truth distribu- tion with prediction distribution. In real-world scenarios for ambiguous medical image segmentation, ground truth distributions are characterized by only a set of samples. In practice, the GED metric has been shown to reward sam- ple diversity regardless of the generated samples’ fidelity or their match with ground truths, which can be potentially harmful in clinical applications [30]. In medical practice, individual assessments are manually combined into a single diagnosis and evaluated in terms of sensitivity. When real- time group assessment occurs, the participant generates a consensus among themselves. Lastly, the minimum agree- ment and maximum agreement among radiologists are also considered in clinical settings. Inspired by the current prac- tice in collective insight medicine, we coin a new metric, namely the CI score ( Collective Insight) that considers total sensitivity, general consensus, and variation among radiolo- gists. In summary, the following are the major contributions of this work: • We propose a novel diffusion-based framework: Col- lectively Intelligent Medical Diffusion (CIMD), that realistically models heterogeneity of the segmentation masks without requiring any additional network to pro- vide prior information during inference unlike previ- ous ambiguous segmentation works. • We revisit and analyze the inherent problem of the current evaluation metric, GED for ambiguous models and explain why this metric is insufficient to capture the performance of the ambiguous models. We intro- duce a new metric inspired by collective intelligence medicine, coined as the CI Score ( Collective Insight). • We demonstrate across three medical imaging modal- ities that CIMD performs on par or better than the existing ambiguous image segmentation networks in terms of quantitative standards while producing supe- rior qualitative results.
Shen_Progressive_Transformation_Learning_for_Leveraging_Virtual_Images_in_Training_CVPR_2023	Abstract To effectively interrogate UAV-based images for detect- ing objects of interest, such as humans, it is essential to acquire large-scale UAV-based datasets that include human instances with various poses captured from widely varying viewing angles. As a viable alternative to laborious and costly data curation, we introduce Progressive Transfor- mation Learning (PTL), which gradually augments a train- ing dataset by adding transformed virtual images with en- hanced realism. Generally, a virtual2real transformation generator in the conditional GAN framework suffers from quality degradation when a large domain gap exists be- tween real and virtual images. To deal with the domain gap, PTL takes a novel approach that progressively iterates the following three steps: 1) select a subset from a pool of vir- tual images according to the domain gap, 2) transform the selected virtual images to enhance realism, and 3) add the transformed virtual images to the training set while remov- ing them from the pool. In PTL, accurately quantifying the domain gap is critical. To do that, we theoretically demon- strate that the feature representation space of a given object detector can be modeled as a multivariate Gaussian dis- tribution from which the Mahalanobis distance between a virtual object and the Gaussian distribution of each object category in the representation space can be readily com- puted. Experiments show that PTL results in a substantial performance increase over the baseline, especially in the small data and the cross-domain regime.	1. Introduction Training an object detector usually requires a large-scale training image set so that the detector can acquire the abil- ity to detect objects’ diverse appearances. This desire for a large-scale training set is bound to be greater for object cat- egories with more diverse appearances, such as the human category whose appearances vary greatly depending on its pose or viewing angles. Moreover, a person’s appearance Transformation Candidate SelectionVirtual2Real TransformationSet Update Real images w/ conditional GANReal images Virtual images - -+: add -: subtract + + Virtual imagesselect generatorgeneratorFigure 1. Overview of Progressive Transformation Learning. becomes more varied in images captured by an unmanned aerial vehicle (UA V), leading to a wide variety of camera viewing angles compared to ground-based cameras, mak- ing the desire for a large-scale training set even greater. In this paper, we aim to satisfy this desire, especially when the availability of UA V-based images to train a human detector is scarce, where this desire is more pressing. As an intuitive way to expand the training set, one might consider synthesizing virtual images to imitate real-world images by controlling the optical and physical conditions in a virtual environment. Virtual images are particularly use- ful for UA V-based object detection since abundant object instances can be rendered with varying UA V locations and camera viewing angles along with ground-truth informa- tion (e.g., bounding boxes, segmentation masks) that comes free of charge. Therefore, a large-scale virtual image set covering diverse appearances of human subjects that are rarely shown in existing UA V-based object detection bench- marks [1,3,53] can be conveniently acquired by controlling entities and parameters in a virtual environment, such as poses, camera viewing angles, and illumination conditions. To make virtual images usable for training real-world ob- ject detection models, recent works [19, 29, 38, 39] trans- form virtual images to look realistic. They commonly use the virtual2real generator [36] trained with the conditional GAN framework to transform images in the source domain to have the visual properties of images in the target do- main. Here, virtual and real images are treated as the source and target domains, respectively. However, the large dis- 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. 835 crepancy in visual appearance between the two domains, referred to as the “domain gap”, result in the degraded transformation quality of the generator. In fact, the afore- mentioned works using virtual images validate their meth- ods where the domain gap is not large (e.g., digit detec- tion [19]) or when additional information is available (e.g., animal pose estimation with additional keypoint informa- tion [29, 38, 39]). In our case, real and virtual humans in UA V-based images inevitably have a large domain gap due to the wide variety of human appearances. To address the large domain gap, one critical question inherent in our task is how to measure accurately the do- main gap. Consequently, we estimate the domain gap in the representation space of a human detector trained on the real images. The representation space of the detector is learned such that test samples, which have significantly different properties than training samples from the perspective of the detector, are located away from the training samples. In this paper, we show that the feature distribution of object entities belonging to a certain category, such as the human category, in the representation space can be modeled with a multivariate Gaussian distribution if the following two con- ditions are met: i) the detector uses the sigmoid function to normalize the final output and ii) the representation space is constructed using the output of the penultimate layer of the detector. This idea is inspired by [28], which shows that softmax-based classifiers can be modeled as multivari- ate Gaussian distributions. In this paper, we show that the proposition is also applicable to sigmoid-based classifiers, which are widely used by object detectors. Based on this modeling, when the two aforementioned conditions are met, the human category in the representation space can be rep- resented by two parameters (i.e., mean and covariance) of a multivariate Gaussian distribution that can be computed on the training images. With the empirically calculated mean and covariance, the domain gap from a single virtual hu- man image to real human images (i.e., the training set) can be measured using the Mahalanobis distance [35]. To add virtual images to the training set to include more diverse appearances of objects while preventing the trans- formation quality degradation caused by large domain gaps, we introduce Progressive Transformation Learning (PTL) (Figure 1). PTL progressively expands the training set by adding virtual images through iterating the three steps: 1) transformation candidate selection, 2) virtual2real transfor- mation, and 3) set update. When selecting transformation candidates from a virtual image pool, we use weighted ran- dom sampling, which gives higher weights to images with smaller domain gaps. The weight takes an exponential term with one hyperparameter controlling the ratio between im- ages with smaller domain gaps and images with more di- verse appearances. Then, the virtual2real transformation generator is trained via the conditional GAN, taking the se-lected transformation candidates as the “source” and the im- ages in the training set as the “target”. After transforming the transformation candidates by applying the virtual2real transformation generator, the training set is expanded with the transformed candidates while the original candidates are excluded from the pool of virtual images. The main contribution of this paper is that we have val- idated the utility of virtual images in augmenting training data via PTL coupled with carefully designed comprehen- sive experiments. We first use the task of low-shot learn- ing, where adequately expanding datasets has notable ef- fects. Specifically, PTL provides better accuracy on three UA V-view human detection benchmarks than other previ- ous works that leverage virtual images in training, as well as methods that only use real images. Then, we validate PTL on the cross-domain detection task where training and test sets are from distinct domains and virtual images can serve as a bridge between these two sets. The experimen- tal results indicate that a high-performance human detection model can be effectively learned via PTL, even with the sig- nificant lack of real-world training data.
Ren_VolRecon_Volume_Rendering_of_Signed_Ray_Distance_Functions_for_Generalizable_CVPR_2023	Abstract The success of the Neural Radiance Fields (NeRF) in novel view synthesis has inspired researchers to propose neural implicit scene reconstruction. However, most exist- ing neural implicit reconstruction methods optimize per- scene parameters and therefore lack generalizability to new scenes. We introduce VolRecon, a novel generalizable implicit reconstruction method with Signed Ray Distance Function (SRDF). To reconstruct the scene with fine de- tails and little noise, VolRecon combines projection features aggregated from multi-view features, and volume features interpolated from a coarse global feature volume. Using a ray transformer, we compute SRDF values of sampled points on a ray and then render color and depth. On DTU dataset, VolRecon outperforms SparseNeuS by about 30% in sparse view reconstruction and achieves comparable ac- curacy as MVSNet in full view reconstruction. Furthermore, our approach exhibits good generalization performance on the large-scale ETH3D benchmark. Code is available at https://github.com/IVRL/VolRecon/ .	1. Introduction The ability to reconstruct 3D geometries from images or videos is crucial in various applications in robotics [16, 43, 52] and augmented/virtual reality [29,35]. Multi-view stereo (MVS) [13, 15, 39, 47, 54, 55] is a commonly used technique for this task. A typical MVS pipeline involves multiple steps, i.e., multi-view depth estimation, filtering, and fusion [5,13]. Recently, there has been a growing interest in neural im- plicit representations for various 3D tasks, such as shape modeling [28, 36], surface reconstruction [49, 56], and novel view synthesis [30]. NeRF [30], a seminal work in this area, employs Multi-Layer Perceptrons (MLP) to model a radiance field, producing volume density and radiance estimates for a given position and viewing direction. While NeRF’s scene representation and volume rendering approach has proven effective for tasks such as novel view synthesis, it cannot *Equal contribution Input Reconstruction SparseNeuS Ours (VolRecon ) Figure 1. Generalizable implicit reconstructions from three views (top). The state-of-the-art method SparseNeuS [26] produces over- smoothed surfaces (left), while our (VolRecon) reconstructs finer details (right). Best viewed on a screen when zoomed in. generate accurate surface reconstruction due to difficulties in finding a universal density threshold for surface extrac- tion [56]. To address this, researchers have proposed neural implicit reconstruction using the Signed Distance Function (SDF) for geometry representation and modeling the vol- ume density function [49, 56]. However, utilizing SDF with only color supervision leads to unsatisfactory reconstruction quality compared to MVS methods [15, 54] due to a lack of geometry supervision and potential radiance-geometry ambi- guities [51,62]. As a result, subsequent works have sought to improve reconstruction quality by incorporating additional priors, such as sparse Struction-from-Motion (SfM) point clouds [11], dense MVS point clouds [60], normals [48, 59], and depth maps [59]. Many neural implicit reconstruction methods are re- stricted to optimizing one model for a particular scene and cannot be applied to new, unseen scenes, i.e., across-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. 16685 generalization. However, the ability to generalize learned priors to new scenes is valuable in challenging scenarios such as reconstruction with sparse views [4, 50, 58]. In or- der to achieve across-scene generalization in neural implicit reconstruction, it is insufficient to simply input the spatial coordinate of a point as NeRF. Instead, we need to incorpo- rate information about the scene, such as the points’ pro- jection features on the corresponding images [4, 50, 58]. SparseNeuS [26] recently achieved across-scene general- ization in implicit reconstruction with global feature vol- umes [4]. Despite achieving promising results, SparseNeuS is limited by the resolution of the feature volume due to the memory constraints [22, 32], leading to over-smoothing surfaces even with a higher resolution feature volume, Fig. 1. In this paper, we propose V olRecon, a novel framework for generalizable neural implicit reconstruction using the Signed Ray Distance Function (SRDF). Unlike SDF, which defines the distance to the nearest surface along any direc- tions, SRDF [63] defines the distance to the nearest surface along a given ray. We utilize a projection-based approach to gather local information about surface location. We first project each point on the ray into the feature map of each source view to interpolate multi-view features. Then, we ag- gregate the multi-view features to projection features using a view transformer. However, when faced with challenging situations such as occlusions and textureless surfaces, de- termining the surface location along the ray with only local information is difficult. To address this, we construct a coarse global feature volume that encodes global shape priors like SparseNeuS [26, 32]. We use the interpolated features from the global feature volume, i.e.,volume features , and pro- jection features of all the sampled points along the ray to compute their SRDF values, with a ray transformer. Similar to NeuS [49], we model the density function with SRDF and then estimate the image and depth map with volume rendering. Extensive experiments on DTU [1] and ETH3D [40] verify the effectiveness and generalization ability of our method. On DTU, our method outperforms the state-of- the-art method SparseNeuS [26] by 30% in sparse view reconstruction and 22% in full view reconstruction. Further- more, our method performs better than the MVS baseline COLMAP [39]. Compared with MVSNet [54], a seminal learning-based MVS method, our method performs better in the depth evaluation and has comparable accuracy in full- view reconstruction. On the ETH3D benchmark [40], we show that our method has a good generalization ability to large-scale scenes. In summary, our contributions are as follows: •We propose V olRecon, a new pipeline for generalizable implicit reconstruction that produce detailed surfaces. • Our novel framework comprises a view transformer toaggregate multi-view features and a ray transformer to compute SRDF values of all the points along a ray. •We introduce a combination of local projection features and global volume features, which enables the recon- struction of surfaces with fine details and high quality.
Si_Angelic_Patches_for_Improving_Third-Party_Object_Detector_Performance_CVPR_2023	Abstract Deep learning models have shown extreme vulnerabil- ity to distribution shifts such as synthetic perturbations and spatial transformations. In this work, we explore whether we can adopt the characteristics of adversarial attack meth- ods to help improve robustness of object detection to dis- tribution shifts such as synthetic perturbations and spa- tial transformations. We study a class of realistic object detection settings wherein the target objects have control over their appearance. To this end, we propose a reversed Fast Gradient Sign Method (FGSM) to obtain these angelic patches that significantly increase the detection probabil- ity, even without pre-knowledge of the perturbations. In de- tail, we apply the patch to each object instance simultane- ously, strengthening not only classification, but also bound- ing box accuracy. Experiments demonstrate the efficacy of the partial-covering patch in solving the complex bounding box problem. More importantly, the performance is also transferable to different detection models even under severe affine transformations and deformable shapes. To the best of our knowledge, we are the first object detection patch that achieves both cross-model efficacy and multiple patches. We observed average accuracy improvements of 30% in the real-world experiments. Our code is available at: https: //github.com/averysi224/angelic_patches .	1. Introduction Deep learning models have been heavily deployed in many safety-critical settings such as autonomous vehicles. However, these models have been notoriously fragile to mild perturbations. For example, natural corruptions like weather conditions and simple lightning effects can signif- icantly degrade the performance of state-of-the-art mod- els [11, 14]. Similarly, the performance under small spa- tial transformations exhibits a large gap compared to clean benchmarks [2, 7, 15]. On the other hand, a set of carefully designed adversarial examples [10] are able to manipulate the prediction behavior arbitrarily without notice of human eyes. The untrustworthiness of deep learning systems leads Angelic Patch Foggy Clean Foggy Patched Figure 1. In this paper, we demonstrate that optimizing a partial- cover patch for pre-trained object detectors can improve robust- ness and significantly boost performance for both the classification and bounding box regression accuracy. In the left picture, one of the three people is mis-detected after applying the fog corruption. However, in the right image with our angelic patch , the detector was able to detect all three people with even more accurate bound- ing boxes. Consider all three people wearing an angelic raincoat, we could save lives from the foggy-blinded autonomous car! to high stake failures and devastating consequences. Two main streams of algorithms investigate solving these problems. One is to improve out-of-distribution behav- ior by adding more robustness interventions and diverse data during training. However, they do not fully close the gap between standard model performance and perturbed re- sults [8]. Others applied domain adaptation over covari- ate shift achieving reasonable performance [23], yet this method does not generalize on unseen domains. Whereas the misspecified test time distribution occurs dominantly in practice. Motivated and inspired by the efficacy of these perturba- tion/adversarial methods above, we ask the question: Can we adopt the characteristics of adversarial attack methods to help improve perturbation-robustness? We propose to reconsider the problem setup itself and study in a scenario where the target objects are in control of their appearance. To simplify, we build the objects instead of the models to improve detection reliability. As a concrete example, con- sider a pedestrian interacting with autonomous cars that use deep learning models for detection. Our approach is to pro- vide a wearable patch designed to improve the visibility of these people to these models (Figure 1). Such practices 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. 24638 Step 1: Apply patch. Step 2: Apply corruption. (a) An example of corruption-aware angelic patch training procedure.Step 1: Apply patch. (b) An example of corruption-agnostic angelic patch training. Figure 2. Examples of the two considered methods for constructing angelic patches. In the corruption-aware setting, we compute loss by the predictions of the corrupted patched images. At test time, we test on the same type of corruption; In the corruption-agnostic setting, we compute loss by the predictions of the corrupted clean images. At test time, we test on arbitrary corruption. are already common for dealing with human drivers—e.g., wearing bright or reflective clothing at night. Furthermore, though the adversarial attacks and robust- ness of classifiers are well-studied, we focus on the less studied yet practically broadly used object detection setting. In detail, the detectors learn not only the object class infor- mation but also learn about the localization and the size of an object, e.g. bounding boxes. Besides, the object detec- tion problem is essentially multiple instances, causing im- plicit interaction between objects with even occlusions. To our knowledge, we are the first to systematically investigate this setting. We validated our method on both the single- stage detector and the two-stage detector that are with the proposal network as well as cross model experiments. Our contribution is three-fold: First, we propose the novel angelic patches with a Reversed Fast Signed Gra- dient Method to improve the performance of both single- stage and two-stage third-party object detectors. Second, we demonstrate the efficacy of our framework on a wide va- riety of detection settings including dozens of synthetic cor- ruptions and affine transformations without additional aug- mentation during training. Third, we are the first defense physical patch that achieves cross-model validation on sev- eral state-of-the-art unseen models. We extensively evalu- ated our approach with both programmatic patches as well as real-world experiments. We believe our approach identi- fies a highly practical valuable strategy that can be used in a broad range of applications. In the following sections, we start with a review of re- lated work, then give the proposed angelic patch framework and experiment results. After that, we conclude the paper.
Shen_Disentangling_Orthogonal_Planes_for_Indoor_Panoramic_Room_Layout_Estimation_With_CVPR_2023	Abstract Based on the Manhattan World assumption, most exist- ing indoor layout estimation schemes focus on recovering layouts from vertically compressed 1D sequences. However, the compression procedure confuses the semantics of differ- ent planes, yielding inferior performance with ambiguous interpretability. To address this issue, we propose to disentangle this 1D representation by pre-segmenting orthogonal (vertical and horizontal) planes from a complex scene, explicitly captur- ing the geometric cues for indoor layout estimation. Con- sidering the symmetry between the floor boundary and ceil- ing boundary, we also design a soft-flipping fusion strategy to assist the pre-segmentation. Besides, we present a fea- ture assembling mechanism to effectively integrate shallow and deep features with distortion distribution awareness. To compensate for the potential errors in pre-segmentation, we further leverage triple attention to reconstruct the dis- entangled sequences for better performance. Experiments on four popular benchmarks demonstrate our superiority over existing SoTA solutions, especially on the 3DIoU met- ric. The code is available at https://github.com/ zhijieshen-bjtu/DOPNet .	1. Introduction Indoor panoramic layout estimation refers to recon- structing 3D room layouts from omnidirectional images. Since the panoramic vision system captures the whole-room contextual information, we can estimate the complete room layout with a single panoramic image. However, inferring the 3D information from a 2D image is an ill-posed prob- lem. Besides, the 360° field-of-view (FoV) of panoramas brings severe distortions that increase along the latitude. Both issues are challenging for indoor layout estimation. †Corresponding author (a) Compressing Compressing Compressing C Multi -scale features Multi -scale features Features assembling Disentangling (b) Compressing Compressing Compressing Compressing Compressing ResizeResize ResizeSampling Horizontal plane Vertical planesFigure 1. (a) The commonly used architecture. (b) The proposed one. Compared with the traditional pipeline, ours has two advan- tages: (1) Disentangling the 1D representation into two separate sequences with different plane semantics. (2) Adaptively integrat- ing shallow and deep features with distortion awareness via a fea- ture assembling mechanism rather than simple concatenation. Different from outdoor scenarios, the indoor room has the following properties: (1) The indoor scenes conform to the Manhattan World assumption (The floors and ceilings are all flat planes, and the walls are perpendicular to them). (2) The room layout is always described as the corners or the floor boundary and ceiling boundary. These characteris- tics could be used as potential priors to guide the design of a reasonable layout estimation method. Based on the Manhattan World assumption, previous approaches [19, 20, 8, 14, 21] prefer to estimate the lay- out from a 1D sequence. They advocate compressing the extracted 2D feature maps in height dimension to obtain a 1D sequence, of which every element in this sequence share the same distortion magnitude (Fig. 1a). We argue that this compressed representation does not eliminate the panoramic distortions because there is no explicit distortion processing when extracting 2D feature maps before com- pression. Moreover, this strategy roughly mixes the vertical and horizontal planes together, confusing the semantics of different planes that are crucial for layout estimation. On the other hand, some researchers devoted themselves 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. 17337 to adopting different projection formats to boost the perfor- mance, e.g., the bird’s view of the room [29] and cubemap projection [27]. These projection-based schemes weaken the negative effect of the distortions. Nevertheless, frequent projection operations between different formats raise com- putational complexity. In addition, there exists an inevitable domain gap between the feature maps from different for- mats. To address the above problems, we propose a novel ar- chitecture (Fig.1b) that disentangles the orthogonal planes in advance to capture an explicit geometric cue. Follow- ing [21], our room layout can be recovered from the pre- dicted horizon-depth map and the room height. Therefore, the “clean” depth-relevant features and height-relevant fea- tures can both help with the layout estimation. To obtain such “clean” features free from the disturbance of decora- tions and furniture, we disentangle the widely used 1D rep- resentation into two separate sequences — the horizontal and vertical ones. Specifically, we pre-segment the vertical plane (i.e., walls) and horizontal planes (i.e., floors and ceil- ings) from the whole-room contextual information. Then these orthogonal planes are compressed into two 1D repre- sentations. Especially, based on the symmetry property be- tween the floor boundary and ceiling boundary, we design a soft-flipping fusion strategy to assist this process. Moreover, we propose an assembling mechanism to fuse multi-scale features with distortion awareness effectively. To eliminate the negative effect of distortion, we com- pute the attention among distortion-relevant positions fol- lowing distortion distribution patterns. Meanwhile, cross- scale interaction is carried out to integrate shallow geo- metric structures and deep semantic features. Considering the error of pre-segmentation, we further leverage triple at- tention to reconstruct the two 1D sequences. Particularly, we adopt graph-based attention to generate discriminative channels, self-attention to rebuild long-range dependencies, and cross-attention to provide the missing information for different sequences. To demonstrate our effectiveness, we conduct extensive experiments on four popular datasets — MatterportLayout [29], Zind [6], Stanford 2D-3D [1], and PanoContext [26]. The qualitative and quantitative results show that the pro- posed solution outperforms other SoTA methods. Our con- tributions can be summarized as follows: • We propose to disentangle orthogonal planes to cap- ture an explicit geometric cue for indoor 360° room layout estimation, with a soft-flipping fusion strategy to assist this procedure. • We design a cross-scale distortion-aware assembling mechanism to perceive distortion distribution as well as integrate shallow geometric structures and deep se- mantic features.• On popular benchmarks, our solution outperforms other SoTA schemes, especially on the metric of in- tersection over the union of 3D room layouts.
Sanghi_CLIP-Sculptor_Zero-Shot_Generation_of_High-Fidelity_and_Diverse_Shapes_From_Natural_CVPR_2023	Abstract Recent works have demonstrated that natural language can be used to generate and edit 3D shapes. However, these methods generate shapes with limited fidelity and di- versity. We introduce CLIP-Sculptor, a method to address these constraints by producing high-fidelity and diverse 3D shapes without the need for (text, shape) pairs during training. CLIP-Sculptor achieves this in a multi-resolution approach that first generates in a low-dimensional latent space and then upscales to a higher resolution for im- proved shape fidelity. For improved shape diversity, we use a discrete latent space which is modeled using a trans- former conditioned on CLIP’s image-text embedding space. We also present a novel variant of classifier-free guid- ance, which improves the accuracy-diversity trade-off. Fi- nally, we perform extensive experiments demonstrating that CLIP-Sculptor outperforms state-of-the-art baselines.	1. Introduction In recent years, there has been rapid progress in text- conditioned image generation [31, 32, 35], which has been driven by advances in multimodal understanding learned from web-scaled paired (text, image) data. These advanceshave led to applications in domains ranging from content creation [21, 22, 31] to human–robot interaction [40]. Un- fortunately, developing the analogue of a text-conditioned 3D shape generator is challenging because it is difficult to obtain (text, 3D shape) pairs at large scale. Prior work has attempted to address this problem by collecting text-shape paired data [2, 5, 9, 23, 26], but these approaches have been limited to a small number of object categories. A promising way around this data bottleneck is to use weak supervision from large-scale vision/language mod- els such as CLIP [30]. One approach is to directly op- timize a 3D representation such that (text, image render) pairs are aligned when projected into the CLIP embedding space. Prior work has applied this approach to stylize 3D meshes [25, 41] and to create abstract “dreamlike” objects using neural radiance fields [18] or meshes [19]. However, neither of the aforementioned methods produce realistic ob- ject geometry, and they can require expensive optimization. Another approach, more in line with text-to-image gener- ators [31, 32], is to train a conditional generative model. The CLIP-Forge system [37] builds such a model without paired (text, shape) data by using rendered images of shapes at training time and leveraging the CLIP embedding space to bridge the modalities of image and text at inference time. CLIP-Forge demonstrates compelling zero-shot generation 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. 18339 Table 1. High-level comparison between zero-shot text-to-shape generation methods. Inference time is calculated using a single NVIDIA Tesla V100 GPU. Method Inference Fidelity Diversity DreamFields [18] >24 hrs Low Single Clip-Mesh [19] 30 min Low Single Clip-Forge [37] 6.41 ms Medium Medium CLIP-Sculptor 0.91 sec High High abilities but produces low-fidelity shapes which do not cap- ture the full diversity of shapes found in the training distri- bution. In this paper, we propose CLIP-Sculptor, a text- conditioned 3D shape generative model that outperforms the state of the art by improving shape diversity and fi- delity with only (image, shape) pairs as supervision. CLIP- Sculptor’s novelty lies in its multi-resolution, voxel-based conditional generation scheme. Without (text, shape) pairs, CLIP-Sculptor learns to generate 3D shapes of common object categories from the text-image joint embedding of CLIP. To achieve high fidelity outputs, CLIP-Sculptor adopts a multi-resolution approach: it first generates a low- resolution latent grid representation that captures the se- mantics from text/image, then upscales it to a higher res- olution latent grid representation with a super-resolution model, and finally decodes the output geometry. To gen- erate diverse shapes, CLIP-Sculptor adopts discrete latent representations which are obtained using a vector quanti- zation scheme that avoids posterior collapse. To further improve shape fidelity and diversity, CLIP-Sculptor uses a masked transformer architecture. We additionally propose a novel annealed strategy for the classifier-free guidance [16]. To sum up, we make the following contributions: • CLIP-Sculptor, a multi-resolution text-conditional shape generative model that achieves both high fidelity and diversity without the need for (text, shape) pairs. • A novel variant of classifier-free guidance for gener- ative models, with an annealed guidance schedule to achieve better quality for a given diversity level.
Rony_Proximal_Splitting_Adversarial_Attack_for_Semantic_Segmentation_CVPR_2023	Abstract Classiﬁcation has been the focal point of research on ad- versarial attacks, but only a few works investigate methods suited to denser prediction tasks, such as semantic segmenta- tion. The methods proposed in these works do not accurately solve the adversarial segmentation problem and, therefore, overestimate the size of the perturbations required to fool models. Here, we propose a white-box attack for these mod- els based on a proximal splitting to produce adversarial perturbations with much smaller ℓ∞norms. Our attack can handle large numbers of constraints within a nonconvex minimization framework via an Augmented Lagrangian ap- proach, coupled with adaptive constraint scaling and mask- ing strategies. We demonstrate that our attack signiﬁcantly outperforms previously proposed ones, as well as classiﬁca- tion attacks that we adapted for segmentation, providing a ﬁrst comprehensive benchmark for this dense task.	1. Introduction Research on white-box adversarial attacks has mostly fo- cused on classiﬁcation tasks, with several methods proposed over the years for ℓp-norms [ 8,9,20–22,27,30,32,35,36,42]. In contrast, the literature on adversarial attacks for segmenta- tion tasks has been much scarcer, with few works proposing attacks [ 13,33,39]. The lack of studies on adversarial attacks in segmentation may appear surprising because of the promi- nence of this computer vision task in many applications where a semantic understanding of image contents is needed. In many safety-critical application areas, it is thus crucial to assess the robustness of the employed segmentation models. Although segmentation is treated as a per-pixel classiﬁ- cation problem, designing adversarial attacks for semantic segmentation is much more challenging for several reasons. First, from an optimization perspective, the problem of ad- versarial example generation is more difﬁcult. In a clas- siﬁcation task, producing minimal adversarial examples is a nonconvex constrained problem with a single constraint. In a segmentation task, this optimization problem now hasmultiple constraints, since at least one constraint must be addressed for each pixel in the image. For a dataset such as Cityscapes [ 19], the images have a size of 2048×1024 , resulting in more than 2 million constraints. Consequently, most attacks originally designed for classiﬁcation cannot be directly extended to segmentation. For instance, penalty methods such as C&W [ 9] cannot tackle multiple constraints since they rely on a binary search of the penalty weight. Second, the computational and memory cost of gener- ating adversarial examples can be prohibitive. White-box adversarial attacks usually rely on computing the gradient of a loss w.r.t. the input. In segmentation tasks, the dense outputs result in high memory usage to perform a backward propagation of the gradient. For reference, computing the gradients of the logits w.r.t. the input requires ∼22GiB of memory for FCN HRNetV2 W48 [ 38] on Cityscapes with a2048×1024 image. Additionally, most recent classiﬁ- cation adversarial attacks require between 100 and 1 000 iterations, resulting in a run-time of up to a few seconds per image [ 34,36] (on GPU) depending on the dataset and model. For segmentation, this increases to tens or even hundreds of seconds per image with larger models. In this article, we propose an adversarial attack to pro- duce minimal adversarial perturbations w.r.t. the ℓ∞-norm, for deep semantic segmentation models. Building on Aug- mented Lagrangian principles, we introduce adaptive strate- gies to handle a large number of constraints ( i.e.>106). Furthermore, we tackle the nonsmooth ℓ∞-norm minimiza- tion with a proximal splitting instead of gradient descent. In particular, we show that we can efﬁciently compute the prox- imity operator of the sum of the ℓ∞-norm and the indicator function of the space of possible perturbations. This results in an adversarial attack that signiﬁcantly outperforms the DAG attacks [ 39], in addition to several classiﬁcation attacks that we were able to adapt for segmentation. We propose a methodology to evaluate adversarial attacks in segmentation, and compare the different approaches on the Cityscapes [ 19] and Pascal VOC 2012 [ 23] datasets with a diverse set of archi- tectures, including well-known DeepLabV3+ models [ 11], the recently proposed transformer-based SegFormer [ 40], and the robust DeepLabV3 DDC-AT model from [ 41]. 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. 20524 (a) Pascal VOC 2012: ∥δ∥∞=0.67/255 (b) Cityscapes: ∥δ∥∞=0.53/255 Figure 1. Untargeted adversarial examples for FCN HRNetV2 W48 on Pascal VOC 2012 and Cityscapes. In both cases, more than 99% of pixels are incorrectly classiﬁed. For each dataset, left is the original image and its predicted segmentation, middle is the ampliﬁed perturbation and the ground truth segmentation and right is the adversarial image with its predicted segmentation. For Pascal VOC 2012, the predicted classes are TV monitor (blue), person (beige) and chair (bright red). proposed approach yields outstanding performances for all models and datasets considered in our experiments. For in- stance, our attack ﬁnds untargeted adversarial perturbations withℓ∞-norms lower than 1/255on average for all models on Cityscapes. With this attacker, we provide better means to as- sess the robustness of deep segmentation models, which has often been overestimated until now, as the existing attacks could only ﬁnd adversarial examples with larger norms.
Siddiqui_Panoptic_Lifting_for_3D_Scene_Understanding_With_Neural_Fields_CVPR_2023	Abstract We propose Panoptic Lifting, a novel approach for learn- ing panoptic 3D volumetric representations from images of in-the-wild scenes. Once trained, our model can render color images together with 3D-consistent panoptic segmen- tation from novel viewpoints. Unlike existing approaches which use 3D input directly or indirectly, our method re- quires only machine-generated 2D panoptic segmentation masks inferred from a pre-trained network. Our core con- tribution is a panoptic lifting scheme based on a neural field representation that generates a unified and multi-view con- sistent, 3D panoptic representation of the scene. To ac- count for inconsistencies of 2D instance identifiers across views, we solve a linear assignment with a cost based on the model’s current predictions and the machine-generated segmentation masks, thus enabling us to lift 2D instances to 3D in a consistent way. We further propose and ablate contributions that make our method more robust to noisy, machine-generated labels, including test-time augmenta- tions for confidence estimates, segment consistency loss, bounded segmentation fields, and gradient stopping. Ex- perimental results validate our approach on the challeng- ing Hypersim, Replica, and ScanNet datasets, improving by 8.4, 13.8, and 10.6% in scene-level PQ over state of the art.	1. Introduction Robust panoptic 3D scene understanding models are key to enabling applications such as VR, robot navigation, or self-driving, and more. Panoptic image understanding – the task of segmenting a 2D image into categorical “stuff” areas and individual “thing” instances – has experienced tremen- dous progress over the past years. These advances can be at- tributed to continuously improved model architectures and the availability of large-scale labeled 2D datasets, leading to state-of-the-art 2D panoptic segmentation models [6,19,43] that generalize well to unseen images captured in the wild. Single-image panoptic segmentation, unfortunately, is still insufficient for tasks requiring coherency and consis- tency across multiple views. In fact, panoptic masks often contain view-specific imperfections and inconsistent clas- sifications, and single-image 2D models naturally lack the ability to track unique object identities across views (see Fig. 2). Ideally, such consistency would stem from a full, 3D understanding of the environment, but lifting machine- generated 2D panoptic segmentations into a coherent 3D panoptic scene representation remains a challenging task. Project page: nihalsid.github.io/panoptic-lifting/ 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. 9043 Recent works [11,18,40,45] have addressed panoptic 3D scene understanding from 2D images by leveraging Neu- ral Radiance Fields (NeRFs) [22], gathering semantic scene data from multiple sources. Some works [11, 40] rely on ground truth 2D and 3D labels, which are expensive and time-consuming to acquire. The work of Kundu et al. [18] instead exploits machine-generated 3D bounding box detec- tion and tracking together with 2D semantic segmentation, both computed using off-the-shelf models. However, this approach is limited by the fact that 3D detection models, when compared to 2D panoptic segmentation ones, strug- gle to generalize beyond the data they were trained on. This is in large part due to the large difference in scale between 2D and 3D training datasets. Furthermore, dependence on multiple pre-trained models increases complexity and intro- duces potentially compounding sources of error. In this work we introduce Panoptic Lifting, a novel for- mulation which represents a static 3D scene as a panoptic radiance field (see Sec. 3.2). Panoptic Lifting supports ap- plications like novel panoptic view synthesis and scene edit- ing, while maintaining robustness to a variety of diverse in- put data. Our model is trained from only 2D posed images and corresponding, machine-generated panoptic segmenta- tion masks, and can render color, depth, semantics, and 3D- consistent instance information for novel views of the scene. Starting from a TensoRF [4] architecture that encodes density and view-dependent color information, we intro- duce lightweight output heads for learning semantic and instance fields. The semantic field, represented as a small MLP, is directly supervised with the machine-generated 2D labels. An additional segment consistency loss guides this supervision to avoid optima that fragment objects in the presence of label inconsistencies. The 3D instance field is modelled by a separate MLP, holding a fixed number of class-agnostic, 3D-consistent surrogate object identifiers. Losses for both the fields are weighted by confidence esti- mates obtained by test-time augmentation on the 2D panop- tic segmentation model. Finally, we discuss specific tech- niques, e.g. bounded segmentation fields and stopping semantics-to-geometry gradients (see Sec. 3.3), to further limit inconsistent segmentations. In summary, our contributions are: • A novel approach to panoptic radiance field represen- tation that models the radiance, semantic class and in- stance id for each point in the space for a scene by directly lifting machine-generated 2D panoptic labels. • A robust formulation to handle inherent noise and in- consistencies in machine-generated labels, resulting in a clean, coherent and view-consistent panoptic seg- mentations from novel views, across diverse data. Figure 2. Predictions from state-of-the-art 2D panoptic segmen- tation methods such as Mask2Former [6] are typically noisy and inconsistent when compared across views of the same scene. Typ- ical failure modes include conflicting labels (e.g. sofa predicted as a bed above) and segmentations (e.g. labeled void above). Fur- thermore, instance identities are not preserved across frames (rep- resented as different colors).
Schiappa_A_Large-Scale_Robustness_Analysis_of_Video_Action_Recognition_Models_CVPR_2023	Abstract We have seen a great progress in video action recognition in recent years. There are several models based on convolu- tional neural network (CNN) and some recent transformer based approaches which provide top performance on exist- ing benchmarks. In this work, we perform a large-scale robustness analysis of these existing models for video ac- tion recognition. We focus on robustness against real-world distribution shift perturbations instead of adversarial per- turbations. We propose four different benchmark datasets, HMDB51-P ,UCF101-P ,Kinetics400-P , and SSv2-P to per- form this analysis. We study robustness of sixstate-of-the-art action recognition models against 90different perturbations. The study reveals some interesting findings, 1) transformer based models are consistently more robust compared to CNN based models, 2) Pretraining improves robustness for Trans- former based models more than CNN based models, and 3) All of the studied models are robust to temporal perturba- tions for all datasets but SSv2; suggesting the importance of temporal information for action recognition varies based on the dataset and activities. Next, we study the role of augmen- tations in model robustness and present a real-world dataset, UCF101-DS , which contains realistic distribution shifts, to further validate some of these findings. We believe this study will serve as a benchmark for future research in robust video action recognition1.	1. Introduction Robustness of deep learning models against real-world distribution shifts is crucial for various applications in vision, such as medicine [4], autonomous driving [41], environment monitoring [60], conversational systems [36], robotics [68] and assistive technologies [5]. Distribution shifts with re- spect to training data can occur due to the variations in en- vironment such as changes in geographical locations, back- ground, lighting, camera models, object scale, orientations, *The authors contributed equally as first authors to this paper. †Corresponding author:[email protected] ‡The authors contributed equally as supervisors to this paper. 1More details available at bit.ly/3TJLMUF . Figure 1. Performance against robustness of action recognition models on UCF101-P. y-axis: relative robustness γr, x-axis: ac- curacy on clean videos, model names appended with P indicate pre-training, and the size of circle indicates FLOPs. motion patterns, etc. Such distribution shifts can cause the models to fail when deployed in a real world settings [26]. For example, an AI ball tracker that replaced human camera operators was recently deployed in a soccer game and repeat- edly confused a soccer ball with the bald head of a lineman, leading to a bad experience for viewers [13]. Robustness has been an active research topic due to its importance for real-world applications [4, 41, 60]. However, most of the effort is directed towards images [7, 25, 26]. Video is a natural form of input to the vision systems that function in the real world. Therefore studying robustness in videos is an important step towards developing reliable systems for real world deployment. In this work we perform a large-scale analysis on robustness of existing deep models for video action recognition against common real world spatial and temporal distribution shifts. Video action recognition provides an important test sce- nario to study robustness in videos given there are sufficient, large benchmark datasets and well developed deep learning models. Although the existing approaches have made im- pressive progress in action recognition, there are several fun- damental questions that still remain unanswered in the field. Do these approaches enable effective temporal modeling, the crux of the matter for action recognition approaches? Are these approaches robust to real-world corruptions like noise 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. 14698 (temporally consistent and inconsistent), blurring effects, etc? Do we really need heavy architectures for robustness, or are light-weight models good enough? Are the recently introduced transformer-based models, which give state of the art accuracy on video datasets, more robust? Does pre- training play a role in model robustness? This study aims at finding answers to some of these critical questions. Towards this goal, we present multiple benchmark datasets to conduct robustness analysis in video action recognition. We utilize four different widely used action recognition datasets including HMDB51 [34], UCF101 [54], Kinetics-400 [8], and SSv2 [43] and propose four corresponding benchmarks; HMDB51-P ,UCF101-P , Kinetics400-P , and SSv2-P . In order to create this benchmark, we introduce 90 different common perturbations which in- clude, 20 different noise corruptions, 15 blur perturbations, 15 digital perturbations, 25 temporal perturbations, and 15 camera motion perturbations as a benchmark. The study covers 6 different deep architectures considering different aspects, such as network size (small vs large), network archi- tecture (CNN vs Transformers), and network depth (shallow vs deep). This study reveals several interesting findings about ac- tion recognition models. We observe that recent transformer based models are not only better in performance , but they are also more robust than CNN models against most distri- bution shifts (Figure 1). We also observe that pretraining is more beneficial to transformers compared to CNN based models in robustness. We find that all the models are very robust against the temporal perturbations with minor drop in performance on the Kinetics, UCF and HMDB. However, on the SSv2 dataset, behavior of the models is different whereas the performance drops on different temporal perturbations. These observations show interesting phenomena about the video action recognition datasets, i.e., the importance of tem- poral information varies based on the dataset and activities . Next, we study the role of training with data augmenta- tions in model robustness and analyze the generalization of these techniques to novel perturbations. To further study the capability of such techniques, we propose a real-world dataset, UCF101-DS , which contains realistic distribution shifts without simulation. This dataset also helps us to bet- ter understand the behavior of CNN and Transformer-based models under realistic scenarios. We believe such findings will open up many interesting research directions in video action recognition and will facilitate future research on video robustness which will lead to more robust architectures for real-world deployment. We make the following contributions in this study, •A large-scale robustness analysis of video action recog- nition models to different real-world distribution shifts. •Provide insights including comparison of transformervs CNN based models, effect of pre-training, and effect of temporal perturbations on video robustness. •Four large-scale benchmark datasets to study robustness for video action recognition along with a real-world dataset with realistic distribution shifts.
Saragadam_WIRE_Wavelet_Implicit_Neural_Representations_CVPR_2023	Abstract Implicit neural representations (INRs) have recently ad- vanced numerous vision-related areas. INR performance depends strongly on the choice of activation function em- ployed in its MLP network. A wide range of nonlinearities have been explored, but, unfortunately, current INRs de- signed to have high accuracy also suffer from poor robust- ness (to signal noise, parameter variation, etc.). Inspired by harmonic analysis, we develop a new, highly accurate and robust INR that does not exhibit this tradeoff. Our Wavelet Implicit neural REpresentation (WIRE) uses as its ac- tivation function the complex Gabor wavelet that is well- known to be optimally concentrated in space–frequency and to have excellent biases for representing images. A wide range of experiments (image denoising, image inpainting, super-resolution, computed tomography reconstruction, im- age overfitting, and novel view synthesis with neural radi- ance fields) demonstrate that WIRE defines the new state of the art in INR accuracy, training time, and robustness.	1. Introduction Implicit neural representations (INRs), which learn a continuous function over a set of points, have emerged as a promising general-purpose signal processing framework. An INR consists of a multilayer perceptron (MLP) com- bining linear layers and element-wise nonlinear activation functions. Thanks to the MLP, INRs do not share the local- ity biases that limit the performance of convolutional neural networks (CNNs). Consequently, INRs have advanced the state of the art in numerous vision-related areas, including computer graphics [22, 27, 28], image processing [10], in- verse problems [42], and signal representations [41]. Currently, INRs still face a number of obstacles that limit their use. First, for applications with high-dimensional data such as 3D volumes, training an INR to high accuracy can still take too long (tens of seconds) for real time applica- tions. Second, INRs are not robust to signal noise or insuffi- cient measurements. Indeed, most works on INRs in the lit- WIRE Gauss SIREN Ground truth 1.0 0.0 Spatially spread -out errorSpatially compact but large errorSpatially compact and small error Error mapNonlinearity for Implicit Neural Representations Approximation accuracy with various nonlinearities Figure 1. Wavelet implicit neural representation (WIRE). We propose a new nonlinearity for implicit neural representations (INRs) based on the continuous complex Gabor wavelet that has high representation capacity for visual signals. The top row shows two commonly used nonlinearities: SIREN with sinusoidal non- linearity and Gaussian nonlinearity, and WIRE that uses a contin- uous complex Gabor wavelet. WIRE benefits from the frequency compactness of sine, and spatial compactness of a Gaussian non- linearity. The bottom row shows error maps for approximating an image with strong edges. SIREN results in global ringing arti- facts while Gaussian nonlinearity leads to compact but large error at edges. WIRE produces results with the smallest and most spa- tially compact error. This enables WIRE to learn representations accurately, while being robust to noise and undersampling of data. erature assume virtually no signal noise and large amounts of data. We find in our own experiments that current INR methods are ineffective for tasks such as denoising or super- resolution. Finally, INRs still have room for improvement in representational accuracy for fine details. In this paper, we develop a new, faster, more accurate, and robust INR that addresses these issues and takes INR performance to the next level. To achieve this, we take in- spiration from harmonic analysis and reconsider the nonlin- ear activation function used in the MLP. Recent work has shown that an INR can be interpreted as a structured signal 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
Ren_TinyMIM_An_Empirical_Study_of_Distilling_MIM_Pre-Trained_Models_CVPR_2023	Abstract Masked image modeling (MIM) performs strongly in pre- training large vision Transformers (ViTs). However, small models that are critical for real-world applications can- not or only marginally benefit from this pre-training ap- proach. In this paper, we explore distillation techniques to transfer the success of large MIM-based pre-trained mod- els to smaller ones. We systematically study different op- tions in the distillation framework, including distilling tar- gets, losses, input, network regularization, sequential dis- tillation, etc, revealing that: 1) Distilling token relations is more effective than CLS token- and feature-based distil- lation; 2) An intermediate layer of the teacher network as target perform better than that using the last layer when the depth of the student mismatches that of the teacher; 3) Weak regularization is preferred; etc. With these find- ings, we achieve significant fine-tuning accuracy improve- ments over the scratch MIM pre-training on ImageNet-1K classification, using all the ViT-Tiny, ViT-Small, and ViT- base models, with +4.2%/+2.4%/+1.4% gains, respectively. Our TinyMIM model of base size achieves 52.2 mIoU in AE20K semantic segmentation, which is +4.1 higher than the MAE baseline. Our TinyMIM model of tiny size achieves 79.6% top-1 accuracy on ImageNet-1K image classifica- tion, which sets a new record for small vision models of the same size and computation budget. This strong perfor- mance suggests an alternative way for developing small vision Transformer models, that is, by exploring better train- ing methods rather than introducing inductive biases into architectures as in most previous works. Code is available at https://github.com/OliverRensu/TinyMIM.	1. Introduction Masked image modeling (MIM), which masks a large portion of the image area and trains a network to recover the original signals for the masked area, has proven to be a very effective self-supervised method for pre-training vision Transformers [2, 11, 17, 49]. Thanks to its strong fine-tuning performance, MIM has now been a main-stream pre-training *Corresponding author: [email protected]. ViT-T ViT-S ViT-B7074788286Scratch MAE TinyMIM -0.6+3.6+0.7+3.1+2.4+3.8Acc. 72.279.981.2Figure 1. Comparison among TinyMIM (ours), MAE [17] and training from scratch by using ViT-T, -S and -B on ImageNet-1K. We report top-1 accuracy. We adopt DeiT [41] when training from scratch. For the first time, we successfully perform masked image modeling pre-training for smaller ViTs. ModelParam. Flops Top-1mIoU(M) (G) (%) DeiT-T [41] 5.5 1.3 72.2 38.0 PVT-T [43] 13.0 1.9 75.1 39.8 CiT-T [36] 5.5 1.3 75.3 38.5 Swin [29] 8.8 1.2 76.9 40.4 EdgeViT-XS [31] 6.4 1.1 77.5 42.1 MobileViTv1-S [30] 4.9 2.0 78.4 42.7 MobileViTv3-S [42] 4.8 1.8 79.3 43.1 TinyMIM⋆-T (Ours) 5.8 1.3 79.6 45.0 Table 1. Comparison with state-of-the-art tiny Transformers with architecture variants. The parameters indicate the backbone pa- rameter excluding the parameters of the last classification layer in classification or the decoder in segmentation. We report top-1 accuracy on ImageNet-1K classification and mIoU on ADE20K segmentation. method for vision Transformers, and numerous follow-ups have been carried out in this research line, such as study- ing how to set decoding architectures [21], reconstruction targets [10, 32, 45, 59], etc., as well as revealing its proper- ties [46, 48, 50]. 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. 3687 Method ViT-T ViT-S ViT-B ViT-L Scratch 72.2 79.9 81.2 82.6 MAE 71.6 80.6 83.6 85.9 Gap -0.6 +0.7 +2.4 +3.3 Table 2. Comparison between MAE pre-trained ViTs and ViTs trained from scratch by using ViT-T, -S, -B and -L on ImageNet- 1K. We adopt DeiT when training from scratch. We report top-1 accuracy. As model size shrinks, the superiority of MAE gradually vanishes. MAE even hurts the performance of ViT-T. However, as shown in Table 2, MIM pre-training [17] mainly effects for relatively large models. When the model size is as small as ViT-Tiny (5 million parameters), which is critical for real-world applications, MIM pre-training can even hurt the fine-tuning accuracy on ImageNet-1K classifi- cation. In fact, the accuracy drops by -0.6 compared to the counterpart trained from scratch. This raises a question: can small models also benefit from MIM pre-training, and how can this be achieved? In addition, the existing study on small vision Transform- ers mainly focus on introducing certain inductive bias into architecture design [6, 22, 30, 31]. The additional architec- tural inductive biases facilitate optimization yet limit the expressive capacity. It’s natural to ask whether we can boost plain small vision Transformers to perform just as well. In this work, we present TinyMIM, which answers the above questions. Instead of directly training small ViT mod- els using a MIM pretext task, TinyMIM uses distillation technology [20] to transfer the knowledge of larger MIM pre-trained models to smaller ones. Distillation endows the nice properties of larger MIM pre-trained models to smaller ones while avoiding solving a “too” difficult MIM task. Not- ing that knowledge distillation has been well developed, especially for supervised models [15], our main work is to systematically study for the first time the effects of different design options in a distillation framework when using MIM pre-trained models as teachers. Specifically, we consider dis- tillation targets, data augmentation, network regularization, auxiliary losses, macro distillation strategy, etc., and draw several useful findings: •Distillation targets . There are two main findings re- lated to distillation targets: 1) Distilling token relations is more effective than distilling the CLS token and fea- ture maps. 2) Using intermediate layers as the target may perform better than using the last layer, and the optimal target layer for different down-stream tasks, e.g., classification and segmentation, can be different. •Data and network regularization . Weak augmentation and regularization is preferred: 1) The performance of using a masked image is worse than using the originalimage; 2) Relatively small drop path rate (0 for teacher and 0.1 for student) performs best. •auxiliary losses . We find that an auxiliary MIM loss does not improve fine-tuning accuracy. •Macro distillation strategy . We find that using a se- quential distillation strategy, i.e., “ViT-B →ViT-S → ViT-T”, performs better than that distilling directly from ViT-B to ViT-T. By selecting the best framework options, we achieve sig- nificant fine-tuning accuracy improvements over the direct MIM pre-training on ImageNet-1K classification, using ViT models of different sizes, as shown in Figure 1. Specifi- cally, the gains of TinyMIM on the ViT-Tiny, ViT-Small, and ViT-base models are +4.2%/+2.4%/+1.4%, respectively. In particular, our TinyMIM⋆-T model with knowledge distillation during finetune-tuning achieves a top-1 accuracy of 79.6% on ImageNet-1K classification (see Table 1), which performs better than all previous works that develop small vision Transformer models by introducing architectural in- ductive biases or smaller feature resolutions. It sets a new accuracy record using similar model size and computation budget. On ADE20K semantic segmentation, TinyMIM⋆-T achieves 45.0 mIoU, which is +1.9 higher than the second best method, MobileViTv3-S [42]. The strong fine-tuning accuracy by TinyMIM⋆-T suggests an alternative way for developing small vision Transformer models, that is, by exploring better training methods rather than introducing inductive biases into architectures as most previous works have done.
Shen_Self-Supervised_3D_Scene_Flow_Estimation_Guided_by_Superpoints_CVPR_2023	Abstract 3D scene ﬂow estimation aims to estimate point-wise motions between two consecutive frames of point clouds. Superpoints, i.e., points with similar geometric features, are usually employed to capture similar motions of local regions in 3D scenes for scene ﬂow estimation. However, in existing methods, superpoints are generated with the ofﬂine clustering methods, which cannot characterize local regions with similar motions for complex 3D scenes well, leading to inaccurate scene ﬂow estimation. To this end, we propose an iterative end-to-end superpoint based scene ﬂow estimation framework, where the superpoints can be dynamically updated to guide the point-level ﬂow predic- tion. Speciﬁcally, our framework consists of a ﬂow guided superpoint generation module and a superpoint guided ﬂow reﬁnement module. In our superpoint generation module, we utilize the bidirectional ﬂow information at the previ- ous iteration to obtain the matching points of points and superpoint centers for soft point-to-superpoint association construction, in which the superpoints are generated for pairwise point clouds. With the generated superpoints, we ﬁrst reconstruct the ﬂow for each point by adaptively aggregating the superpoint-level ﬂow, and then encode the consistency between the reconstructed ﬂow of pairwise point clouds. Finally, we feed the consistency encod- ing along with the reconstructed ﬂow into GRU to reﬁne point-level ﬂow. Extensive experiments on several different datasets show that our method can achieve promising performance. Code is available at https://github. com/supersyq/SPFlowNet .	1. Introduction Scene ﬂow estimation is one of the vital components of numerous applications such as 3D reconstruction [10], Corresponding authors Yaqi Shen, Le Hui, Jin Xie, and Jian Yang are with PCA Lab, Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, and Jiangsu Key Lab of Image and Video Understanding for Social Security, School of Computer Science and Engineering, Nanjing University of Science and Technology, China. (a) Other clustering based methods Flow RefinementOffline Clustering Flow Estimationfixed superpoints not learnable learnable learnable (b) Our approachFlow Estimation Flow RefinementSuperpoint Generation dynamic superpointsFigure 1. Comparison with other clustering-based methods. (a) Other clustering based methods utilize ofﬂine clustering algorithms to split the point clouds into some ﬁxed superpoints for subsequent ﬂow reﬁnement, which is not learnable. (b) Our method embeds the differentiable clustering (superpoint generation) into our pipeline and generates dynamic superpoints at each iteration. We visualize part of the scene in FlyingThings3D [38] for better visualization. Different colors indicate different superpoints and red lines indicate the ground truth ﬂow. autonomous driving [37], and motion segmentation [2]. Estimating scene ﬂow from stereo videos and RGB-D images has been studied for many years [17, 19]. Recently, with the rapid development of 3D sensors, estimating scene ﬂow from two consecutive point clouds has receiving more and more attention. However, due to the irregularity and sparsity of point clouds, scene ﬂow estimation from point clouds is still a challenging problem in real scenes. In recent years, many 3D scene ﬂow estimation methods have been proposed [11, 34, 37, 56, 57, 59]. Most of these methods [34, 56] rely on dense ground truth scene ﬂow as supervision for model training. However, col- lecting point-wise scene ﬂow annotations is expensive and time-consuming. To avoid the expensive point-level an- notations, some efforts have been dedicated to weakly- supervised and self-supervised scene ﬂow estimation [9, 23, 46, 60]. For example, both Rigid3DSceneFlow [9] and LiDARSceneFlow [7] propose a weakly-supervised scene ﬂow estimation framework, which only take the ego- 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. 5271 motion and background masks as inputs. Especially, they utilize the DBSCAN clustering algorithm [8] to segment the foreground points into local regions with ﬂow rigid- ity constraints. In addition, RigidFlow [31] ﬁrst utilizes the off-line oversegmentation method [32] to decompose the source point clouds into some ﬁxed supervoxels, and then estimates the rigid transformations for supervoxels as pseudo scene ﬂow labels for model training. In summary, these clustering based methods utilize ofﬂine clustering algorithms with hand-crafted features (i.e., coordinates and normals) to generate the superpoints and use the consistent ﬂow constraints on these ﬁxed superpoints for scene ﬂow estimation. However, for some complex scenes, the ofﬂine clustering methods may cluster points with different ﬂow patterns into the same superpoints. Figure 1(a) shows that [32] falsely clusters points with the entirely different ﬂow into the same superpoint colored in purple (highlighted by the dotted circle). Thus, applying ﬂow constraints to the incorrect and ﬁxed superpoints for ﬂow estimation will mislead the model to generate false ﬂow results. To address this issue, we propose an iterative end-to- end superpoint guided scene ﬂow estimation framework (dubbed as “SPFlowNet”), which consists of an online su- perpoint generation module and a ﬂow reﬁnement module. Our pipeline jointly optimizes the ﬂow guided superpoint generation and superpoint guided ﬂow reﬁnement for more accurate ﬂow prediction (Figure 1(b)). Speciﬁcally, we ﬁrst utilize farthest point sampling (FPS) to obtain the initial superpoint centers, including the coordinate, ﬂow, and feature information. Then, we use the superpoint-level and point-level ﬂow information in the previous iteration to obtain the matching points of points and superpoint centers. With the pairs of points and superpoint centers, we can learn the soft point-to-superpoint association map. And we utilize the association map to adaptively aggregate the coordinates, features, and ﬂow values of points for superpoint center updating. Next, based on the updated superpoint-wise ﬂow values, we reconstruct the ﬂow of each point via the generated association map. Furthermore, we encode the consistency between the reconstructed ﬂow of pairwise point clouds. Finally, we feed the reconstructed ﬂow along with the consistency encoding into a gated recurrent unit to reﬁne the point-level ﬂow. Extensive experiments on several benchmarks show that our approach achieves state- of-the-art performance. Our main contributions are summarized as follows: We propose a novel end-to-end self-supervised scene ﬂow estimation framework, which iteratively gener- ates dynamic superpoints with similar ﬂow patterns and reﬁnes the point-level ﬂow with the superpoints. Different from other ofﬂine clustering based methods, we embed the online clustering into our model todynamically segment point clouds with the guidance from pseudo ﬂow labels generated at the last iteration. A superpoint guided ﬂow reﬁnement layer is intro- duced to reﬁne the point-wise ﬂow with superpoint- level ﬂow information, where the superpoint-wise ﬂow patterns are adaptively aggregated into the point-level with the learned association map. Our self-supervised scene ﬂow estimation method out- performs state-of-the-art methods by a large margin.
Sarkar_Parameter_Efficient_Local_Implicit_Image_Function_Network_for_Face_Segmentation_CVPR_2023	Abstract Face parsing is defined as the per-pixel labeling of im- ages containing human faces. The labels are defined to identify key facial regions like eyes, lips, nose, hair, etc. In this work, we make use of the structural consistency of the human face to propose a lightweight face-parsing method using a Local Implicit Function network, FP-LIIF. We propose a simple architecture having a convolutional encoder and a pixel MLP decoder that uses 1/26thnum- ber of parameters compared to the state-of-the-art models and yet matches or outperforms state-of-the-art models on multiple datasets, like CelebAMask-HQ and LaPa. We do not use any pretraining, and compared to other works, our network can also generate segmentation at different reso- lutions without any changes in the input resolution. This work enables the use of facial segmentation on low-compute or low-bandwidth devices because of its higher FPS and smaller model size.	1. Introduction Face parsing is the task of assigning pixel-wise labels to a face image to distinguish various parts of a face, like eyes, nose, lips, ears, etc. This segregation of a face image enables many use cases, such as face image editing [20, 46, 57], face e-beautification [37], face swapping [16, 35, 36], face completion [23]. Since the advent of semantic segmentation through the use of deep convolutional networks [31], a multitude of research has investigated face parsing as a segmentation problem through the use of fully convolutional networks [13, 14, 25, 26, 28, 29]. In order to achieve better results, some methods [14, 28] make use of conditional random fields (CRFs), in addition to CNNs. Other methods [25,27], *Work done during internship at Adobe Figure 1. The simple architecture of Local Implicit Image repre- sentation base FP-LIIF: A light convolutional encoder of modified resblocks followed by a pixel only MLP decoder focus on a two-step approach that predicts bounding boxes of facial regions (nose, eyes, hair. etc.) followed by segmentation within the extracted regions. Later works like AGRNET [48] and EAGR [49] claim that earlier ap- proaches do not model the relationship between facial com- ponents and that a graph-based system can model these statistics, leading to more accurate segmentation. In more recent research, works such as FaRL [59] in- vestigate pretraining on a human face captioning dataset. They pre-train a Vision Transformer (ViT) [8] and finetune on face parsing datasets and show improvement in com- parison to pre-training with classification based pre-training like ImageNet [44], etc., or no pre-training at all. The cur- rent state-of-the-art model, DML CSR [58], tackles the face parsing task using multiple concurrent strategies including multi-task learning, graph convolutional network (GCN), and cyclic learning. The Multi-task approach handles edge discovery in addition to face segmentation. The 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. 20970 GCN is used to provide global context instead of an aver- age pooling layer. Additionally, cyclic learning is carried out to arrive at an ensemble model and subsequently per- form self-distillation using the ensemble model in order to learn in the presence of noisy labels. In this work, we perform face segmentation by taking advantage of the consistency seen in human facial struc- tures. We take our inspiration from various face modeling works [1, 12, 61] that can reconstruct a 3D model of a face from 2D face images. These works show it is possible to create a low-dimensional parametric model of the human face in 3D. This led us to conclude that 2D modeling of the human face should also be possible with low dimen- sion parametric model. Recent approaches, like NeRF [34] and Siren [47] demonstrated that it is possible to reconstruct complex 3D and 2D scenes with implicit neural represen- tation. Many other works [2, 11, 43, 56] demonstrate that implicit neural representation can also model faces both in 3D and 2D. However, to map 3D and 2D coordinates to the RGB space, the Nerf [34] and Siren [47] variants of the models require training a separate network for every scene. This is different from our needs, one of which is that we must map an RGB image into label space and require a sin- gle network for the whole domain. That brings us to another method known as LIIF [3], which is an acronym for a Lo- cal Implicit Image Function and is used to perform image super-resolution. They learn an approximation of a contin- uous function that can take in any RGB image with low res- olution and output RGB values at the sub-pixel level. This allows them to produce an enlarged version of the input im- age. Thus, given the current success of learning implicit representations and the fact that the human face could be modeled using a low-dimension parametric model, we came to the conclusion that a low parameter count LIIF-inspired model should learn a mapping from a face image to its la- bel space or segmentation domain. In order to test this hy- pothesis, we modify a low-parameter version of EDSR [24] encoder such that it can preserve details during encoding. We also modify the MLP decoder to reduce the comput- ing cost of our decoder. Finally, we generate a probability distribution in the label space instead of RGB values. We use the traditional cross-entropy-based losses without any complicated training mechanisms or loss adaptations. An overview of the architecture is depicted in Figure 1, and more details are in Section 3. Even with a parameter count that is 1/26thcompared to DML CSR [58], our model at- tains state-of-the-art F1 and IoU results for CelebAMask- HQ [21] and LaPa [29] datasets. Some visualizations of our outputs are shared in Figure 3 and Figure 4. To summarise, our key contributions are as follows: • We propose an implicit representation-based simple and lightweight neural architecture for human face se- mantic segmentation.• We establish new state-of-the-art mean F1 and mean IoU scores on CelebAMask-HQ [21] and LaPa [29]. • Our proposed model has a parameter count of 1/26th or lesser compared to the previous state-of-the-art model. Our model’s SOTA configuration achieves an FPS of 110 compared to DML CSR’s FPS of 76.
Shin_Deep_Depth_Estimation_From_Thermal_Image_CVPR_2023	Abstract Robust and accurate geometric understanding against adverse weather conditions is one top prioritized condi- tions to achieve a high-level autonomy of self-driving cars. However, autonomous driving algorithms relying on the vis- ible spectrum band are easily impacted by weather and lighting conditions. A long-wave infrared camera, also known as a thermal imaging camera, is a potential res- cue to achieve high-level robustness. However, the miss- ing necessities are the well-established large-scale dataset and public benchmark results. To this end, in this pa- per, we first built a large-scale Multi-Spectral Stereo (MS2) dataset, including stereo RGB, stereo NIR, stereo thermal, and stereo LiDAR data along with GNSS/IMU informa- tion. The collected dataset provides about 195K synchro- nized data pairs taken from city, residential, road, campus, and suburban areas in the morning, daytime, and nighttime under clear-sky, cloudy, and rainy conditions. Secondly, we conduct an exhaustive validation process of monocu- lar and stereo depth estimation algorithms designed on visible spectrum bands to benchmark their performance in the thermal image domain. Lastly, we propose a uni- fied depth network that effectively bridges monocular depth and stereo depth tasks from a conditional random field approach perspective. Our dataset and source code are available at https://github.com/UkcheolShin/ MS2-MultiSpectralStereoDataset .	1. Introduction Recently, a number of researches have been conducted for accurate and robust geometric understanding in self- driving cars based on the widely-used benchmark datasets, such as KITTI [15], DDAD [17], and nuScenes [4]. Mod- ern computer vision algorithms deploy a deep neural net- work and data-driven machine learning technique to achieve high-level accuracy, which needs large-scale datasets. How- ever, from the perspective of robustness in real-world, the algorithms mostly rely on visible spectrum images and are easily degenerated by weather and lighting conditions. (a) Depth from thermal images via unified depth network (b) RGB (Reference) (c) Thermal image (d) Depth map Figure 1. Depth from thermal images in various environments. Our proposed network can estimate both monocular and stereo depth maps regardless of given a single or stereo thermal image via unified network architecture. Furthermore, depth estimation results from thermal images show high-level reliability and robust- ness under day-light, low-light, and rainy conditions. Therefore, recent works have actively investigated alter- native sensors such as Near-Infrared (NIR) cameras [39], LiDARs [16, 51], radars [14, 32], and long-wave infrared (LWIR) cameras [35, 45] to achieve reliable and robust geometric understanding in adverse conditions. Among these alternative and complementary sensors, LWIR cam- era ( i.e., thermal camera) has become more popular be- cause of its competitive price, adverse weather robustness, and unique modality information ( i.e., temperature). There- fore, various thermal image based computer vision solu- tions [3, 21–23, 27, 35, 45, 47–50] to achieve high-level ro- bustness have been actively attracting attention recently. 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. 1043 Table 1. Comprehensive comparison of multi-modal datasets . Compared to previous datasets [6, 8, 24, 25, 28, 53], the proposed Multi- Spectral Stereo (MS2) dataset provides about 195K synchronized and rectified multi-spectral stereo sensor data ( i.e., RGB, NIR, thermal, LiDAR, and GNSS/IMU data) covering diverse locations ( e.g., city, campus, residential, road, and suburban), times ( e.g., morning, daytime, and nighttime), and weathers ( e.g., clear-sky, cloudy, and rainy). Dataset Year Environment PlatformTotal # ofLiDAR IMURGB NIR Thermal Weather Data Pairs Mono Stereo Mono Stereo Mono Stereo RAW Daytime Nighttime Rain CATS [53] 2017 In/Outdoor Handheld 1.4K ✓ ✓ ✓ ✓ ✕ ✕ ✓ ✓ ✓ ✓ ✓ ✕ KAIST [6] 2018 Outdoor Vehicle Unknown ✓ ✓ ✓ ✓ ✕ ✕ ✓ ✕ ✓ ✓ ✓ ✕ ViViD [25] 2019 In/Outdoor Handheld 5.3K/4.3K ✓ ✓ ✓ ✕ ✕ ✕ ✓ ✕ ✓ ✓ ✓ ✕ MultiSpectralMotion [8] 2021 In/Outdoor Handheld 121K/27.3K ✓ ✓ ✓ ✕ ✓ ✕ ✓ ✕ ✓ ✓ ✓ ✕ ViViD++ [24] 2022 Outdoor Vehicle 56K ✓ ✓ ✓ ✕ ✕ ✕ ✓ ✕ ✓ ✓ ✓ ✕ OdomBeyondVision [28] 2022 IndoorHandheld/ 71K/✓ ✓ ✓ ✓ ✓ ✓ ✓ ✕ ✓ ✓ ✓ ✕UGV/UA V 117K/21K Ours 2022 Outdoor Vehicle 195K ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ However, the missing necessities are the well- established large-scale dataset and public benchmark results. The publicly available datasets for autonomous driving are overwhelmingly composed of the visible spec- trum band ( i.e., RGB images), but it very rarely includes other spectrum bands, such as the NIR band and LWIR band. Especially, despite the advantage of the LWIR band, just a few LWIR datasets have been recently released. However, these datasets are indoor oriented [8, 25, 28], small scale [25, 53], publicly unavailable [6], or limited sensor diversity [6, 24]. Therefore, the necessity is getting increase to design a large-scale multi-sensor driving dataset to investigate the feasibility and challenges associated with an autonomous driving perception system from multi-spectral sensors. The other necessity is thoroughly validating vision ap- plications on the LWIR band. Estimating a depth map from monocular and stereo images is one fundamental task for geometric understanding. Despite numerous recent stud- ies in depth estimation, these works have mainly focused on depth estimation using RGB images. However, thermal images, which typically have lower resolution, less texture, and more noise than RGB images, could pose a challenge for stereo-matching algorithms. This means that the perfor- mances of these previous works in thermal image domains are uncertain and may not be guaranteed. To this end, in this paper, we provide a large-scale multi- spectral dataset along with exhaustive experimental results and a new perspective of depth unification to encourage ac- tive research of various geometry algorithms from multi- spectral data to achieve high-level performance, reliability, and robustness against hostile conditions. Our contributions can be summarized as follows: • We provide a large-scale Multi-Spectral Stereo (MS2) dataset, including stereo RGB, stereo NIR, stereo ther- mal, and stereo LiDAR data along with GNSS/IMU data. Our dataset provides about 195K synchronized data pairs taken from city, residential, road, campus, and suburban areas in the morning, daytime, and night- time under clear-sky, cloudy, and rainy conditions.• We perform exhaustive validation and investigate that monocular and stereo depth estimation algorithms originally designed for visible spectral bands work rea- sonably in thermal spectral bands. • We propose a unified depth network that bridges monocular depth and stereo depth estimation tasks from the perspective of a conditional random field ap- proach.
Ju_Human-Art_A_Versatile_Human-Centric_Dataset_Bridging_Natural_and_Artificial_Scenes_CVPR_2023	Abstract Humans have long been recorded in a variety of forms since antiquity. For example, sculptures and paintings were the primary media for depicting human beings before the in- vention of cameras. However, most current human-centric computer vision tasks like human pose estimation and hu- man image generation focus exclusively on natural images in the real world. Artificial humans, such as those in sculp- tures, paintings, and cartoons, are commonly neglected, making existing models fail in these scenarios. As an abstraction of life, art incorporates humans in both natural and artificial scenes. We take advantage of it and introduce the Human-Art dataset to bridge related tasks in natural and artificial scenarios. Specifically, Human-Art contains 50khigh-quality images with over 123kperson instances from 5natural and 15artificial scenarios, which are annotated with bounding boxes, keypoints, self-contact points, and text information for humans represented in both 2D and 3D. It is, therefore, comprehensive and versatile for various downstream tasks. We also provide a rich set of baseline results and detailed analyses for related tasks, including human detection, 2D and 3D human pose estima- tion, image generation, and motion transfer. As a challeng- ing dataset, we hope Human-Art can provide insights for relevant research and open up new research questions.	1. Introduction " Art is inspired by life but beyond it." Human-centric computer vision (CV) tasks such as hu- man detection [39], pose estimation [30], motion trans- fer [59], and human image generation [35] have been in- tensively studied and achieved remarkable performances in the past decade, thanks to the advancement of deep learning techniques. Most of these works use datasets [14,23,27,39] *Equal contribution. †Work in part done during an internship at IDEA. ‡Corresponding author.that focus on humans in natural scenes captured by cameras due to the practical demands and easy accessibility. However, besides being captured by cameras, humans are frequently present and recorded in various other forms, from ancient murals on walls to portrait paintings on pa- per to computer-generated virtual figures in digital form. However, existing state-of-the-art (SOTA) human detection and pose estimation models [64, 71] trained on commonly used datasets such as MSCOCO [27] generally work poorly on these scenarios. For instance, the average precision of such models can be as high as 63.2% and 79.8% on natural scenes but drops significantly to 12.6% and 28.7% on the sculpture scene. A fundamental reason is the domain gap between natural and artificial scenes. Also, the scarcity of datasets with artificial human scenes significantly restricts the development of tasks such as anime character image generation [9, 67, 74], character rendering [29], and char- acter motion retargeting [1, 37, 68] in computer graphics and other areas. With the growing interest in virtual reality (VR), augmented reality (AR), and metaverse, this problem is exacerbated and demands immediate attention. There are a few small datasets incorporating humans in artificial environments in the literature. Sketch2Pose [4] and ClassArch [33] collect images in sketch and vase paint- ing respectively. Consequently, they are only applicable to the corresponding context. People-Art [61] is a human de- tection dataset that consists of 1490 paintings. It covers arti- ficial scenes in various painting styles, but its categories are neither mutually exclusive nor collectively comprehensive. More importantly, the annotation type and image number in People-Art are limited, and hence this dataset is mainly used for testing (instead of training) object detectors. Art presents humans in both natural and artificial scenes in various forms, e.g., dance, paintings, and sculptures. In this paper, we take advantage of the classification of vi- sual arts to introduce Human-Art , a versatile human-centric dataset, to bridge the gap between natural and artificial scenes. Human-Art is hierarchically structured and includes high-quality human scenes in rich scenarios with precise 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. 618 Acrobatics Cosplay Dance Drama Movie Garage Kits Relief Sculpture Kids Drawing Mural Oil Painting Shadow Play Sketch Stained Glass Ukiyoe Generated Cartoon Digital Art Ink Painting WatercolorNatural HumanArtificial Human Acrobatics Shadow PlaySculpture Oil PaintingRelief Garage Kits MovieDramaDanceCosplay Digital Art CartoonUkiyoe Stained Glass SketchMural Kids Drawing Generated Watercolor Ink Painting Visual Art Painting Sculpture Drama Dance Movie Acrobatics Shadow Play Oil PaintingRelief Garage KitsCosplay Digital Art CartoonUkiyoe Stained Glass SketchMural Kids Drawing GeneratedWatercolor Ink Painting Human DetectionPose EstimationMotion TransferImage GenerationSupported TasksFigure 1. Human-Art is a versatile human-centric dataset to bridge the gap between natural and artificial scenes. It includes 20high-quality scenes, including natural and artificial humans in both 2D representation (yellow dashed boxes) and 3D representation (blue solid boxes). manual annotations. Specifically, it is composed of 50k images with about 123kperson instances in 20artistic cate- gories, including 5natural and 15artificial scenarios in both 2D and 3D, as shown in Fig. 1. To support both recog- nition and generation tasks, Human-Art provides precise manual annotations containing human bounding boxes, 2D keypoints, self-contact points, and text descriptions. It can compensate for the lack of scenarios in prior datasets (e.g., MSCOCO [27]), link virtual and real worlds, and introduce new challenges and opportunities for human-centric areas. Human-Art has the following unique characteristics: •Rich scenario :Human-Art focuses on scenes missing in mainstream datasets (e.g., [27]), which covers most human-related scenarios. Challenging human appear- ances, diverse contexts, and various poses largely com- plement the scenario deficiency of existing datasets and will open up new challenges and opportunities. •High quality : We guarantee inter-category variabil- ity and intra-category diversity in style, author, origin, and age. The 50k images are manually selected from 1,000k carefully collected images using standardized data collection, filtering, and consolidating processes. •Versatile annotations :Human-Art provides carefully manual annotations of 2D human keypoints, human bounding boxes, and self-contact points to support var- ious downstream tasks. Also, we provide accessible text descriptions to enable multi-modality learning. With Human-Art , we conduct comprehensive experi- ments and analysis on various downstream tasks including human detection, human pose estimation, human mesh re- covery, image generation, and motion transfer. Although training on Human-Art can lead to a separate 31% and 21% performance boost on human detection and human pose es- timation, results demonstrate that human-related CV tasks still have a long way to go before reaching maturity.
Kang_Superclass_Learning_With_Representation_Enhancement_CVPR_2023	Abstract In many real scenarios, data are often divided into a handful of artificial super categories in terms of ex- pert knowledge rather than the representations of images. Concretely, a superclass may contain massive and vari- ous raw categories, such as refuse sorting. Due to the lack of common semantic features, the existing classifica- tion techniques are intractable to recognize superclass with- out raw class labels, thus they suffer severe performance damage or require huge annotation costs. To narrow this gap, this paper proposes a superclass learning framework, called SuperClass Learning with Representation Enhance- ment(SCLRE), to recognize super categories by leverag- ing enhanced representation. Specifically, by exploiting the self-attention technique across the batch, SCLRE col- lapses the boundaries of those raw categories and enhances the representation of each superclass. On the enhanced representation space, a superclass-aware decision bound- ary is then reconstructed. Theoretically, we prove that by leveraging attention techniques the generalization error of SCLRE can be bounded under superclass scenarios. Exper- imentally, extensive results demonstrate that SCLRE outper- forms the baseline and other contrastive-based methods on CIFAR-100 datasets and four high-resolution datasets.	1. Introduction In recent decades, basic-level raw categorization (e.g. cats vs dogs, apples vs bananas) has greatly developed [9, 27] while high-level or super-coarse-grained visual cat- egorization (e.g., recyclable waste vs kitchen waste, crea- tures vs non-creatures) has received little attention. In many real scenarios, there often exist a handful of high-level cat- egories, wherein numerous images from diverse basic-level categories share one common label. We tend to define this kind of super-coarse-grained class as Superclass. *Corresponding Author EmbeddingSpace SharedFeature LocalFeatureLocalFeature DataSpace“Apple”Domain DataSpaceFruit DomainToyDomain DataSpace “Rotten”Domain(a) Image Feature Domain(b) Separate Fine Domain(c) Superclass Concept DomainFigure 1. Illustration of superclass learning. The samples from a same superclass will be scatteredly distributed in the embed- ding space. The process of superclass learning is to break old domains and construct new domains. Red indicates the superclass of kitchen waste, and blue indicates the superclass of recyclable waste. Refuse sorting, as an example, is such a recognization problem with four superclasses, i.e., kitchen waste, recy- clable waste, hazardous waste, and others waste. One task of refuse sorting is to accurately collect various items, such as rotten fruits, bones, raw vegetables, and eggshells, into kitchen waste. Traditional recognition needs to identify what exact basic-level categories they are, then sort them out. Obviously, it is wasteful and unrealistic for superclass identification. Essentially, high-level superclasses contain two charac- teristics, remarkably distinct from basic-level classes. First, the basic-level classes contained in superclass problems are usually scattered and share few common features. As de- picted in the top-left corner subgraph of Fig. 1, the fruit apple, bone, and eggs are remote from each other in fea- ture spaces, though all of them belong to kitchen waste. 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. 24060 Second, the instances from two distinct superclasses may share common features. Just as illustrated in Fig. 1, fruit apple, from kitchen waste, and toy apple, from recyclable waste, are close to each other as they share common seman- tic features. Obviously, the above-mentioned characteristics indicate that the smoothness assumption [27] under basic- level classification (nearby images tend to have the same label) does not hold in the superclass scenarios. Thus, the existing classification techniques based on smoothness as- sumption are not practically deployable and scalable and they may suffer severe performance damage in superclass settings. Accordingly, it is valuable and promising to inves- tigate superclass identification. To tackle the superclass problems, we have to address two main challenges. First, we need to break the origi- nal decision boundaries of basic-level classes and disclose a superclass-aware boundary at the basic class level. As depicted in bottom subgraph (a) of Fig. 1, the boundary of the apple domain is original, however, it is useless and even harmful for the refuse sorting as both fruit apples and toy apples belong to this domain. To get the required domain boundary, the apple domain needs to be separated into the fruit domain and toy domain by leveraging their individ- ual local features, as depicted in the bottom subgraph (b) of Fig. 1. In superclass scenarios, it is not enough to in- vestigate the boundary at a basic class level. Consequently, the second challenge is to reconstruct a decision boundary at a superclass level. To achieve this end, it is necessary to merge the domain of classes, such as fruit apples, eggs, and bones into a new rotten superclass domain, as depicted in the bottom subgraph (c) of Fig. 1. In this paper, we propose a SuperClassLearning frame- work with Representation Enhancement. Considering that the semantic representation at the basic class level is not workable for superclass recognition, we propose one cross- instance attention module which could seize the representa- tion across the instances with the same superclass label. By leveraging contrastive adjustment loss, the attention mech- anism enhances this representation. Moreover, to overcome the imbalance distribution of superclasses, we adopt target adjustment loss to reconstruct a superclass-aware decision boundary on the enhanced representation space. In summary, this paper makes the following contribu- tions: • We propose an under-study but realistic problem, su- perclass identification, that has notably distinct distri- bution from basic-level categorization. • We propose a novel representation enhancement method by leveraging cross-instance attention and then exploit it in superclass identification. And by theo- retical analyses, we verify that this self-attention tech- nique can bound the generalization error of superclassrecognition. • Extensive experiments demonstrate that SCLRE out- performs the SOTA classification techniques on one artificial superclass dataset and three real datasets. The remainder of the paper is organized as follows: in Sec. 2 we briefly review related work and in Sec. 3 we de- scribe our method for superclass recognition. Then, exten- sive experiments and generalization error analysis are con- ducted in Secs. 4 and 5. Finally, Sec. 6 draws a brief con- clusion.
Ju_Distilling_Vision-Language_Pre-Training_To_Collaborate_With_Weakly-Supervised_Temporal_Action_Localization_CVPR_2023	Abstract Weakly-supervised temporal action localization (WTAL) learns to detect and classify action instances with only cat- egory labels. Most methods widely adopt the off-the-shelf Classification-Based Pre-training (CBP) to generate video features for action localization. However, the different opti- mization objectives between classification and localization, make temporally localized results suffer from the serious in- complete issue. To tackle this issue without additional anno- tations, this paper considers to distill free action knowledge from Vision-Language Pre-training (VLP), as we surpris- ingly observe that the localization results of vanilla VLP have an over-complete issue, which is just complementary to the CBP results. To fuse such complementarity, we pro- pose a novel distillation-collaboration framework with two branches acting as CBP and VLP respectively. The frame- work is optimized through a dual-branch alternate training strategy. Specifically, during the B step, we distill the confi- dent background pseudo-labels from the CBP branch; while during the F step, the confident foreground pseudo-labels are distilled from the VLP branch. As a result, the dual- branch complementarity is effectively fused to promote one strong alliance. Extensive experiments and ablation studies on THUMOS14 and ActivityNet1.2 reveal that our method significantly outperforms state-of-the-art methods.	1. Introduction Temporal action localization (TAL), which aims to lo- calize and classify action instances from untrimmed long videos, has been recognized as an indispensable component of video understanding [13,72,92]. To avoid laborious tem- poral boundary annotations, the weakly-supervised setting (WTAL) [31, 62, 64, 78], i.e.only video-level category la- bels are available, has gained increasing attentions. To date in the literature, almost all WTAL methods rely on Classification-Based Pre-training (CBP) for video fea- → FP & FN → TP & TNCBPTrue Negative (TN) False Negative (FN) →→ (B) VLP branchCBP branch CollaborationDistill BackgroundDistill ForegroundVLPInput Videos GT(A) True Positive (TP) False Positive (FP) →→Figure 1. (A) Complementarity. Most works use Classification- Based Pre-training (CBP) for localization, bringing high TN yet serious FN. Vanilla Vision-Language Pre-training (VLP) confuses action and background, bringing high TP yet serious FP. (B) Our distillation-collaboration framework distills foreground from the VLP branch while background from the CBP branch, and pro- motes mutual collaboration, bringing satisfactory results. ture extraction [5, 73]. A popular pipeline is to train an ac- tion classifier with CBP features, then threshold the frame- level classification probabilities for final localization re- sults. As demonstrated in Figure 1 (A), these CBP meth- ods suffer from the serious incompleteness issue ,i.e.only detecting sparse discriminative action frames and incurring high false negatives . The main reason is that the optimiza- tion objective of classification pre-training, is to find sev- eral discriminative frames for action recognition, which is far from the objective of complete localization. As a re- sult, features from CBP are inevitably biased towards par- tial discriminative frames. To solve the incompleteness is- sue, many efforts [33, 43, 56, 96] have been attempted, but most of them are trapped in a ‘performance-cost dilemma’, namely, solely digging from barren category labels to keep costs low. Lacking location labels fundamentally limits the performance, leaving a huge gap from strong supervision. To jump out of the dilemma, this paper raises one novel question: is there free action knowledge available, to help 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. 14751 complete detection results while maintain cheap annotation overhead at the same time? We naturally turn our sights to the prevalent Vision-Language Pre-training (VLP) [22, 66]. VLP has demonstrated great success to learn joint visual- textual representations from large-scale web data. As lan- guage covers rich information about objects, human-object interactions, and object-object relationships, these learned representations could provide powerful human-object co- occurrence priors: valuable gifts for action localization. We here take one step towards positively answering the question, i.e.fill the research gap of distilling action priors from VLP, namely CLIP [66], to solve the incomplete issue for WTAL. As illustrated in Figure 1 (A), we first naively evaluate the temporal localization performance of VLP by frame-wise classification. But the results are far from satis- factory, suffering from the serious over-complete issue ,i.e. confusing multiple action instances into a whole, causing high false positives . We conjecture the main reasons are: (1) due to data and computational burden, almost all VLPs are trained using image-text pairs. Hence, VLP lacks suffi- cient temporal knowledge and relies more on human-object co-occurrence for localization, making it struggle to distin- guish the actions with visually similar background contexts; (2) some background contexts have similar (confusing) tex- tual semantics to actions, such as run-up vs.running. Although simply steering VLP for WTAL is infeasible, we fortunately observe the complementary property be- tween CBP and VLP paradigms: the former localizes high true negatives but serious false negatives, while the latter has high true positives but serious false positives. To lever- age the complementarity, as shown in Figure 1 (B), we de- sign a novel distillation-collaboration framework that uses two branches to play the roles of CBP and VLP, respec- tively. The design rationale is to distill background knowl- edge from the CBP branch, while foreground knowledge from the VLP branch, for strong alliances. Specifically, we first warm up the CBP branch using only category super- vision to initialize confident background frames, and then optimize the framework via an alternating strategy. Dur- ing B step , we distill background pseudo-labels for the VLP branch to solve the over-complete issue, hence obtaining high-quality foreground pseudo-labels. During F step , we leverage high-quality pseudo-labels for the CBP branch to tackle the incomplete issue. Besides, in each step, we intro- duce both confident knowledge distillation and representa- tion contrastive learning for pseudo-label denoising, effec- tively fusing complementarity for better results. On two standard benchmarks: THUMOS14 and Activi- tyNet1.2, our method improves the average performance by 3.5% and 2.7% over state-of-the-art methods. We also con- duct extensive ablation studies to reveal the effectiveness of each component, both quantitatively and qualitatively. To sum up, our contributions lie in three folds:•We pioneer the first exploration in distilling free action knowledge from off-the-shelf VLP to facilitate WTAL; •We design a novel distillation-collaboration framework that encourages the CBP branch and VLP branch to comple- ment each other, by an alternating optimization strategy; •We conduct extensive experiments and ablation studies to reveal the significance of distilling VLP and our superior performance on two public benchmarks.
Kang_Distilling_Self-Supervised_Vision_Transformers_for_Weakly-Supervised_Few-Shot_Classification__Segmentation_CVPR_2023	Abstract We address the task of weakly-supervised few-shot im- age classification and segmentation, by leveraging a Vision Transformer (ViT) pretrained with self-supervision. Our proposed method takes token representations from the self- supervised ViT and leverages their correlations, via self- attention, to produce classification and segmentation pre- dictions through separate task heads. Our model is able to effectively learn to perform classification and segmen- tation in the absence of pixel-level labels during train- ing, using only image-level labels. To do this it uses at- tention maps, created from tokens generated by the self- supervised ViT backbone, as pixel-level pseudo-labels. We also explore a practical setup with “mixed” supervision, where a small number of training images contains ground- truth pixel-level labels and the remaining images have only image-level labels. For this mixed setup, we propose to im- prove the pseudo-labels using a pseudo-label enhancer that was trained using the available ground-truth pixel-level la- bels. Experiments on Pascal-5iand COCO-20idemonstrate significant performance gains in a variety of supervision settings, and in particular when little-to-no pixel-level la- bels are available.	1. Introduction Few-shot learning [22, 23, 77] is the problem of learn- ing to perform a prediction task using only a small number of examples ( i.e. supports) of the completed task, typically in the range of 1-10 examples. This problem setting is ap- pealing for applications where large amounts of annotated data are expensive or impractical to collect. In computer vi- sion, few-shot learning has been actively studied for image classification [24, 36, 46, 66, 70, 85] and semantic segmen- tation [17, 58, 61, 88, 89]. In this work, we focus on the combined task of few-shot classification and segmentation (FS-CS) [33], which aims to jointly predict (i) the presence of each support class, i.e., multi-label classification, and (ii) *Work done during an internship at FAIR. Weakly-supervisedfew-shotlearner3-way 1-shot support setwith no ground-truth masks query image 32class 1class 2class 3 Figure 1. Weakly-supervised few-shot classification and seg- mentation. In this paper, we explore training few-shot models using little-to-no ground-truth segmentation masks. its pixel-level semantic segmentation. Few-shot learners are typically trained either by meta- learning [24, 56, 63, 71] or by transfer learning with fine- tuning [11, 12, 16]. Both these paradigms commonly as- sume the availability of a large-scale and fully-annotated training dataset. For FS-CS, we need Ground-Truth (GT) segmentation masks for query and support images during training. We also need these GT masks for support im- ages during testing. A natural alternative to collecting such expensive segmentation masks would be to instead em- ploy a weaker supervision, e.g. image-level [52,59] or box- level labels [13, 30], for learning, i.e., to adopt a weakly- supervised learning approach [5, 41, 43, 51]. Compared to conventional weakly-supervised learning however, weakly- supervised few-shot learning has rarely been explored in the literature and is significantly more challenging. This is be- cause in few-shot learning, object classes are completely disjoint between training and testing, resulting in models which are susceptible to severe over-fitting to the classes seen during training. In this work we tackle this challenging scenario, address- ing weakly-supervised FS-CS where only image-level la- bels, i.e. class labels, are available ( cf.Fig. 1). Inspired 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. 19627 recent work by Caron et al. [9], which showed that pixel- level semantic attention emerges from self-supervised vi- sion transformers (ViTs) [18], we leverage attention maps from a frozen self-supervised ViT to generate pseudo-GT segmentation masks. We train an FS-CS learner on top of the frozen ViT using the generated pseudo-GT masks as pixel-level supervision. The FS-CS learner is a small trans- former network that takes features from the ViT as input and is trained to predict the pseudo-GT masks. Our com- plete model thus learns to segment using its own intermedi- ate byproduct, with no segmentation labeling cost, in a form of distillation [2, 28, 91, 94] between layers within a model. We also explore a practical training setting where a lim- ited number of training images have both image-level and pixel-level annotations, while the remaining images have only image-level labels. For this mixed-supervised setting, we propose to train an auxiliary mask enhancer that refines attention maps into pseudo-GT masks. The enhancer is su- pervised using the small number of available GT masks and is used to generate the final pseudo-GT masks for the train- ing images that do not have GT masks. Lastly, to e ffectively address this weakly-supervised FS- CS task, we propose a Classification-Segmentation Trans- former (CST) architecture. CST takes as input ViT tokens for query and support images, computes correlations be- tween them, and then predicts classification and segmen- tation outputs through separate task heads; the classifica- tion head is trained with class labels while the segmenta- tion head is trained with either GT masks or pseudo-GT masks, depending on availability. Unlike prior work us- ing ViTs that generate prediction for a single task using ei- ther the class token [18, 74] or image tokens [50, 65], CST uses each of them for each task prediction, which proves ad- vantageous for both tasks. CST achieves moderate-to-large performance gains for FS-CS on all three supervision lev- els (image-level-, mixed-level, and pixel-level), when com- pared to prior state-of-the-art methods. Our contributions are the following: i. We introduce a powerful baseline for few-shot classi- fication and segmentation (FS-CS) using only image- level supervision, which leverages localized semantic information encoded in self-supervised ViTs [9]. ii. We propose a learning setup for FS-CS with mixed su- pervision, and present an auxiliary mask enhancer to improve performance compared to image-level super- vision only. iii. We design Classification-Segmentation Transformer, and a multi-task learning objective of classification and segmentation, which is beneficial for both tasks, and al- lows for flexibly tuning the degree of supervision.
Kang_Meta-Learning_With_a_Geometry-Adaptive_Preconditioner_CVPR_2023	Abstract Model-agnostic meta-learning (MAML) is one of the most successful meta-learning algorithms. It has a bi-level opti- mization structure where the outer-loop process learns a shared initialization and the inner-loop process optimizes task-specific weights. Although MAML relies on the stan- dard gradient descent in the inner-loop, recent studies have shown that controlling the inner-loop’s gradient descent with a meta-learned preconditioner can be beneficial. Existing preconditioners, however, cannot simultaneously adapt in a task-specific and path-dependent way. Additionally, they do not satisfy the Riemannian metric condition, which can enable the steepest descent learning with preconditioned gradient. In this study, we propose Geometry-Adaptive Pre- conditioned gradient descent (GAP) that can overcome the limitations in MAML; GAP can efficiently meta-learn a preconditioner that is dependent on task-specific param- eters, and its preconditioner can be shown to be a Rieman- nian metric. Thanks to the two properties, the geometry- adaptive preconditioner is effective for improving the inner- loop optimization. Experiment results show that GAP out- performs the state-of-the-art MAML family and precondi- tioned gradient descent-MAML (PGD-MAML) family in a variety of few-shot learning tasks. Code is available at: https://github.com/Suhyun777/CVPR23-GAP .	1. Introduction Meta-learning, or learning to learn , enables algorithms to quickly learn new concepts with only a small number of samples by extracting prior-knowledge known as meta- knowledge from a variety of tasks and by improving the gen- eralization capability over the new tasks. Among the meta- learning algorithms, the category of optimization-based meta-learning [8, 17, 20, 21, 48] has been gaining popularity *Equal contribution. †Corresponding author.due to its flexible applicability over diverse fields including robotics [55, 59], medical image analysis [40, 54], language modeling [37,42], and object detection [46,61]. In particular, Model-Agnostic Meta-Learning (MAML) [20] is one of the most prevalent gradient-based meta-learning algorithms. Many recent studies have improved MAML by adopting a Preconditioned Gradient Descent (PGD1) for inner-loop optimization [34, 36, 44, 49, 53, 57, 66]. In this paper, we collectively address PGD-based MAML algorithms as the PGD-MAML family. A PGD is different from the ordinary gradient descent because it performs a preconditioning on the gradient using a preconditioning matrix P, also called a preconditioner . A PGD-MAML algorithm meta-learns not only the initialization parameter θ0of the network but also the meta-parameter ϕof the preconditioner P. For the inner-loop optimization, Pwas kept static in most of the previous works (Figure 1(b)) [34, 36, 44, 57, 66]. Some of the previous works considered adapting the preconditioner Pwith the inner-step k(Figure 1(c)) [49] and some others with the individual task (Figure 1(d)) [53]. They achieved performance improvement by considering individual tasks and inner-step, respectively, and showed that both factors were valuable. However, both factors have not been considered simultaneously in the existing studies. When a parameter space has a certain underlying struc- ture, there exists a Riemannian metric corresponding the pa- rameter space [3, 4]. If the preconditioning matrix is the Rie- mannian metric, the preconditioned gradient is known to be- come the steepest descent on the parameter space [2 –4,6,27]. An illustration of a toy example is shown in Figure 2. The optimization path of an ordinary gradient descent is shown in Figure 2(a). Compared to the ordinary gradient descent, a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can actually harm the optimization. For the example in Figure 2(b), the preconditioner affects the optimization into an undesirable direction and negatively affects the gradient descent. On the 1PGD in our work should not be confused with Projected Gradient Descent [13]. 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. 16080 (a) MAML (b) Non-adaptive P(ϕ) (c) Adaptive P(k;ϕ) (d) Adaptive P(Dtr τ;ϕ) (e) Adaptive P(θτ,k;ϕ) Figure 1. Diagram of MAML and PGD-MAML family. For the inner-loop adaptation in each diagram, the dotted lines of the same color indicate that they use a common preconditioning matrix (preconditioner). (a) MAML adaptation: no preconditioner is used ( i.e.,P=I). (b)P(ϕ): a constant preconditioner is used in the inner-loop where the preconditioner’s meta-parameter ϕis meta-learned. (c) P(k;ϕ): a constant preconditioner is used for each inner-step k. Preconditioner for each step is meta-learned, but P(k, ϕ)is not task-specific. (d) P(Dtr τ;ϕ): a constant preconditioner is used for each task. Preconditioner for each task is meta-learned, but P(Dtr τ;ϕ)is not dependent on k. (e) GAP adapts P(θτ,k;ϕ): a fully adaptive preconditioner is used where it is task-specific andpath-dependent . Instead of saying ‘dependent onk’, we specifically say it is path-dependent because the exact dependency is on the task-specific parameter set θτ,kthat is considerably more informative than k. (a) Gradient Descent (GD) (b) Preconditioned GD (Non-Riemannian metric) (c) Preconditioned GD (Riemannian metric) Figure 2. A toy example for illustrating the effect of Riemannian metric (see Supplementary Section A). When the curvature of the parameter space is poorly conditioned, (a) gradient descent can suf- fer from the difficulty of finding the solution, (b) a preconditioned gradient descent with a preconditioner that does not satisfy the Riemannian metric condition can suffer further, and (c) a precondi- tioned gradient descent with a preconditioner that is a Riemannian metric can perform better. contrary, if the preconditioner is the Riemannian metric cor- responding the parameter space, the preconditioned gradient descent can become the steepest descent and can exhibit a better optimization behavior as shown in Figure 2(c). While the Riemannian metric condition ( i.e., positive definiteness) is a necessary condition for steepest descent learning, the existing studies on PGD-MAML family did not consider constraining preconditioners to satisfy the condition for Rie- mannian metric. In this study, we propose a new PGD method named Geometry Aaptive Preconditioned gradient descent (GAP).Specifically, GAP satisfies two desirable properties which have not been considered before. First, GAP’s precondi- tioner PGAPis a fully adaptive preconditioner that can adapt to the individual task ( task-specific ) and to the optimization- path ( path-dependent ). The full adaptation is made possible by having the preconditioner depend on the task-specific parameter θτ,k(Figure 1(e)). Second, we prove that PGAP is a Riemannian metric. To this end, we force the meta- parameters of PGAPto be positive definite. Thus, GAP guar- antees the steepest descent learning on the parameter space corresponding to PGAP. Owing to the two properties, GAP enables a geometry-adaptive learning in inner-loop optimiza- tion. For the implementation of GAP, we utilize the Singular Value Decomposition (SVD) operation to come up with our preconditioner satisfying the desired properties. For the recently proposed large-scale architectures, computational overhead can be an important design factor and we provide a low-computational approximation, Approximate GAP , that can be proven to asymptotically approximate the operation of GAP. To demonstrate the effectiveness of GAP, we empiri- cally evaluate our algorithm on popular few-shot learning tasks; few-shot regression, few-shot classification, and few- shot cross-domain classification. The results show that GAP outperforms the state-of-the-art MAML family and PGD- MAML family. Themain contributions of our study can be summarized as follows: •We propose a new preconditioned gradient descent method called GAP, where it learns a preconditioner that enables a geometry-adaptive learning in the inner- loop optimization. •We prove that GAP’s preconditioner has two desir- able properties: (1) It is both task-specific and path- dependent ( i.e., dependent on task-specific parameter θτ,k). (2) It is a Riemannian metric. 16081 •For large-scale architectures, we provide a low- computational approximation that can be theoretically shown to approximate the GAP method. •For popular few-shot learning benchmark tasks, we empirically show that GAP outperforms the state-of- the-art MAML family and PGD-MAML family.
Kang_Scaling_Up_GANs_for_Text-to-Image_Synthesis_CVPR_2023	Abstract The recent success of text-to-image synthesis has taken the world by storm and captured the general public’s imag- ination. From a technical standpoint, it also marked a dras- tic change in the favored architecture to design generative image models. GANs used to be the de facto choice, with techniques like StyleGAN. With DALL ·E 2, autoregressive and diffusion models became the new standard for large- scale generative models overnight. This rapid shift raises a fundamental question: can we scale up GANs to benefit from large datasets like LAION? We find that na ¨ıvely in- creasing the capacity of the StyleGAN architecture quickly becomes unstable. We introduce GigaGAN, a new GAN ar- chitecture that far exceeds this limit, demonstrating GANs as a viable option for text-to-image synthesis. GigaGAN offers three major advantages. First, it is orders of mag- nitude faster at inference time, taking only 0.13 seconds to synthesize a 512px image. Second, it can synthesize high-resolution images, for example, 16-megapixel images in 3.66 seconds. Finally, GigaGAN supports various latent space editing applications such as latent interpolation, style mixing, and vector arithmetic operations.	1. Introduction Recently released models, such as DALL ·E 2 [53], Ima- gen [59], Parti [73], and Stable Diffusion [58], have ushered in a new era of image generation, achieving unprecedented levels of image quality and model flexibility. The now- dominant paradigms, diffusion models and autoregressive models, both rely on iterative inference. This is a double- edged sword, as iterative methods enable stable training with simple objectives but incur a high computational cost during inference. Contrast this with Generative Adversarial Networks (GANs) [5,17,33,51], which generate images through a sin- gle forward pass and thus inherently efficient. While such models dominated the previous “era” of generative mod- eling, scaling them requires careful tuning of the networkarchitectures and training considerations due to instabilities in the training procedure. As such, GANs have excelled at modeling single or multiple object classes, but scaling to complex datasets, much less an open world, has remained challenging. As a result, ultra-large models, data, and com- pute resources are now dedicated to diffusion and autore- gressive models. In this work, we ask – can GANs continue to be scaled up and potentially benefit from such resources, or have they plateaued? What prevents them from further scaling, and can we overcome these barriers? We first experiment with StyleGAN2 [34] and observe that simply scaling the backbone causes unstable training. We identify several key issues and propose techniques to stabilize the training while increasing the model capacity. First, we effectively scale the generator’s capacity by re- taining a bank of filters and taking a sample-specific linear combination. We also adapt several techniques commonly used in the diffusion context and confirm that they bring similar benefits to GANs. For instance, interleaving both self-attention (image-only) and cross-attention (image-text) with the convolutional layers improves performance. Furthermore, we reintroduce multi-scale training, find- ing a new scheme that improves image-text alignment and low-frequency details of generated outputs. Multi-scale training allows the GAN-based generator to use parameters in low-resolution blocks more effectively, leading to better image-text alignment and image quality. After careful tun- ing, we achieve stable and scalable training of a one-billion- parameter GAN (GigaGAN) on large-scale datasets, such as LAION2B-en [63]. Our results are shown in Figure 1. In addition, our method uses a multi-stage approach [12, 76]. We first generate at 64×64and then upsample to 512×512. These two networks are modular and robust enough to be used in a plug-and-play fashion. We show that our text-conditioned GAN-based upsampling network can be used as an efficient, higher-quality upsampler for a base diffusion model such as DALL ·E 2, despite never having seen diffusion images at training time (Figures 1). Together, these advances enable our GigaGAN to go far beyond previous GANs: 36 ×larger than Style- GAN2 [34] and 6 ×larger than StyleGAN-XL [62] 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. 10124 A portrait of a human growing colorful flowers from her hair. Hyperrealistic oil painting. Intricate details. a blue Porsche 356 parked in front of a yellow brick wall.Eiffel Tower, landscape photographyA painting of a majestic royal tall ship in Age of Discovery. A golden luxury motorcycle parked at the King's palace. 35mm f/4.5.A hot air balloon in shape of a heart. Grand Canyonlow poly bunny with cute eyes A cube made of denim on a wooden table GigaGAN (4K) Input Real-ESRGAN SD Upscaler GigaGAN (1K) GigaGAN Upscaler (4096px, 16Mpix, 3.66s)Input artwork ( 128px) Real-ESRGAN (1024px, 0.06s) SD Upscaler (1024px, 7.75s) GigaGAN Upscaler (1024px, 0.13s) Figure 1. Our model, GigaGAN, shows GAN frameworks can also be scaled up for general text-to-image synthesis and super- resolution tasks, generating a 512px output at an interactive speed of 0.13s, and 4096px within 3.7s. Selected examples at 2K resolution (text-to-image synthesis) and 1k or 4k resolutions (super-resolution) are shown. For the super-resolution task, we use the caption of “Portrait of a colored iguana dressed in a hoodie.” and compare our model with the text-conditioned upscaler of Stable Diffusion [57] and unconditional Real-ESRGAN [26]. Please zoom in for more details. See our arXiv paper and website for more uncurated comparisons. 10125 XMC-GAN [75]. While our 1B parameter count is still lower than the largest synthesis models, such as DALL ·E 2 (5.5B), and Parti (20B), we have not yet observed a qual- ity saturation regarding the model size. GigaGAN achieves a zero-shot FID of 9.09 on COCO2014, lower than the FID of DALL ·E 2, Parti-750M, and Stable Diffusion. Furthermore, GigaGAN has three major practical ad- vantages compared to diffusion and autoregressive models. First, it is orders of magnitude faster, generating a 512px image in 0.13 seconds. Second, it can synthesize ultra high- res images at 4k resolution in 3.66 seconds. Third, it is endowed with a controllable, latent vector space that lends itself to well-studied controllable image synthesis applica- tions, such as prompt mixing (Figure 4), style mixing (Fig- ure 5), and prompt interpolation (Figures A7 and A8). In summary, our model is the first GAN-based method that successfully trains a billion-scale model on billions of real-world complex Internet images. This suggests that GANs are still a viable option for text-to-image synthe- sis and should be considered for future aggressive scaling. Please visit our website for additional results.
Kamath_A_New_Path_Scaling_Vision-and-Language_Navigation_With_Synthetic_Instructions_and_CVPR_2023	Abstract Recent studies in Vision-and-Language Navigation (VLN) train RL agents to execute natural-language navi- gation instructions in photorealistic environments, as a step towards robots that can follow human instructions. How- ever, given the scarcity of human instruction data and lim- ited diversity in the training environments, these agents still struggle with complex language grounding and spa- tial language understanding. Pretraining on large text and image-text datasets from the web has been extensively explored but the improvements are limited. We investi- gate large-scale augmentation with synthetic instructions. We take 500+ indoor environments captured in densely- sampled 360panoramas, construct navigation trajecto- ries through these panoramas, and generate a visually- grounded instruction for each trajectory using Marky [63], a high-quality multilingual navigation instruction genera- tor. We also synthesize image observations from novel view- points using an image-to-image GAN [27]. The resulting dataset of 4.2M instruction-trajectory pairs is two orders of magnitude larger than existing human-annotated datasets, and contains a wider variety of environments and view- points. To efﬁciently leverage data at this scale, we train a simple transformer agent with imitation learning. On the challenging RxR dataset, our approach outperforms all ex- isting RL agents, improving the state-of-the-art NDTW from 71.1 to 79.1 in seen environments, and from 64.6 to 66.8 in unseen test environments. Our work points to a new path to improving instruction-following agents, emphasizing large- scale training on near-human quality synthetic instructions.	1. Introduction Developing intelligent agents that follow human instruc- tions is a long-term, formidable challenge in AI [66]. A recent focus addressing this problem space is Vision-and- Language Navigation (VLN) [3, 9]. Navigation is an ideal Equal Contribution.yWork done while at Google Research. Figure 1. Simpler agents with more data: We investigate large- scale augmentation using 500+ environments annotated with syn- thetic instructions that approach human quality. test bed for studying instruction-following, since the task can be simulated photo-realistically at scale and evaluation is straightforward. However, datasets that capture the lin- guistic diversity and idiosyncrasies of real human instruc- tors are small and expensive to collect. Shortages of human-annotated training data for other vision-and-language tasks have been partially addressed by pretraining transformers on up to billions of image- text pairs. This has underpinned dramatic improvements in image captioning [65, 70], visual question answering [59], phrase grounding [26, 35], text-to-video retrieval, video question answering [32] and text-to-image synthesis [49, 69]. However, these are all static image or video tasks, whereas VLN agents interact with 3D environments. In VLN, pretraining on large image-text and text-only datasets has been thoroughly explored [21, 22, 40, 45], but improve- ments are more limited. Arguably, progress in VLN has plateaued while still leaving a large gap between machine and human performance [73]. We hypothesize that static image-text and text-only datasets – despite their size – lack the spatially grounded and action-oriented language needed for effective VLN pretraining. Consider instructions from the Room-across-Room (RxR) dataset [30], which illus- trate that wayﬁnding requires an understanding of allocen- 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. 10813 tric and egocentric spatial expressions ( near a grey console table behind you ), verbs ( climb the stairs ), imperatives and negations ( do not enter the room in front ) and temporal conditions ( walk until you see an entrance on your left ). Such expressions are rarely found in image-text datasets. Though similar expressions are found in text-only corpora, their meaning as it relates to the physical world is hard to infer from text alone (without sensorimotor context) [6]. To address this problem, we investigate large-scale augmentation with synthetic in-domain data, i.e., model- generated navigation instructions for trajectories in realistic 3D environments using previously developed components [27, 63]. We construct a large dataset using Marky [63], which generates VLN instructions that approach the qual- ity of human instructors. [63] released the 1M Marky instruction-trajectory pairs situated in 61 Matterport3D [7] environments. To increase the diversity of the environments (and thus the scenes and objects available in them), we au- tomatically annotate an additional 491 environments from the Gibson dataset [67]. Gibson environments have been underutilized in prior VLN work due to the lack of navi- gation graphs indicating navigable trajectories through its densely-sampled 360panoramas. We train a model that classiﬁes navigable directions for Matterport3D and use it to construct the missing navigation graphs. We sample 3.2M trajectories from these graphs and annotate them with Marky. To further increase the variability of trajectories, we synthesize image observations from novel viewpoints using an image-to-image GAN [27]. The resulting dataset is two orders of magnitude larger than existing human-annotated ones, and contains a wider variety of scenes and viewpoints. We have released our Gibson navigation graphs and the Marky-Gibson dataset.2 With orders of magnitude more training examples and environments, we explore VLN agent performance with imitation learning (IL), i.e., behavioral cloning and DAG- GER [53] IL can take advantage of high-throughput trans- former frameworks such as T5 [48] and thus efﬁciently train on 4.2M instructions (accumulating over 700M steps of ex- perience). This is a departure from most prior VLN work in low-data settings, e.g. [10] report that pure IL under- performs by 8.5% success rate compared to agents trained with both IL and online reinforcement learning (RL) algo- rithms such as A3C [44]. However, IL outperforms RL in related tasks with sufﬁcient training data [50]. Online RL also requires interacting with the environment at each step; this precludes efﬁcient data prefetching and paral- lel preprocessing and thus imposes unavoidable overhead compared to IL. Empirically, we conﬁrm that training ex- isting models such as HAMT [10] on 4.2M instructions is infeasible without ground-up re-engineering, though we do ﬁnd incorporating 10K additional synthetic instructions into 2//github.com/google-research-datasets/RxR/tree/main/marky-mT5HAMT training modestly improves performance. Training with IL aligns with the trend towards large-scale multi-task vision-and-language models trained with supervised learn- ing; these have uniﬁed tasks as diverse as visual question answering, image captioning, object detection, image clas- siﬁcation, OCR and text reasoning [12] – and could include VLN in future. Experimentally, in detailed ablation studies we show that adding Gibson environments, synthesizing additional image observations from novel viewpoints, increasing the capacity of the transformer, and ﬁnetuning with DAGGER all im- prove agent performance. On the challenging RxR dataset – which contains multilingual instructions with a median trajectory length of 15m – our best agent using only imita- tion learning outperforms all prior RL agents. Evaluating on novel instruction-trajectories in seen environments (Val- Seen), we improve over the state-of-the-art by 8%, reaching 79.1 NDTW. In new, unseen environments (Test), we im- prove by 2%, achieving 66.8 NDTW. We also show that that self-training with synthetic instructions in new envi- ronments (still without human annotations) improves per- formance by an additional 2% to 68.6 NDTW. Overall, our RxR results point to a new path to improving instruction- following agents, emphasizing large-scale training on near- human quality synthetic instructions. Perhaps surprisingly, on the English-only R2R dataset [3], our IL agent achieves strong but not state-of-the-art results. Marky was trained on RxR, so we attribute this to domain differences between R2R and RxR, underscoring the domain dependence of syn- thetic instructions.
Jung_Devils_on_the_Edges_Selective_Quad_Attention_for_Scene_Graph_CVPR_2023	Abstract Scene graph generation aims to construct a semantic graph structure from an image such that its nodes and edgesrespectively represent objects and their relationships. Oneof the major challenges for the task lies in the presenceof distracting objects and relationships in images; contex-tual reasoning is strongly distracted by irrelevant objects or backgrounds and, more importantly, a vast number of irrel- evant candidate relations. To tackle the issue, we proposethe Selective Quad Attention Network (SQUAT) that learnsto select relevant object pairs and disambiguate them via di-verse contextual interactions. SQUAT consists of two maincomponents: edge selection and quad attention. The edge selection module selects relevant object pairs, i.e., edges inthe scene graph, which helps contextual reasoning, and thequad attention module then updates the edge features us- ing both edge-to-node and edge-to-edge cross-attentions tocapture contextual information between objects and object pairs. Experiments demonstrate the strong performanceand robustness of SQUAT, achieving the state of the art on the Visual Genome and Open Images v6 benchmarks.	1. Introduction The task of scene graph generation (SGG) is to construct a visually-grounded graph from an image such that its nodesand edges respectively represent objects and their relation-ships in the image [ 27,43,46]. The scene graph provides a semantic structure of images beyond individual objects and thus is useful for a wide range of vision problems such asvisual question answering [ 36,37], image captioning [ 53], image retrieval [ 11], and conditional image generation [ 10], where a holistic understanding of the relationships amongobjects is required for high-level reasoning. With recent ad- vances in deep neural networks for visual recognition, SGGhas been actively investigated in the computer vision com- munity. A vast majority of existing methods tackle SGG byﬁrst detecting candidate objects and then performing con-textual reasoning between the objects via message pass-(a) Ground-truth scene graph manhorse glasses bucketBackground Background (b) Fully-connected scene graph horse bucketBackground Backgroundglassesman Node EdgeNode EdgeQueryKey-Value N2N E2NN2E E2E Node Edge (c) Overview of the quad attention module Figure 1. (a) The ground-truth scene graph contains only 4 ground-truth objects and 4 relations between the objects. (b) Only13% of edges in a fully-connected graph have the actual relation-ships according to the ground-truths. (c) The overview of the quadattention. The node features are updated by node-to-node (N2N)and node-to-edge (N2E) attentions, and the edge features are up-dated by edge-to-node (E2N) and edge-to-edge (E2E) attentions. ing [19,21,43] or sequential modeling [ 28,36,50]. Despite these efforts, the task of SGG remains extremely challeng-ing, and even the state-of-the-art methods do not producereliable results for practical usage. While there exist a multitude of challenges for SGG, the intrinsic difﬁculty may lie in the presence of distracting ob-jects and relationships in images; there is a vast number ofpotential but irrelevant relations, i.e., edges, which quadrat- ically increase with the number of candidate objects, i.e., nodes, in the scene graph. The contextual reasoning for SGG in the wild is thus largely distracted by irrelevant ob- jects and their relationship pairs. Let us take a simple ex-ample as in Fig. 1, where 4 objects and 4 relations in its ground-truth scene graph exist in the given image. If our 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. 18664 object detector obtains 6 candidate boxes, 2 of which are from the background (red), then the contextual reasoning,e.g., message passing or self-attention, needs to consider 30 potential relations, 87% of which are not directly related ac-cording to the ground-truth and most of them may thus actas distracting outliers. In practice, the situation is far worse;in the Visual Genome dataset, the standard benchmark forSGG, an image contains 38 objects and 22 relationships onaverage [ 43], which means that only around 1% of object pairs have direct and meaningful relations even when ob-ject detection is perfect. As will be discussed in our exper- iments, we ﬁnd that existing contextual reasoning schemesobtain only a marginal gain at best and often degrade theperformance. The crux of the matter for SGG may lie indeveloping a robust model for contextual reasoning againstirrelevant objects and relations. To tackle the issue, we propose the Selective Quad Atten- tion Network (SQUAT) that learns to select relevant object pairs and disambiguate them via diverse contextual interac-tions with objects and object pairs. The proposed methodconsists of two main components: edge selection and quadattention. The edge selection module removes irrelevant ob-ject pairs, which may distract contextual reasoning, by pre-dicting the relevance score for each pair. The quad attentionmodule then updates the edge features using edge-to-nodeand edge-to-edge cross-attentions as well as the node fea-tures using node-to-node and node-to-edge cross-attentions;it thus captures contextual information between all objectsand object pairs, as shown in Figure 1(c). Compared to pre- vious methods [ 19,21], which perform either node-to-node or node-to-edge interactions, our quad attention providesmore effective contextual reasoning by capturing diverse interactions in the scene graph. For example, in the caseof Fig. 1(a), [‘man’, ‘feeding’, ‘horse’] relates to [‘man’, ‘holding’, ‘bracket’] and [‘horse’, ‘eat from’, ‘bracket’],and vice versa; node-to-node or node-to-edge interactionsare limited in capturing such relations between the edges. Our contributions can be summarized as follows: • We introduce the edge selection module for SGG that learns to select relevant edges for contextual reasoning. • We propose the quad attention module for SGG that performs effective contextual reasoning by updatingnode and edge features via diverse interactions. • The proposed SGG model, SQUAT, outperforms the state-of-the-art methods on both Visual Genome and Open Images v6 benchmarks. In particular, SQUATachieves remarkable improvement on the SGDet set-tings, which is the most realistic and challenging.
Jung_AnyFlow_Arbitrary_Scale_Optical_Flow_With_Implicit_Neural_Representation_CVPR_2023	Abstract To apply optical flow in practice, it is often necessary to resize the input to smaller dimensions in order to reduce computational costs. However, downsizing inputs makes the estimation more challenging because objects and motion ranges become smaller. Even though recent approaches have demonstrated high-quality flow estimation, they tend to fail to accurately model small objects and precise bound- aries when the input resolution is lowered, restricting their applicability to high-resolution inputs. In this paper, we introduce AnyFlow, a robust network that estimates accu- rate flow from images of various resolutions. By repre- senting optical flow as a continuous coordinate-based rep- resentation, AnyFlow generates outputs at arbitrary scales from low-resolution inputs, demonstrating superior perfor- mance over prior works in capturing tiny objects with de- tail preservation on a wide range of scenes. We establish a new state-of-the-art performance of cross-dataset gener- alization on the KITTI dataset, while achieving comparable accuracy on the online benchmarks to other SOTA methods.	1. Introduction Optical flow seeks to estimate per-pixel correspon- dences, characterized as the horizontal and vertical shift, from a pair of images. Specifically, it aims to identify the correspondence across pixels in different images, which is at the heart of numerous computer vision tasks such as video denoising [4, 33, 70], action recognition [55, 60, 62] and object tracking [11, 26, 46]. This is particularly chal- lenging due to the fact that the scene geometry and object motion are combined into a single observation, making the inverse problem of estimating motion highly ill-posed. A common assumption for enabling computationally tractable optical flow is to explicitly account for small mo- tion in local neighborhood and incorporate additional prior to constrain the solution space, such as using total variation *This work was done during the author’s internship at Meta. †Affiliated with Meta at the time of this work. Figure 1. Predicting 50% downsized images on Sintel [5], we compare AnyFlow to RAFT [57]. AnyFlow shows clearer bound- aries, accurate shapes, and better detection of small objects. prior [15, 71] and smooth prior [2, 3, 51]. While effective, these techniques are significantly restricted by the assump- tion and do not generalize well for real-life scenes with so- phisticated geometry and motions. An alternative approach is motivated by the success of deep neural networks. Different from using handcrafted priors, learning-based methods [14, 67, 74] design end-to- end networks to regress the motion field. Sun et al. [53] use the coarse-to-fine approach to characterize the pixel-to- pixel mapping and several methods [20, 22, 44, 67] are de- veloped along this line. The drawback of these approaches is that the accuracy of flow estimation is not only limited by the sampling but also has a critical dependence on a good initial solution because the pixel correspondence is funda- mentally ambiguous. Tedd and Deng [57] resort to iterative update module on top of the correlation volumes, allowing better estimates in tiny and fast-moving objects. Indeed the success of this model relies on having a correlation volume that is high quality in characterizing the feature correspon- dence for sub-pixel level. Inspired by its success, follow- up methods [23, 25, 36, 50, 52] have further improved it in multiple ways. One popular direction is to introduce atten- 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. 5455 tion mechanisms [59] for efficiency [64,72], occlusion han- dling [24], and large displacements [49]. As attention has been found to be effective in estimating long-range depen- dencies, Xu et al. [65] and Huang et al. [19] adopt Vision Transformers [34] to perform global matching and demon- strate their effectiveness. However, these methods naturally become less effective for low-resolution images, where in- accuracy is introduced when computing the correlation vol- umes and the iterative refinements. They tend to fail in ac- curately modeling small objects and precise boundary when the input resolution is lowered. This inherently limits the applicability for mobile devices in particular, where resiz- ing the input to small size is often necessary to reduce the computational cost. In this paper, we propose AnyFlow, a novel method that is agnostic to the resolution of images. The key insight behind AnyFlow is the use of implicit neural representa- tion (INR) for flow estimation. INRs, such as LIIF [8], have been demonstrated to be effective for image super- resolution, as they model an image as a continuous function without limiting the output to a fixed resolution. This en- ables image generation at arbitrary scales. Inspired by this capability, we design a continuous coordinate-based flow upsampler to infer flow at any desired scale. In addition, we propose a novel warping scheme utilizing multi-scale feature maps together with dynamic correlation lookup to generalize well on diverse shapes of input. As the proposed methods produce a synergistic effect, we demonstrate su- perior performance for tiny objects and detail preservation on a wide range of scenes involving complex geometry and motions. Specifically, we verify the robustness for the input with low resolution. Fig. 1 showcases our technique against RAFT [57]. Our contributions are summarized as follows: • We introduce AnyFlow, a novel network to produce high quality flow estimation for arbitrary size of image. • We present the methods to utilize multi-scale feature maps and search optimal correspondence within pre- dicted range, enabling robust estimation in wide ranges of motion types. • We demonstrate strong performance on cross-dataset generalization by achieving more than 25% of error reduction with only 0.1M additional parameters from the baseline and rank 1st on the KITTI dataset. Together, our contributions provide, for the first time, an approach for arbitrary size of input, which significantly im- proves the quality for low-resolution images and extends the applicability for portable devices.
Kalb_Principles_of_Forgetting_in_Domain-Incremental_Semantic_Segmentation_in_Adverse_Weather_CVPR_2023	Abstract Deep neural networks for scene perception in automated vehicles achieve excellent results for the domains they were trained on. However, in real-world conditions, the do- main of operation and its underlying data distribution are subject to change. Adverse weather conditions, in partic- ular, can significantly decrease model performance when such data are not available during training. Addition- ally, when a model is incrementally adapted to a new do- main, it suffers from catastrophic forgetting, causing a sig- nificant drop in performance on previously observed do- mains. Despite recent progress in reducing catastrophic forgetting, its causes and effects remain obscure. Therefore, we study how the representations of semantic segmenta- tion models are affected during domain-incremental learn- ing in adverse weather conditions. Our experiments and representational analyses indicate that catastrophic forget- ting is primarily caused by changes to low-level features in domain-incremental learning and that learning more gen- eral features on the source domain using pre-training and image augmentations leads to efficient feature reuse in sub- sequent tasks, which drastically reduces catastrophic for- getting. These findings highlight the importance of meth- ods that facilitate generalized features for effective contin- ual learning algorithms.	1. Introduction Semantic segmentation is widely used for environment perception in automated driving, where it aims at recogniz- ing and comprehending images at the pixel level. One fun- damental constraint of the traditional deep learning-based semantic segmentation models is that they are often only trained and evaluated on data collected mostly in clear weather conditions and that they assume that the domain of the training data matches the domain they operate in. How- ever, in the real world, those autonomous driving systems Figure 1. Activation drift between models f1tof0measured by relative mIoU on the first task of the models stitched together at specific layers (horizontal axis). The layers of the encoder are marked in the gray area, the decoder layers in the white area. Layer-stitching reveals that during domain-incremental learning, changes in low-level features are a major cause of forgetting. With an improved training scheme, combining simple augmentations, exchanging normalization layers and using pre-training, the model is optimized to reuse low-level features during incremental learn- ing, leading to significant reduction of catastrophic forgetting. are faced with constantly changing driving environments and variable input data distributions. Specifically, changing weather conditions can have adverse effects on the perfor- mance of segmentation models. Therefore, a semantic segmentation model needs to be adapted to these conditions. A naive solution to this prob- lem would be to incrementally fine-tune the model to new domains with labeled data. However, fine-tuning a neural network to a novel domain will, in most cases, lead to a severe performance drop in previously observed domains. This phenomenon is usually referred to as catastrophic for- getting and is a fundamental challenge when training a neu- ral network on a continuous stream of data. Recently pro- posed methods mostly mitigate this challenge by replaying data from previous domains, re-estimating statistics or even in an unsupervised manner by transferring training images in the style of the novel domain [32, 48, 52]. The focus of our work is to study how the internal representations of se- 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. 19508 mantic segmentation models are affected during domain- incremental learning and how efficient feature reuse can mitigate forgetting without explicit replay of the previous domain. Our main contributions are: 1. We analyze the activation drift that a model’s layers are subjected to when adapting from good to adverse weather conditions by stitching them with the previous task’s network. We reveal that the major cause of for- getting is a shift of low-level representations in the first convolution layer that adversely affects the population statistics of the following BatchNorm Layer. 2. Using different augmentation strategies to match the target domains in color statistics or in the frequency domain, we reveal that learning color-invariant fea- tures stabilizes the representations in early layers, as they don’t change when the model is adapted to a new domain. 3. With a combination of pre-training, augmentations and exchanged normalization layers, we achieve an over- all reduction of forgetting of ∼20% mIoU compared to fine-tuning without using any form of replay and prove the effectiveness of pre-training and augmenta- tions which are often overlooked in continual learning.
Kalischek_BiasBed_-_Rigorous_Texture_Bias_Evaluation_CVPR_2023	Abstract The well-documented presence of texture bias in modern convolutional neural networks has led to a plethora of al- gorithms that promote an emphasis on shape cues, often to support generalization to new domains. Yet, common datasets, benchmarks and general model selection strate- gies are missing, and there is no agreed, rigorous evaluation protocol. In this paper, we investigate difficulties and limi- tations when training networks with reduced texture bias. In particular, we also show that proper evaluation and mean- ingful comparisons between methods are not trivial. We introduce BiasBed , a testbed for texture- and style-biased training, including multiple datasets and a range of exist- ing algorithms. It comes with an extensive evaluation pro- tocol that includes rigorous hypothesis testing to gauge the significance of the results, despite the considerable train- ing instability of some style bias methods. Our extensive experiments, shed new light on the need for careful, sta- tistically founded evaluation protocols for style bias (and beyond). E.g., we find that some algorithms proposed in the literature do not significantly mitigate the impact of style bias at all. With the release of BiasBed , we hope to fos- ter a common understanding of consistent and meaning- ful comparisons, and consequently faster progress towards learning methods free of texture bias. Code is available at https://github.com/D1noFuzi/BiasBed	1. Introduction Visual object recognition is fundamental to our daily lives. Identifying and categorizing objects in the environ- ment is essential for human survival, indeed our brain is able to assign the correct object class within a fraction of a second, independent of substantial variations in appearance, illumination and occlusion [28]. Recognition mainly takes place along the ventral pathway [7], i.e., visual perception induces a hierarchical processing stream from local patterns to complex features, in feed-forward fashion [28]. Signalsare filtered to low frequency and used in parallel also in a top-down manner [2], emphasizing the robustness and in- variance to deviations in appearance. Inspired by our brain’s visual perception, convolutional neural architectures build upon the hierarchical intuition, stacking multiple convolutional layers to induce feed- forward learning of basic concepts to compositions of com- plex objects. Indeed, early findings suggested that neu- rons in deeper layers are activated by increasingly complex shapes, while the first layers are mainly tailored towards low-level features such as color, texture and basic geometry. However, recent work indicates the opposite: convolutional neural networks (CNNs) exhibit a strong bias to base their decision on texture cues [11–13, 17], which heavily influ- ences their performance, in particular under domain shifts, which typically affect local texture more than global shape. This inherent texture bias has led to a considerable body of work that tries to minimize texture and style bias and shift towards “human-like” shape bias in object recognition. Common to all approaches is the principle of incorporating adversarial texture cues into the learning pipeline – either directly in input space [13, 15, 27] or implicitly in feature space [20,22,30,33]. Perturbing the texture cues in the input forces the neural network to make “texture-free” decisions, thus focusing on the objects’ shapes that remain stable dur- ing training. While texture cues certainly boost perfor- mance in fine-grained classification tasks, where local tex- ture patterns may indicate different classes, they can cause serious harm when the test data exhibit a domain shift w.r.t. training. In this light, texture bias can be seen as a main rea- son for degrading domain generalization performance, and various algorithms have been developed to improve general- ization to new domains with different texture properties (re- spectively image styles), e.g., [13, 15, 20, 22, 27, 30, 33, 40]. However, while a considerable number of algorithms have been proposed to address texture bias, they are neither evaluated on common datasets nor with common evaluation metrics. Moreover they often employ inconsistent model selection criteria or do not report at all how model selection 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. 22221 is done. With the present paper we promote the view that: • the large domain shifts induced by texture-biased train- ing cause large fluctuations in accuracy, which call for a particularly rigorous evaluation; • model selection has so far been ignored in the litera- ture; together with the high training volatility, this may have lead to overly optimistic conclusions from spuri- ous results; • in light of the difficulties of operating under drastic do- main shifts, experimental results should include a no- tion of uncertainty in the evaluation metrics. Motivated by these findings, we have implemented Bias- Bed, an open-source PyTorch [35] testbed that comes with six datasets of different texture and shape biases, four ad- versarial robustness datasets and eight fully implemented algorithms. Our framework allows the addition of new al- gorithms and datasets with few lines of code, including full flexibility of all parameter settings. In order to tackle the previously discussed limitations and difficulties for evalu- ating such algorithms, we have added a carefully designed evaluation pipeline that includes training multiple runs and employing multiple model selection methods, and we re- port all results based on sound statistical hypothesis tests – all run with a single command. In addition to our novel framework, the present paper makes the following contribu- tions: • We highlight shortcomings in the evaluation protocols used in recent work on style bias, including the ob- servation that there is a very high variance in the per- formance of different runs of the same algorithm, and even between different checkpoints in the same run with similar validation scores. • We develop and openly release a testbed that rig- orously compares different algorithms using well- established hypothesis testing methods. This testbed includes several of the most prominent algorithms and datasets in the field, and is easily extensible. • We observe in our results that current algorithms on texture-bias datasets fail to surpass simple ERM in a statistically significant way, which is the main moti- vation for this work and for using rigorous hypothesis tests for evaluating style bias algorithms. In Sec. 2, we provide a comprehensive overview of exist- ing work on reducing texture bias. Furthermore we describe the main forms of statistical hypothesis testing. We con- tinue in Sec. 3 with a formal introduction to biased learning, and systematically group existing algorithms into families with common properties. In Sec. 4, we investigate previous evaluation practices in detail, discuss their limitations, andgive a formal introduction to hypothesis testing. Sec. 5 de- scribes our proposed BiasBed evaluation framework in de- tail, while Sec. 6 presents experiments, followed by a dis- cussion (Sec. 7), conclusions (Sec. 8) and a view towards the broader impact of our work (Sec. 9).
Jung_On_the_Importance_of_Accurate_Geometry_Data_for_Dense_3D_CVPR_2023	Abstract Learning-based methods to solve dense 3D vision prob- lems typically train on 3D sensor data. The respectively used principle of measuring distances provides advantages and drawbacks. These are typically not compared nor discussed in the literature due to a lack of multi-modal datasets. Texture-less regions are problematic for structure from motion and stereo, reflective material poses issues for active sensing, and distances for translucent objects are in- tricate to measure with existing hardware. Training on in- accurate or corrupt data induces model bias and hampers generalisation capabilities. These effects remain unnoticed if the sensor measurement is considered as ground truth during the evaluation. This paper investigates the effect of sensor errors for the dense 3D vision tasks of depth estima- tion and reconstruction. We rigorously show the significant impact of sensor characteristics on the learned predictions and notice generalisation issues arising from various tech- nologies in everyday household environments. For evalu- ation, we introduce a carefully designed dataset1compris- ing measurements from commodity sensors, namely D-ToF , I-ToF , passive/active stereo, and monocular RGB+P . Our study quantifies the considerable sensor noise impact and paves the way to improved dense vision estimates and tar- geted data fusion.	1. Introduction Our world is 3D. Distance measurements are essential for machines to understand and interact with our environ- ment spatially. Autonomous vehicles [23, 30, 50, 58] need this information to drive safely, robot vision requires dis- tance information to manipulate objects [15,62,72,73], and AR realism benefits from spatial understanding [6, 31]. A variety of sensor modalities and depth predic- 1dataset available at https://github.com/Junggy/HAMMER-dataset Figure 1. Other datasets for dense 3D vision tasks reconstruct the scene as a whole in one pass [8,12,56], resulting in low quality and accuracy (cf. red boxes). On the contrary, our dataset scans the background and every object in the scene separately a priori and annotates them as dense and high-quality 3D meshes. Together with precise camera extrinsics from robotic forward-kinematics, this enables a fully dense rendered depth as accurate pixel-wise ground truth with multimodal sensor data, such as RGB with po- larization, D-ToF, I-ToF and Active Stereo. Hence, it allows quan- tifying different downstream 3D vision tasks such as monocular depth estimation, novel view synthesis, or 6D object pose estima- tion. tion pipelines exist. The computer vision community thereby benefits from a wide diversity of publicly available datasets [23, 51, 52, 57, 60, 61, 65], which allow for evalua- tion of depth estimation pipelines. Depending on the setup, different sensors are chosen to provide ground truth (GT) depth maps, all of which have their respective advantages and drawbacks determined by their individual principle of distance reasoning. Pipelines are usually trained on the data without questioning the nature of the depth sensor used for supervision and do not reflect areas of high or low confi- dence of the GT. Popular passive sensor setups include multi-view stereo cameras where the known or calibrated spatial relationship 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. 780 between them is used for depth reasoning [51]. Correspond- ing image parts or patches are photometrically or struc- turally associated, and geometry allows to triangulate points within an overlapping field of view. Such photometric cues are not reliable in low-textured areas and with little ambient light where active sensing can be beneficial [52,57]. Active stereo can be used to artificially create texture cues in low- textured areas and photon-pulses with a given sampling rate are used in Time-of-Flight (ToF) setups either directly (D- ToF) or indirectly (I-ToF) [26]. With the speed of light, one can measure the distance of objects from the return time of the light pulse, but unwanted multi-reflection artifacts also arise. Reflective and translucent materials are measured at incorrect far distances, and multiple light bounces distort measurements in corners and edges. While ToF signals can still be aggregated for dense depth maps, a similar setup is used with LiDAR sensors which sparsely measure the dis- tance using coordinated rays that bounce from objects in the surrounding. The latter provides ground truth, for instance, for the popular outdoor driving benchmark KITTI [23]. While LiDAR sensing can be costly, radar [21] provides an even sparser but more affordable alternative. Multiple modalities can also be fused to enhance distance estimates. A common issue, however, is the inherent problem of warp- ing onto a common reference frame which requires the in- formation about depth itself [27, 37]. While multi-modal setups have been used to enhance further monocular depth estimation using self-supervision from stereo and temporal cues [25, 60], its performance analysis is mainly limited to average errors and restricted by the individual sensor used. An unconstrained analysis of depth in terms of RMSE com- pared against a GT sensor only shows part of the picture as different sensing modalities may suffer from drawbacks. Where are the drawbacks of current depth-sensing modalities - and how does this impact pipelines trained with this (potentially partly erroneous) data? Can self- or semi-supervision overcome some of the limitations posed currently? To objectively investigate these questions, we provide multi modal sensor data as well as highly accurate annotated depth so that one can analyse the deterioration of popular monocular depth estimation and 3D reconstruc- tion methods (see Fig. 1) on areas of different photometric complexity and with varying structural and material prop- erties while changing the sensor modality used for training. To quantify the impact of sensor characteristics, we build a unique camera rig comprising a set of the most popular in- door depth sensors and acquire synchronised captures with highly accurate ground truth data using 3D scanners and aligned renderings. To this end, our main contributions can be summarized as follows: 1. We question the measurement quality from commodity depth sensor modalities and analyse their impact as supervision signals for the dense 3D vision tasks ofdepth estimation and reconstruction. 2. We investigate performance on texture-varying mate- rial as well as photometrically challenging reflec- tive, translucent and transparent areas where learning methods systematically reproduce sensor errors . 3. To objectively assess and quantify different data sources, we contribute an indoor dataset compris- ing an unprecedented combination of multi-modal sensors , namely I-ToF, D-ToF, monocular RGB+P, monochrome stereo, and active light stereo together with highly accurate ground truth.
Kang_Benchmarking_Self-Supervised_Learning_on_Diverse_Pathology_Datasets_CVPR_2023	Abstract Computationalpathologycanleadtosavinghumanlives, butmodelsareannotationhungryandpathologyimagesare notoriouslyexpensivetoannotate. Self-supervisedlearning (SSL) has shown to be an effective method for utilizing un- labeleddata,anditsapplicationtopathologycouldgreatly benefit its downstream tasks. Yet, there are no principled studiesthatcompareSSLmethodsanddiscusshowtoadapt them for pathology. To address this need, we execute the largest-scale study of SSL pre-training on pathology image data,todate. Ourstudyisconductedusing4representative SSLmethodsondiversedownstreamtasks. Weestablishthat large-scale domain-aligned pre-training in pathology con- sistently out-performs ImageNet pre-training in standard SSL settings such as linear and fine-tuning evaluations, as well as in low-label regimes. Moreover, we propose a set of domain-specific techniques that we experimentally show leads to a performance boost. Lastly, for the first time, we applySSLtothechallengingtaskofnucleiinstancesegmen- tationandshowlargeandconsistentperformanceimprove- ments. We release the pre-trained model weights1.	1. Introduction Thecomputationalanalysisofmicroscopicimagesofhu- man tissue – also known as computational pathology – has emerged as an important topic of research, as its clinical implementations can result in the saving of human lives *The first two authors contributed equally. 1https://lunit-io.github.io/research/publications/pathology_sslby improving cancer diagnosis [49] and treatment [42]. Deep Learning and Computer Vision methods in pathol- ogyallowforobjectivity[15],large-scaleanalysis[20],and triaging [5] but often require large amounts of annotated data [52]. However, the annotation of pathology images re- quiresspecialistswithmanyyearsofclinicalresidency[37], resulting in scarce labeled public datasets and the need for methods to train effectively on them. When annotated data is scarce for a given Computer Vi- siontask,onecommonandpracticalsolutionistofine-tune a model that was pre-trained in a supervised manner us- ing the ImageNet dataset [19,34]. This paradigm of trans- ferlearning[34]wasrecentlychallengedbyself-supervised learning (SSL), which trains on large amounts of unlabeled data only, yet out-performs supervised pre-training on Ima- geNet [8,10,26]. In the field of pathology, large unlabeled datasetsareabundant[4,37,38,57]incontrasttothelackof annotateddatasets[52]. IfweweretoapplySSLeffectively to this huge amount of unlabeled data, downstream pathol- ogy tasks could benefit greatly even if they contain limited amount of annotated training data. Naturally, we ask the question: How well does self-supervised learning help in improving the performance of pathology tasks? ImageNet pre-trained weights are commonly used in medical imaging and are known to be helpful in attaining high task performance [30,32,43,59]. Due to the differ- ence between natural images and medical images, large- scale domain-aligned pre-training has the potential to push performance beyond ImageNet pre-training [39]. Accord- ingly, recent works show that SSL pre-training on pathol- 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. 3344 ogy data can improve performance on downstream pathol- ogy tasks [3,16,23,55]. Our study aims to expand on these previous works by evaluating multiple SSL methods on di- versedownstream pathologytasks. In addition,we propose techniquestoadaptSSLmethodsthatweredesignedfornat- ural image data, to better learn from pathology data. To understand how to adapt existing SSL methods to workonpathologyimagedata,wemustidentifyseveralkey differences between natural and pathology imagery. Unlike natural images, pathology images can be rotated arbitrar- ily (impossible to determine a canonical orientation) and exhibit fewer variations in color. Also, pathology images canbeinterpreteddifferentlydependingonthefield-of-view (FoV) due to the multiple hierarchies and contextual differ- ences involved in each task. We propose to overcome these differenceswhenadaptingSSLmethodsforpathologydata, viachangestothetrainingdataaugmentationschemeinpar- ticular, during pre-training. Inthispaper,wecarryoutanin-depthanalysisof4recent andrepresentativeSSLmethods;MoCov2[12],SwAV[7], Barlow Twins [61], and DINO [8], when applied to large- scale pathology data. For this purpose, we source 19 mil- lion image patches from Whole Slide Images (WSI) in The Cancer Genome Atlas (TCGA) dataset [57] and apply our domain-specific techniques in training the SSL methods on this data. The evaluations are conducted for 2 different downstream tasks over 5 datasets: (1) pathological image classificationusingBACH[1],CRC[31],MHIST[56],and PatchCamelyon [54] datasets, and (2) nuclei instance seg- mentationandclassificationusingtheCoNSePdataset[25]. Ourlarge-scalestudyyieldsseveralusefulcontributions: (a)weconductthelargest-scalestudyofSSLpre-trainingon pathologyimagedatatodate,andshowitsbenefitoverusing ImageNet pre-trained weights on diverse downstream tasks (see Fig. 1), (b) we propose a set of carefully designed data curation and data augmentation techniques that can further improve downstream performance, (c) we demonstrate that SSL is label-efficient, and is therefore a practical solution in pathology where gathering annotation is particularly ex- pensive,and(d)forthefirsttime,weapplySSLtothedense predictiontaskofnucleiinstancesegmentationandshowits valueunderdiverseevaluationsettings. Wereleaseourpre- trained model weights at https://lunit-io.github. io/research/publications/pathology_ssl to fur- ther contribute to the research community.
Jones_Self-Supervised_Representation_Learning_for_CAD_CVPR_2023	Abstract Virtually every object in the modern world was created, modified, analyzed and optimized using computer aided de- sign (CAD) tools. An active CAD research area is the use of data-driven machine learning methods to learn from the massive repositories of geometric and program represen- tations. However, the lack of labeled data in CAD’s na- tive format, i.e., the parametric boundary representation (B-Rep), poses an obstacle at present difficult to overcome. Several datasets of mechanical parts in B-Rep format have recently been released for machine learning research. How- ever, large-scale databases are mostly unlabeled, and la- beled datasets are small. Additionally, task-specific label sets are rare and costly to annotate. This work proposes to leverage unlabeled CAD geometry on supervised learn- ing tasks. We learn a novel, hybrid implicit/explicit surface representation for B-Rep geometry. Further, we show that this pre-training both significantly improves few-shot learn- ing performance and achieves state-of-the-art performance on several current B-Rep benchmarks.	1. Introduction Almost every human-made object that exists today started its life as a model in a CAD system. As the preva-lent method of creating 3D shapes, repositories of CAD models are extensive. Further, CAD models have a robust structure, including geometric and program representations that have the potential to expose design and manufactur- ing intent. Learning from CAD data can therefore enable a variety of applications in design automation and design- and fabrication-aware shape reconstruction and reverse en- gineering. An important challenge in learning from CAD is that most of this data does not have labels that can be lever- aged for inference tasks. Manually labeling B-Rep data is time consuming and expensive, and its specialized format requires CAD expertise, making it impractical for large col- lections. In this work we ask: how can we leverage large databases ofunlabeled CAD geometry for analysis and modeling tasks that typically require labels for learning? Our work is driven by a simple, yet fundamental observa- tion: the CAD data format was not developed to enable easy visualizing or straightforward geometric interpretation: it is a format designed to be compact, have infinite resolu- tion, and allow easy editing. Indeed, CAD interfaces con- sistently run sophisticated algorithms to convert the CAD representation into geometric formats for rendering. Driven by this observation, our key insight is to leverage large col- 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. 21327 lections of unlabeled CAD data to learn to geometrically interpret the CAD data format. We then leverage the net- works trained over the geometric interpretation task in su- pervised learning tasks where only small labeled collections are available. In other words, we use geometry as a model ofself-supervision and apply it to few-shot-learning. Specifically, we learn to rasterize local CAD geometry using an encoder-decoder structure. The standard CAD format encodes geometry as parametric boundary repre- sentations (B-Reps). B-Reps are graphs where the nodes are parametric geometry (surfaces, curves, and points) and edges denote the topological adjacency relationships be- tween the geometry. Importantly, the parametric geome- try associated with each node is unbounded , and bounds are computed from the topological relationships: curves bound- ing surfaces and points bounding curves. As shown in Fig- ure 2, the geometry of a B-Rep face is computed by clipping the surface primitive to construct a surface patch, where the clipping mask is constructed from adjacent edges. SurfaceBoundary S(u, v): ℝ2 → ℝ3 d(u, v): ℝ2 → ℝ Inside Outside(a) B-R ep Face (c) Im plicit Bound ary (d) Implicit Bound ary (SDF) (b) Explicit Sur face 0011 v uM(u, v): ℝ2 → { in, out} 0011 v u -0.4 -0.2 0.0 0.2 0.4 uv Figure 2. A B-Rep face (a) is a surface patch cut from a geometric primitive surface (b). The adjacent edges define a clipping mask (c), which we learn an SDF (d). Thus, B-reps are constructed piecewise by explicitly de- fined surfaces with implicitly defined boundaries. This ob- servation drives our proposed learning architecture, which reconstructs faces by jointly decoding the explicit surface parameterization as well as the implicit surface boundary. Our proposed encoder uses message passing on the topo- logical graph to capture the boundary information to encode B-Rep faces. To handle graph heterogeneity (nodes com- prised of faces, edges, and, vertices), we use a hierarchical message passing architecture inspired by the Structured B- Rep GCN [16]. Our decoder uses the learned embeddings as latent codes for two per-face neural function evaluators: one mapping from R2→R3that encodes the face’s para- metric surface (Figure 2 (b)), and one mapping R2→Rthat encodes the face’s boundary as a signed distance field (SDF) within the parametric surface (Figure 2 (c,d)). We apply our proposed model of B-Rep self-supervision to learn specialized B-Rep tasks from very small sets of la- beled data—10s to 100s of examples vs 10k to 100k. To do this, we use the embeddings learned on self-supervision as input features to supervised tasks. We evaluate our ap- proach on three tasks and datasets from prior work [3,6,21] and validate our findings across varying training set sizes. We show that our model consistently outperforms prior su- pervised approaches, significantly improving performance on smaller training sets. By using less data, our approach also proves substantially faster to train, making possible ap- plications that depend on training speed. We believe that our differentiable CAD rasterizer paves the way to many exciting future applications, and show one possibility by prototyping a reverse engineering example.
Kang_Soft-Landing_Strategy_for_Alleviating_the_Task_Discrepancy_Problem_in_Temporal_CVPR_2023	Abstract Temporal Action Localization (TAL) methods typically operate on top of feature sequences from a frozen snippet encoder that is pretrained with the Trimmed Action Classiﬁ- cation (TAC) tasks, resulting in a task discrepancy problem. While existing TAL methods mitigate this issue either by re- training the encoder with a pretext task or by end-to-end ﬁne- tuning, they commonly require an overload of high memory and computation. In this work, we introduce Soft-Landing (SoLa) strategy, an efﬁcient yet effective framework to bridge the transferability gap between the pretrained encoder and the downstream tasks by incorporating a light-weight neural network, i.e., a SoLa module, on top of the frozen encoder . We also propose an unsupervised training scheme for the SoLa module; it learns with inter-frame Similarity Match- ing that uses the frame interval as its supervisory signal, eliminating the need for temporal annotations. Experimen- tal evaluation on various benchmarks for downstream TAL tasks shows that our method effectively alleviates the task discrepancy problem with remarkable computational efﬁ- ciency.	1. Introduction Our world is full of untrimmed videos, including a plethora of Y outube videos, security camera recordings, and online streaming services. Analyzing never-ending video streams is thus one of the most promising directions of com- puter vision research in this era [29]. Amongst many long- form video understanding tasks, the task of ﬁnding action instances in time and classifying their categories, known as Temporal Action Localization (T AL), has received intense attention from both the academia and the industry in recent years; T AL is considered to be the fundamental building block for more sophisticated video understanding tasks since it plays the basic role of distinguishing frame-of-interest from irrelevant background frames [19, 34, 40]./g7/g25/g17/g26/g18/g19/g28/g9/g25/g27/g31/g30/g1/g8/g28/g15/g24/g19/g29/g14/g5/g10/g3/g21/g19/g15/g18/g14/g5/g10/g3/g21/g19/g15/g18 /g13/g26/g10/g15/g13/g30/g15/g25/g18/g15/g28/g18/g1/g14/g5/g10 /g6/g22/g20/g20/g19/g28/g19/g25/g30/g1 /g11/g28/g19/g30/g19/g32/g30/g1/g14/g15/g29/g23/g29/g15/g2 /g7/g25/g18/g3/g30/g26/g3/g7/g25/g18/g1/g14/g5/g10 /g16/g2 /g9/g25/g27/g31/g30/g1/g8/g28/g15/g24/g19/g29 /g4/g1/g12/g19/g29/g19/g15/g28/g17/g21/g1/g8/g26/g17/g31/g29 /g10/g3/g1/g5/g23/g16/g20/g21/g16/g17/g14 /g1/g9/g4/g6/g1/g7/g13/g21/g15/g18/g12/g20 /g11/g3 /g1/g8/g18/g6/g10 /g2/g18/g22/g19/g20/g3/g4/g1/g13/g25/g22/g27/g27/g19/g30/g1/g8/g19/g15/g30/g31/g28/g19/g29/g7/g25/g17/g26/g18/g19/g28 /g4/g1/g11/g15/g28/g15/g24/g19/g30/g19/g28/g29/g1/g8/g28/g26/g33/g19/g25 /g4/g1/g7/g25/g21/g15/g25/g17/g19/g18/g1/g13/g25/g22/g27/g27/g19/g30/g1/g8/g19/g15/g30/g31/g28/g19/g29 /g22/g22/g2 /g22/g22/g22/g2/g1/g22/g2 Figure 1. (a-i) Standard T AL framework assumes a “frozen snip- pet encoder” and only focuses on designing a good T AL head, causing the task discrepancy problem. Straightforward approaches to alleviating the issue include devising (a-ii) a temporally sensi- tive pretext task [1, 32, 41], and (a-iii) an end-to-end T AL training procedure [20, 33]. However, both approaches break the aforemen- tioned frozen snippet encoder assumption. On the other hand, (b) our SoLa strategy shares the “frozen snippet encoder assumption” with the standard T AL framework by providing a smooth linkage between the frozen encoder and the downstream head. It offers gen- eral applicability and more importantly, exceptional computational efﬁciency. Despite its importance, training a T AL model has a unique computational challenge that hinders the naive extension of conventional image processing models, mainly due to the large size of the model input. For instance, videos in the wild can be several minutes or even hours long, implying that loading the whole video to a device for processing isoften infeasible. In this context, the prevailing convention in processing a long-form video for T AL is to divide the video into non-overlapping short snippets and deal with the snippet-wise feature sequences. Speciﬁcally, a standard train- 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. 6514 ing pipeline for the long-form video understanding tasks consists of two steps: (i) Train the snippet encoder with a large-scale action recognition dataset (e.g., Kinetics400), which is often different from the dataset for the downstream task; (ii) Train the downstream head (e.g., T AL) that takes the snippet feature sequences extracted from the pretrained encoder. An issue here is that the mainstream pretext task for the snippet-wise video encoder is “Trimmed” Action Classi- ﬁcation (T AC), which does not handle action boundaries and background frames. Although the current pipeline achieves remarkable performance in T AL tasks due to the power of large action recognition datasets, recent works [1, 32, 33, 41] point out the task discrepancy problem that is inherent in this two-staged approach. The task discrepancy problem,ﬁrst introduced in [33], comes from the pretrained snippet encoder’s insensitivity to different snippets within the same action class. It results in a temporally invariant snippet feature sequence, making it hard to distinguish foreground actions from backgrounds. A straightforward approach to the problem is adopting a temporally sensitive pretext task to train the snippet encoder [1, 32], or devising an end-to-end framework [20, 33], which are brieﬂy described in Figure 1 (a). However, as all previous methods involve retraining the snippet encoder, an excessive use of memory and computa- tion is inevitable. To tackle the task discrepancy problem, we propose a new approach, namely Soft-Landing (SoLa) strategy, which is neither memory nor computationally expensive. SoLa strategy is a novel method which incorporates a light-weight neural network, i.e., Soft-Landing (SoLa) module, between the pretrained encoder and the downstream head. The prop- erly trained SoLa module will act like a middleware between the pretext and the downstream tasks, mitigating the task dis- crepancy problem (Figure 1 (b)). Since the task adaptation is solely done by the SoLa module, the parameters of thepretrained encoder are ﬁxed in our SoLa strategy. The use of a frozen encoder signiﬁcantly differentiates our approach from previous methods that mainly focus on designing an appropriate training methodology for a snippet encoder. In addition, our SoLa strategy only requires an access to the pre- extracted snippet feature sequence, being fully compatible with the prevailing two-stage T AL framework. We also propose Similarity Matching , an unsupervised training scheme for the SoLa module that involves neither frame-level data manipulation nor temporal annotations. Our training strategy circumvents the need for strong frame-level data augmentation which most existing unsupervised repre- sentation learning techniques [5, 10] rely on. This strategyperfectly suits our condition where frame-level data aug-mentation is impossible, as we only have an access to pre- extracted snippet features. The new loss is based on a simple empirical observation: “adjacent snippet features are simi- lar, while distant snippet features remain distinct”. Coupledwith the Simsiam [6] framework, Similarity Matching not only prevents the collapse, but also induces temporally sen- sitive feature sequences, resulting in a better performance in various downstream tasks. The contributions of the paper can be summarized as follows: •To tackle the task discrepancy problem, we introduce a novel Soft-Landing (SoLa) strategy, which does not involve retraining of the snippet encoder. As we can directly deploy the “frozen” pretrained snippet encoder without any modiﬁcation, our SoLa strategy offers eas- ier applicability compared to previous works that re- quire snippet encoders to be retrained. •We propose Similarity Matching , a new self- supervised learning algorithm for the SoLa strategy. As frame interval is utilized as its only learning signal, it requires neither data augmentation nor temporal an- notation. •With our SoLa strategy, we show signiﬁcant improve- ment in performance for downstream tasks, outperform- ing many of the recent works that involve computation- ally heavy snippet encoder retraining.
Kang_The_Dialog_Must_Go_On_Improving_Visual_Dialog_via_Generative_CVPR_2023	Abstract Visual dialog (VisDial) is a task of answering a sequence of questions grounded in an image, using the dialog history as context. Prior work has trained the dialog agents solely on VisDial data via supervised learning or leveraged pre- training on related vision-and-language datasets. This paper presents a semi-supervised learning approach for visually- grounded dialog, called Generative Self-Training (GST), to leverage unlabeled images on the Web. Specifically, GST first retrieves in-domain images through out-of-distribution detection and generates synthetic dialogs regarding the im- ages via multimodal conditional text generation. GST then trains a dialog agent on the synthetic and the original Vis- Dial data. As a result, GST scales the amount of training data up to an order of magnitude that of VisDial (1.2M → 12.9M QA data). For robust training of the synthetic di- alogs, we also propose perplexity-based data selection and multimodal consistency regularization. Evaluation on Vis- Dial v1.0 and v0.9 datasets shows that GST achieves new state-of-the-art results on both datasets. We further observe the robustness of GST against both visual and textual ad- versarial attacks. Finally, GST yields strong performance gains in the low-data regime. Code is available at https: //github.com/gicheonkang/gst-visdial .	1. Introduction Recently, there has been extensive research towards de- veloping visually-grounded dialog systems [12, 13, 34, 36] due to their significance in many real-world applications (e.g., helping visually impaired person). Notably, Visual Di- alog (VisDial) [12] has provided a testbed for studying such systems, where a dialog agent should answer a sequence of image-grounded questions. For instance, the agent is ex- pected to answer open-ended questions like “What color is it?” and“How old does she look?” . This task requires a holistic understanding of visual information, linguistic se- *Equal contributionmantics in context ( e.g., it and she), and most importantly, the grounding of these two. Most of the previous approaches in VisDial [9, 10, 18, 20, 25, 26, 30, 31, 35, 49, 54, 55, 64, 67, 78, 84] have trained the dialog agents solely on VisDial data via supervised learning. More recent studies [8,53,77] have employed self-supervised pre-trained models such as BERT [14] or ViLBERT [48] and finetuned them on VisDial data. The models are typi- cally pre-trained to recover masked inputs and predict the semantic alignment between two segments. This pretrain- then-transfer learning strategy has shown promising results by transferring knowledge from the models pre-trained on large-scale data sources [4, 71, 85] to VisDial. Our research question is the following: How can the dia- log agent expand its knowledge beyond what it can acquire via supervised learning or self-supervised pre-training on the provided datasets? Some recent studies have shown that semi-supervised learning and pre-training have complemen- tary modeling capabilities in image [86] and text classifica- tion [16]. Inspired by them, we consider semi-supervised learning (SSL) as a way to address the above question. Let us assume that large amounts of unlabeled images are available. SSL for VisDial can be applied to generate synthetic conversations for the unlabeled images and train the agent with the synthetic data. However, there are two critical challenges to this approach. First, the target output for VisDial ( i.e.,multi-turn visual QA data) is more complex than that of the aforementioned studies [16,86]. Specifically, they have addressed the classification problems, yielding class probabilities as pseudo labels [39]. In contrast, SSL for VisDial should generate a sequence of pseudo queries (i.e.,visual questions) and pseudo labels ( i.e.,corresponding answers) in natural language to train the answering agent. It further indicates that the target output should be generated while considering the multimodal andsequential nature of the visual dialog task. Next, even if SSL yields synthetic dialogs via text generation, there may be noise, such as generating irrelevant questions or incorrect answers to given contexts. A robust training method is required to leverage such noisy synthetic dialog datasets. 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. 6746 In this paper, we study the above challenges in the con- text of SSL, especially self-training [6, 16, 21, 28, 32, 39, 44, 52, 60, 65, 72, 73, 79, 80, 86], where a teacher model trained on labeled data predicts the pseudo labels for unlabeled data. Then, a student model jointly learns on the labeled and the pseudo-labeled datasets. Unlike existing studies in self-training that have mainly studied uni-modal, discrimi- native tasks such as image classification [72, 80, 86] or text classification [16, 32, 52], we extend the idea of self-training to the task of multimodal conditional text generation. To this end, we propose a new learning strategy, called Generative Self-Training (GST), that artificially generates multi-turn visual QA data and utilizes the synthetic data for training. GST first trains the teacher model (answerer) and the visual question generation model (questioner) using VisDial data. It then retrieves a set of unlabeled images from a Web image dataset, Conceptual 12M [7]. Next, the questioner and the teacher generate a series of visual QA pairs for the retrieved images. Finally, the student is trained on the synthetic and the original VisDial data. We also pro- pose perplexity-based data selection (PPL) and multimodal consistency regularization (MCR) to effectively train the student with the noisy dialog data. PPL is to selectively utilize the answers whose perplexity of the teacher is below a threshold. MCR encourages the student to yield consis- tent predictions when the perturbed multimodal inputs are given. As a result, GST successfully augments the synthetic VisDial data (11.7M QA pairs), thus mitigating the need to scale up the size of the human-annotated VisDial data, which is prohibitively expensive and time-consuming. Our key contributions are three-fold. First, we propose Generative Self-Training (GST) that generates multi-turn visual QA data to leverage unlabeled Web images effectively. Second, experiments show that GST achieves new state- of-the-art performance on VisDial v1.0 and v0.9 datasets. We further demonstrate two important results: (1) GST is indeed effective when the human-annotated visual dialog data is extremely scarce (improving up to 11.09 absolute points on NDCG), and (2) PPL and MCR are effective when training the noisy synthetic dialog data. Third, to validate the robustness of GST, we evaluate our proposed method under three different visual and textual adversarial attacks, i.e., FGSM, coreference, and random token attacks. We observe that GST significantly improves the performance compared with the baseline models against all adversarial attacks, especially boosting NDCG scores from 21.60% to 45.43% in the FGSM attack [19].
Kalluri_GeoNet_Benchmarking_Unsupervised_Adaptation_Across_Geographies_CVPR_2023	Abstract In recent years, several efforts have been aimed at im- proving the robustness of vision models to domains and environments unseen during training. An important practi- cal problem pertains to models deployed in a new geography that is under-represented in the training dataset, posing a direct challenge to fair and inclusive computer vision. In this paper, we study the problem of geographic robust- ness and make three main contributions. First, we intro- duce a large-scale dataset GeoNet for geographic adapta- tion containing benchmarks across diverse tasks like scene recognition (GeoPlaces), image classification (GeoImNet) and universal adaptation (GeoUniDA). Second, we inves- tigate the nature of distribution shifts typical to the prob- lem of geographic adaptation and hypothesize that the ma- jor source of domain shifts arise from significant varia- tions in scene context (context shift), object design (de- sign shift) and label distribution (prior shift) across ge- ographies. Third, we conduct an extensive evaluation of several state-of-the-art unsupervised domain adaptation al- gorithms and architectures on GeoNet, showing that they do not suffice for geographical adaptation, and that large-scale pre-training using large vision models also does not lead to geographic robustness. Our dataset is publicly available at https://tarun005.github.io/GeoNet .	1. Introduction In recent years, domain adaptation has emerged as an effective technique to alleviate dataset bias [80] during train- ing and improve transferability of vision models to sparsely labeled target domains [27, 36, 40, 42, 49 –51, 68, 69, 86, 89]. While being greatly instrumental in driving research forward, methods and benchmark datasets developed for domain adap- tation [56, 57, 64, 83] have been restricted to a narrow set of divergences between domains. However, the geographic ori- gin of data remains a significant source of bias, attributable to several factors of variation between train and test data. Train- ing on geographically biased datasets may cause a model to learn the idiosyncrasies of their geographies, preventing Context ShiftDesign Shift Training DomainUSA Testing DomainAsia+(a)Geographic bias manifested in proposed GeoNet dataset DANN MDD MCC CDAN ToAlign MemSAC 10 010203040Rel. Accuracy GainDomainNet (Real->Clipart) GeoImNet (USA->Asia) (b)Unsupervised domain adaptation does not suffice on GeoNet RN-50 (24.0M) ViT-S (21.8M) ViT-B (85.9M) ViT-L (303.3M) Backbone (#params)30405060Accuracy (%)Target Supervised Sup.-IN 1k MOCO-IN 1kSWAV-IN 1k DINO-IN 1k MAE-IN 1kSWAG-IG 3.7B CLIP-LAION 400M (c)Large vision models exhibit cross-domain drops on GeoNet Figure 1. Summary of our contributions . (a): Training computer vision models on geographically biased datasets suffers from poor generaliza- tion to new geographies. We propose a new dataset called GeoNet to study this problem and take a closer look at the various types of domain shifts induced by geographic variations. (b) Prior unsupervised adapta- tion methods that efficiently handle other variations do not suffice for improving geographic transfer. (c) We highlight the limitations of mod- ern convolutional and transformer architectures in addressing geographic bias, exemplified here by USA →Asia transfer on GeoImNet. generalization to novel domains with significantly different geographic and demographic composition. Besides robust- ness, this may have deep impact towards fair and inclusive 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. 15368 computer vision, as most modern benchmark datasets like ImageNet [63] and COCO [47] suffer from a significant US or UK-centric bias in data [24, 73], with poor representation of images from various other geographies like Asia. In this paper, we study the problem of geographic adaptation by introducing a new large-scale dataset called GeoNet, which constitutes three benchmarks – GeoPlaces for scene classification, GeoImNet for object recognition and GeoUniDA for universal domain adaptation. These bench- marks contain images from USA and Asia, which are two distinct geographical domains separated by various cultural, economic, demographic and climatic factors. We addition- ally provide rich metadata associated with each image, such as GPS location, captions and hashtags, to facilitate algo- rithms that leverage multimodal supervision. GeoNet captures the multitude of novel challenges posed by varying image and label distributions across geographies. We analyze GeoNet through new sources of domain shift caused by geographic disparity, namely (i) context shift , where the appearance and composition of the background in images changes significantly across geographies, (ii) design shift, where the design and make of various objects changes across geographies, and (iii) prior shift , caused by different per-category distributions of images in both domains. We illustrate examples of performance drop caused by these fac- tors in Fig. 1a, where models trained on images from USA fail to classify common categories such as running track and mailbox due to context and design shifts, respectively. GeoNet is an order of magnitude larger than previous datasets for geographic adaptation [58, 61], allowing the training of modern deep domain adaptation methods. Im- portantly, it allows comparative analysis of new challenges posed by geographic shifts for algorithms developed on other popular adaptation benchmarks [56, 57, 64, 83]. Specifically, we evaluate the performance of several state-of-the-art un- supervised domain adaptation algorithms on GeoNet, and show their limitations in bridging domain gaps caused by geographic disparities. As illustrated in Fig. 1b for the case of DomainNet [56] vs. GeoNet, state-of-the-art models on DomainNet often lead to accuracies even worse than a source only baseline on GeoNet, resulting in negative relative gain in accuracy (defined as the gain obtained by an adaptation method over a source-only model as a percentage of gap be- tween a source-only model and the target-supervised upper bound). Furthermore, we also conduct a study of modern ar- chitectures like vision transformers and various pre-training strategies, to conclude that larger models with supervised and self-supervised pre-training offer improvements in accu- racy, which however are not sufficient to address the domain gap (Fig. 1c). This highlights that the new challenges intro- duced by geographic bias such as context and design shift are relatively under-explored, where our dataset may motivate further research towards this important problem.In summary, our contribution towards geographic domain adaptation is four-fold: •A new large-scale dataset, GeoNet, with benchmarks for diverse tasks like scene classification and object recogni- tion, with labeled images collected from geographically distant locations across hundreds of categories (Sec. 3). •Analysis of domain shifts in geographic adaptation, which may be more complex and subtle than style or appearance variations (Sec. 3.4). •Extensive benchmarking of unsupervised adaptation al- gorithms, highlighting their limitations in addressing geographic shifts (Sec. 4.2). •Demonstration that large-scale pretraining and recent advances like vision transformers do not alleviate these geographic disparities (Sec. 4.3).
Liao_AttentionShift_Iteratively_Estimated_Part-Based_Attention_Map_for_Pointly_Supervised_Instance_CVPR_2023	Abstract Pointly supervised instance segmentation (PSIS) learns to segment objects using a single point within the object extent as supervision. Challenged by the non-negligible se- mantic variance between object parts, however, the single supervision point causes semantic bias and false segmenta- tion. In this study, we propose an AttentionShift method, to solve the semantic bias issue by iteratively decomposing the instance attention map to parts and estimating fine-grained semantics of each part. AttentionShift consists of two mod- ules plugged on the vision transformer backbone: (i) to- ken querying for pointly supervised attention map genera- tion, and (ii) key-point shift, which re-estimates part-based attention maps by key-point filtering in the feature space. These two steps are iteratively performed so that the part- based attention maps are optimized spatially as well as in the feature space to cover full object extent. Experiments on PASCAL VOC and MS COCO 2017 datasets show that AttentionShift respectively improves the state-of-the-art of by 7.7% and 4.8% under [email protected], setting a solid PSIS baseline using vision transformer.	1. Introduction Instance segmentation is one of the most important vi- sion tasks with a wide range of applications in medical im- age processing [21, 23, 36], human machine interface [3, 7, 24] and advanced driver assistance system [14,37,41]. Nev- ertheless, this task requires great human efforts to annotate instance masks, particular in the era of big data. For exam- ple, it takes more than four years to create the ground-truth instance masks in the MS COCO dataset by a human anno- tator [29]. The large annotation cost hinders the deployment *Equal Contribution (Liao: Idea, Experiment; Guo: Detection Base- line, Writing). †Corresponding Author. Self-attention Map Class Activation Map Image & Supervision Instance Attention Map Attention ShiftPart-based Attention MapBird Image & Supervision Class Activation Map Self-attention Map Instance Attention Map Part-based Attention Map Attention ShiftDogFigure 1. Comparison of existing methods with the pro- posed AttentionShift. Upper: The class activation map (CAM) method [51] suffers partial activation for the lack of spatial constraint. The self-attention map generated by vision trans- former [16] encounters false and missed object parts. Lower: At- tentionShift literately optimizes part-based attention maps (indi- cated by key-points) by shifting the key-points in the feature space to precisely localize the full object extent. (Best viewed in color) of instance segmentation to real-world applications. Pointly supervised instance segmentation (PSIS) [25, 26], where each instance is indicated by a single point, has been a promising approach to solve the annotation cost is- sue. Compared with the precise mask annotation, PSIS re- quires only about 20% annotation cost, which is comparable with the weakly supervised method, while the performance is far beyond the later [2]. Existing methods [25, 26] typically estimate a single pseudo mask for each instance and refine the estimated mask by training a segmentation model. However, such methods ignore the fact that the semantic variance between object parts is non-negligible. For example, a “dog head” 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. 19519 ViT Key-point Shift in Feature SpaceImage Attention Map Part-based Attention Maps Shifted Attention Maps Ground -truth & Estimated PointsPatch Token QueryingPoint SupervisionQuery Tokens Estimated Key-points ViT Key-point Shift in Feature Space Image Attention Map Part-based Attention Maps Shifted Attention Maps Ground -truth & Estimated PointsPatch Token QueryingPoint Supervision Query Tokens Estimated Key-points Figure 2. Flowchart of AttenttionShift. During training, the model first learns the fine-grained semantics by decomposing the instance attention map to parts (indicated by key-points) and performing key-point shift the feature space. It then takes the estimated key-points (each represents a part) as supervisions and querying their locations using the vision transformer. and a “dog tail” belongs to the same semantic of “dog”, but their appearances are quite different, Fig. 1 upper. With a single supervision point, these methods could estimate the semantic of “dog” as that of the most discriminative part (“dog tail” in middle of Fig. 1 upper) or that of the regions around the supervision point (“dog head” in right of Fig. 1 upper), which is termed as semantic bias . We make a statis- tical analysis on ViT and found only 33% of ViT attention maps can cover over 50% of foreground pixels. Despite its ability to model long-range feature dependencies, ViT ap- pears to remain vulnerable to the issue of semantic bias. In this study, we propose AttentionShift to solve the se- mantic bias problem by estimating fine-grained semantics of multiple instance parts and learning an instance segmen- tor under the supervision of estimated fine-grained seman- tics, Fig. 1(lower). Considering that instance parts are un- available during training, AttentionShift adopts an iterative optimization procedure, which spatially decomposes each instance to parts based on key-points defined on mean fea- ture vectors and adaptively updates the key-points in a way like mean shift [11] in the feature space, Fig. 2. Using the vision transformer (ViT) as the backbone, At- tentionShift consists of two steps: (i) token querying of point supervision for instance attention map generation. It takes advantage of the ViT to generate an instance attention map by matching the semantics and locations of patch to- kens with those of the supervision point. (ii) key-point shift which re-estimates the part-based attention map by key- point initialization and filtering in the feature space. These two steps are iteratively performed so that the part-based at- tention map, indicated by key-points, is optimized spatially as well as in the feature space to cover the full object extent.We conduct experiments on commonly used PASCAL VOC and the challenging MS COCO 2017 datasets. At- tentionShift respectively improves the state-of-the-art of by 7.7% and 4.8% under [email protected], demonstrating the poten- tial to fill the performance gap between pointly-supervised instance segmentation and fully supervised instance seg- mentation methods. The contributions of this paper are concluded as follows: • We propose a part-based attention map estimation ap- proach for PSIS, which estimates fine-grained seman- tics of instances to alleviate the semantic bias problem in a systematic fashion. • We represent object parts using key-points and lever- age AttentionShift to operate key-points in the feature space, providing a simple-yet-effective fashion to op- timize part-based attention maps. • AttentionShift achieves state-of-the-art performance, setting a solid PSIS baseline using vision transformer.
Liu_SlowLiDAR_Increasing_the_Latency_of_LiDAR-Based_Detection_Using_Adversarial_Examples_CVPR_2023	Abstract LiDAR-based perception is a central component of au- tonomous driving, playing a key role in tasks such as ve- hicle localization and obstacle detection. Since the safety of LiDAR-based perceptual pipelines is critical to safe au- tonomous driving, a number of past efforts have investi- gated its vulnerability under adversarial perturbations of raw point cloud inputs. However, most such efforts have fo- cused on investigating the impact of such perturbations on predictions (integrity), and little has been done to under- stand the impact on latency (availability), a critical con- cern for real-time cyber-physical systems. We present the first systematic investigation of the availability of LiDAR detection pipelines, and SlowLiDAR, an adversarial per- turbation attack that maximizes LiDAR detection runtime. The attack overcomes the technical challenges posed by the non-differentiable parts of the LiDAR detection pipelines by using differentiable proxies and uses a novel loss function that effectively captures the impact of adversarial perturba- tions on the execution time of the pipeline. Extensive ex- perimental results show that SlowLiDAR can significantly increase the latency of the six most popular LiDAR detec- tion pipelines while maintaining imperceptibility1.	1. Introduction The promise of autonomous transit has stimulated ex- tensive efforts towards the development of self-driving plat- forms [1–3, 5]. A central feature of most such platforms is a LiDAR-based perceptual pipeline (usually integrated with other sensors, such as camera and radar) for critical control tasks, such as localization and object detection [1, 3, 51]. Since errors in localization or obstacle detection can cause the vehicle to crash, these tasks are crucial in ensuring the safety of an autonomous vehicle. As a result, extensive prior research has been devoted to understanding and assur- ing the robustness of the LiDAR-based perceptual pipelines 1Code is available at: https://github.com/WUSTL-CSPL/SlowLiDAR Add Perturb Figure 1. SlowLiDAR attack. The attack goal is to maximize the runtime latency of the state-of-the-art LiDAR detection models with either an adding-based attack or a perturbation-based attack. against adversarial perturbations on its raw point cloud in- puts [22, 51, 53, 54]. A key focus in prior work has been on perceptual prediction accuracy, for example, the ability of the adversary to hide real obstacles or create phantom obsta- cles [15,32,43,44]. Another significant aspect of the robust- ness analysis of LiDAR-based perception that has received little attention is its real-time performance (availability) . In particular, a delay in perceptual processing caused solely by computational quirks of the LiDAR processing pipeline can be just as damaging as a mistaken prediction. For example, a delay in obstacle recognition can cause the vehicle to react to an obstacle too late, failing to avoid a crash. We present the first systematic analysis of the impact of adversarial point cloud perturbations on the runtime la- tency of common LiDAR perceptual pipelines. To this end, we propose SlowLiDAR, an algorithmic framework for adding adversarial perturbations to point cloud data aim- ing to increase the execution time of LiDAR-based ob- ject detection. Specifically, we consider two attack mech- anisms (see Figure 1): 1) perturbation of the 3D coor- dinates of existing points in the raw point cloud ( point perturbation attacks ), and 2) adding points to the raw point cloud ( point addition attacks ). Due to the unique representation and processing procedure of LiDAR detec- 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. 5146 tion pipelines, existing availability attacks on other AI pipelines, such as camera-based detection [39], cannot be directly applied. Specifically, there are three new chal- lenges: C1. Non-differentiable aggregation. LiDAR point clouds are generally sparsely distributed in a large 3D space without an organized pattern. To process such un- structured data, the state-of-the-art LiDAR-based detec- tion models extract aggregated features of the points at the level of either 2D or 3D cells [30, 52, 55]. However, the aggregation operation is by nature non-differentiable [14], which presents new challenges to end-to-end opti- mization. C2. New objective function. Different than ex- isting attacks targeting accuracy, a novel loss function needs to be carefully designed to effectively capture the impact of adversarial perturbations on runtime latency. C3. Large search space. Because points are distributed in a large space, the perturbation search space is quite large, and conventional gradient-based optimization approaches for crafting attacks may yield poor local optima. We address these technical challenges as follows. First, in order to enable end-to-end training of adversarial input perturbations, we develop a differentiable proxy to approx- imate the non-differentiable pre-processing pipelines. Sec- ond, we conduct a response time analysis on the detection pipeline to identify the most vulnerable components, and design a novel loss function to best capture the impact of input perturbations on the runtime latency of the identified components. Third, to tackle the large search space, we propose a probing algorithm to identify high-quality initial- ization for gradient-based attack optimization. We evaluate SlowLiDAR on six popular LiDAR-based detection frameworks that are adopted in modern commer- cial autonomous driving systems. Our experiments show that SlowLiDAR is effective in slowing down the models while retaining comparable imperceptibility. Moreover, the performance of our attacks remains consistent across differ- ent hardware and implementations. Our contributions can be summarized as follows: • We systematically dissect the LiDAR detection models to analyze the attack surface of their runtime latency to adversarial inputs. • We propose the first adversarial perturbation attacks against LiDAR detection with the goal of maximizing runtime latency. To this end, we overcome the non- differentiability of the pre-processing pipelines in or- der to perform end-to-end gradient-based attack opti- mization, and design a novel loss function to capture the impact of input perturbations on execution time. • We evaluate our attacks on six popular LiDAR detec- tion models in modern commercial autonomous driv- ing systems in different hardware and implementations to demonstrate the ability to significantly increase la- tency with comparable imperceptibility.
Liu_Delving_Into_Shape-Aware_Zero-Shot_Semantic_Segmentation_CVPR_2023	Abstract Thanks to the impressive progress of large-scale vision- language pretraining, recent recognition models can clas- sify arbitrary objects in a zero-shot and open-set manner, with a surprisingly high accuracy. However, translating this success to semantic segmentation is not trivial, because this dense prediction task requires not only accurate semantic understanding but also fine shape delineation and existing vision-language models are trained with image-level lan- guage descriptions. To bridge this gap, we pursue shape- aware zero-shot semantic segmentation in this study. In- spired by classical spectral methods in the image segmenta- tion literature, we propose to leverage the eigen vectors of Laplacian matrices constructed with self-supervised pixel- wise features to promote shape-awareness. Despite that this simple and effective technique does not make use of the masks of seen classes at all, we demonstrate that it out- performs a state-of-the-art shape-aware formulation that aligns ground truth and predicted edges during training. We also delve into the performance gains achieved on dif- ferent datasets using different backbones and draw several interesting and conclusive observations: the benefits of pro- moting shape-awareness highly relates to mask compact- ness and language embedding locality. Finally, our method sets new state-of-the-art performance for zero-shot seman- tic segmentation on both Pascal and COCO, with significant margins. Code and models will be accessed at SAZS.	1. Introduction Semantic segmentation has been an established research area for some time now, which aims to predict the categories of an input image in a pixel-wise manner. In real-world ap- plications including autonomous driving [17], medical diag- nosis [31,46] and robot vision and navigation [9,63], an ac- curate semantic segmentation module provides a pixel-wise Pizza* Dining Table Cup* Cat Input Images Ours w/o shape awarenessOurs Ground TruthFigure 1. Without retraining, SAZS is able to precisely segments both seen and unseen objects in the zero-shot setting, largely out- performing a strong baseline. *denotes unseen categories during training). understanding of the input image and is crucial for subse- quent tasks (like decision making or treatment selection). Despite that significant progress has been made in the field of semantic segmentation [6,7,32,49,52,54,57,59,61], most existing methods focus on the closed-set setting in which dense prediction is performed on the same set of cat- egories in training and testing time. Thus, methods that are trained and perform well in the closed-set setting may fail when applied to the open world, as pixels of unseen ob- jects in the open world are likely to be assigned categories that are seen during training, causing catastrophic conse- quences in safety-critical applications such as autonomous driving [62]. Straightforward solutions include fine-tuning or retraining the existing neural networks, but it is impracti- cal to enumerate unlimited unseen categories during retrain- ing, let along large quantities of time and efforts needed. More recent works [4, 14, 24, 27, 40] address this issue by shifting to the zero-shot setting, in which the methods are evaluated with semantic categories that are unseen dur- ing training. While large-scale pre-trained visual-language models such as CLIP [41] or ALIGN [18] shed light on 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. 2999 potential of solving zero-shot tasks with priors contained in large-scale pre-trained model, how to perform dense predic- tion task in this setting is still under-explored. One recent approach by Li et.al. [24] closes the gap by leveraging the shared embedding space for languages and images, but fails to effectively segment regions with fine shape delineation. If the segmented shape of the target object is not accurate, it will be a big safety hazard in practical applications, such as in autonomous driving. Inspired by the classical spectral methods and their in- trinsic capability of enhancing shapeawareness, we pro- pose a novel Shape- Aware Zero-Shot semantic segmenta- tion framework ( SAZS ) to address the task of zero-shot semantic segmentation. Firstly, the framework enforces vision-language alignment on the training set using known categories, which exploits rich language priors in the large- scale pre-trained vision-language model CLIP [41]. Mean- while, the framework also jointly enforces the boundary of predicted semantic regions to be aligned with that of the ground truth regions. Lastly, we leverage the eigenvectors of Laplacian of affinity matrices that is constructed by fea- tures learned in a self-supervised manner, to decompose in- puts into eigensegments. They are then fused with learning- based predictions from the trained model. The fusion out- puts are taken as the final predictions of the framework. As illustrated in Fig. 1, compared with [24], the predic- tions of our approach are better aligned with the shapes of objects. We also demonstrate the effectiveness of our approach with elaborate experiments on PASCAL- 5iand COCO- 20i, the results of which show that our method out- performs former state-of-the-arts [4, 24, 36, 37, 51, 53] by large margins. By examining a) the correlation between shape compactness of target object and IoU and b) the correlation between the language embedding locality and IoU, we discover the large impacts on the performance brought by the distribution of language anchors and ob- ject shapes. Via extensive analyses, we demonstrate the ef- fectiveness and generalization of SAZS framework’s shape perception for segmenting semantic categories in the open world.
Liu_EfficientViT_Memory_Efficient_Vision_Transformer_With_Cascaded_Group_Attention_CVPR_2023	Abstract Vision transformers have shown great success due to their high model capabilities. However, their remarkable performance is accompanied by heavy computation costs, which makes them unsuitable for real-time applications. In this paper, we propose a family of high-speed vision trans- formers named EfficientViT. We find that the speed of ex- isting transformer models is commonly bounded by mem- ory inefficient operations, especially the tensor reshaping and element-wise functions in MHSA. Therefore, we design a new building block with a sandwich layout, i.e., using a single memory-bound MHSA between efficient FFN layers, which improves memory efficiency while enhancing channel communication. Moreover, we discover that the attention maps share high similarities across heads, leading to com- putational redundancy. To address this, we present a cas- caded group attention module feeding attention heads with different splits of the full feature, which not only saves com- putation cost but also improves attention diversity. Compre- hensive experiments demonstrate EfficientViT outperforms existing efficient models, striking a good trade-off between speed and accuracy. For instance, our EfficientViT-M5 sur- passes MobileNetV3-Large by 1.9% in accuracy, while get- ting 40.4% and 45.2% higher throughput on Nvidia V100 GPU and Intel Xeon CPU, respectively. Compared to the recent efficient model MobileViT-XXS, EfficientViT-M2 achieves 1.8% superior accuracy, while running 5.8×/3.7× faster on the GPU/CPU, and 7.4× faster when converted to ONNX format. Code and models are available at here.	1. Introduction Vision Transformers (ViTs) have taken computer vision domain by storm due to their high model capabilities and superior performance [18, 44, 69]. However, the constantly improved accuracy comes at the cost of increasing model sizes and computation overhead. For example, SwinV2 [43] uses 3.0B parameters, while V-MoE [62] taking 14.7B pa- rameters, to achieve state-of-the-art performance on Ima- ∗Work done when Xinyu was an intern of Microsoft Research. Figure 1. Speed and accuracy comparisons between EfficientViT (Ours) and other efficient CNN and ViT models tested on an Nvidia V100 GPU with ImageNet-1K dataset [17]. geNet [17]. Such large model sizes and the accompanying heavy computational costs make these models unsuitable for applications with real-time requirements [40, 78, 86]. There are several recent works designing light and effi- cient vision transformer models [9,19,29,49,50,56,79,81]. Unfortunately, most of these methods aim to reduce model parameters or Flops, which are indirect metrics for speed and do not reflect the actual inference throughput of models. For example, MobileViT-XS [50] using 700M Flops runs much slower than DeiT-T [69] with 1,220M Flops on an Nvidia V100 GPU. Although these methods have achieved good performance with fewer Flops or parameters, many of them do not show significant wall-clock speedup against standard isomorphic or hierarchical transformers, e.g., DeiT [69] and Swin [44], and have not gained wide adoption. To address this issue, in this paper, we explore how to go faster with vision transformers, seeking to find princi- ples for designing efficient transformer architectures. Based on the prevailing vision transformers DeiT [69] and Swin [44], we systematically analyze three main factors that af- fect model inference speed, including memory access, com- putation redundancy, and parameter usage. In particular, we find that the speed of transformer models is commonly memory-bound. In other words, memory accessing de- lay prohibits the full utilization of the computing power in GPU/CPUs [21, 32, 72], leading to a critically negative impact on the runtime speed of transformers [15, 31]. 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. 14420 most memory-inefficient operations are the frequent tensor reshaping and element-wise functions in multi-head self- attention (MHSA). We observe that through an appropri- ate adjustment of the ratio between MHSA and FFN (feed- forward network) layers, the memory access time can be re- duced significantly without compromising the performance. Moreover, we find that some attention heads tend to learn similar linear projections, resulting in redundancy in atten- tion maps. The analysis shows that explicitly decomposing the computation of each head by feeding them with divers
Lattas_FitMe_Deep_Photorealistic_3D_Morphable_Model_Avatars_CVPR_2023	Abstract In this paper, we introduce FitMe, a facial reflectance model and a differentiable rendering optimization pipeline, that can be used to acquire high-fidelity renderable hu- man avatars from single or multiple images. The model consists of a multi-modal style-based generator, that cap- tures facial appearance in terms of diffuse and specular re- flectance, and a PCA-based shape model. We employ a fast differentiable rendering process that can be used in an op- timization pipeline, while also achieving photorealistic fa- cial shading. Our optimization process accurately captures both the facial reflectance and shape in high-detail, by ex- ploiting the expressivity of the style-based latent represen- tation and of our shape model. FitMe achieves state-of-the- art reflectance acquisition and identity preservation on sin- gle “in-the-wild” facial images, while it produces impres- sive scan-like results, when given multiple unconstrained facial images pertaining to the same identity. In contrast with recent implicit avatar reconstructions, FitMe requires only one minute and produces relightable mesh and texture- based avatars, that can be used by end-user applications.	1. Introduction Despite the tremendous steps forward witnessed in the last decade, 3D facial reconstruction from a single uncon- strained image remains an important research problem with an active presence in the computer vision community. Its applications are now wide-ranging, including but not lim- ited to human digitization for virtual and augmented real- ity applications, social media and gaming, synthetic dataset creation, and health applications. However, recent works come short of accurately reconstructing the identity of dif- ferent subjects and usually fail to produce assets that can be used for photorealistic rendering. This can be attributed to the lack of diverse and big datasets of scanned human geom- etry and reflectance, the limited and ambiguous information available on a single facial image, and the limitations of the current statistical and machine learning methods. 3D Morhpable Models (3DMM) [18] have been a stan- dard method of facial shape and appearance acquisition from a single “in-the-wild” image. The seminal 3DMM work in [7] used Principal Component Analysis (PCA), to model facial shape and appearance with variable identity and expression, learned from about 200 subjects. 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. 8629 Rendering-Based GAN Inversion Rendering-Based GAN Tuning Final Rendering Trainable Frozen Renderer Figure 2. FitMe method overview. For a target image I0, we optimize the latent vector Wof the generator G, and the shape identity ps, expression pe, camera pcand illumination plparameters, by combining 3DMM fitting and GAN inversion methods, through accurate differentiable diffuse UDand specular USrendering R. Then, from the optimized Wp, we tune the generator Gweights, through the same rendering process. The reconstructed shape Sand facial reflectance (diffuse albedo AD, specular albedo ASand normals NS), achieve great identity similarity and can be directly used in typical renderers, to achieve photorealistic rendering, as shown on the right-hand side. then, larger models have been introduced, .i.e. the LSFM [10], Basel Face Model [52] and Facescape [71], with thou- sands of subjects. Moreover, recent works have introduced 3DMMs of complete human heads [43, 54, 55] or other fa- cial parts such as ears [54] and tongue [53]. Finally, recent works have introduced extensions ranging from non-linear models [49, 66, 67] to directly regressing 3DMM parame- ters [61,68]. However, such models cannot produce textures capable of photorealistic rendering. During the last decade we have seen considerable im- provements in deep generative models. Generative Adver- sarial Networks (GANs) [30], and specifically progressive GAN architectures [34] have achieved tremendous results in learning distributions of high-resolution 2D images of human faces. Recently, style-based progressive generative networks [35–38] are able to learn meaningful latent spaces, that can be traversed in order to reconstruct and manipulate different attributes of the generated samples. Some methods have also been shown effective in learning a 2D representa- tion of 3D facial attributes, such as UV maps [21,23,24,45]. 3D facial meshes generated by 3DMMs can be utilized in rendering functions, in order to create 2D facial images. Differentiating the rendering process is also required in or- der to perform iterative optimization. Recent advances in differentiable rasterization [44], photorealitic facial shad- ing [41] and rendering libraries [20,25,56], enable the pho- torealistic differentiable rendering of such assets. Unfor- tunately, 3DMM [10, 23, 45] works rely on the lambertian shading model which comes short of capturing the com- plexity of facial reflectance. The issue being, photorealistic facial rendering requires various facial reflectance param- eters instead of a single RGB texture [41]. Such datasets are scarce, small and difficult to capture [27,46,57], despite recent attempts to simplify such setups [39]. Several recent approaches have achieved either high- fidelity facial reconstructions [6,23,45] or relightable facialreflectance reconstructions [16,17,19,40,41,65], including infra-red [47], however, high-fidelity and relightable recon- struction still remains elusive. Moreover, powerful mod- els have been shown to capture facial appearance with deep models [22, 42], but they fail to show single or multi image reconstructions. A recent alternative paradigm uses implicit representations to capture avatar appearance and geometry [11, 69], the rendering of which depends on a learned neu- ral rendering. Despite their impressive results, such implicit representations cannot be used by common renderers and are not usually relightable. Finally, the recently introduced Albedo Morphable Model (AlbedoMM) [65] captures fa- cial reflectance and shape with a linear PCA model, but per- vertex color and normal reconstruction is too low-resolution for photorealistic rendering. AvatarMe++ [40, 41] recon- structs high-resolution facial reflectance texture maps from a single “in-the-wild” image, however, its 3-step process (reconstruction, upsampling, reflectance), cannot be opti- mized directly with the input image. In this work, we introduce FitMe , a fully renderable 3DMM with high-resolution facial reflectance texture maps, which can be fit on unconstrained facial images using ac- curate differentiable renderings. FitMe achieves identity similarity and high-detailed, fully renderable reconstruc- tions, which are directly usable by off-the-shelf rendering applications. The texture model is designed as a multi- modal style-based progressive generator, which concur- rently generates the facial diffuse-albedo ,specular-albedo andsurface-normals . A meticulously designed branched discriminator enables smooth training with modalities of different statistics. To train the model we create a capture- quality facial reflectance dataset of 5k subjects, by fine- tuning AvatarMe++ on the MimicMe [50] public dataset, which we also augment in order to balance skin-tone rep- resentation. For the shape, we use interchangeably a face and head PCA model [54], both trained on large-scale ge- 8630 ometry datasets. We design a single or multi-image fitting method, based on style-based generator projection [35] and 3DMM fitting. To perform efficient iterative fitting (in un- der 1 minute), the rendering function needs to be differen- tiable and fast, which makes models such as path tracing un- usable. Prior works in the field [10,12,23] use simpler shad- ing models (e.g. Lambertian), or much slower optimiza- tion [16]. We add a more photorealistic shading than prior work, with plausible diffuse and specular rendering, which can acquire shape and reflectance capable of photorealistic rendering in standard rendering engines (Fig. 1). The flexi- bility of the generator’s extended latent space and the pho- torealistic fitting, enables FitMe to reconstruct high-fidelity facial reflectance and achieve impressive identity similar- ity, while accurately capturing details in diffuse, specular albedo and normals. Overall, in this work we present: • the first 3DMM capable of generating high-resolution facial reflectance and shape, with an increasing level of detail, that can be photorealistically rendered, • the first branched multi-modal style-based progressive generator of high-resolution 3D facial assets (diffuse albedo, specular albedo and normals), and a suitable multi-modal branched discriminator, • a method to acquire and augment a vast facial re- flectance dataset of, using assets from a public dataset, • a multi-modal generator projection, optimized with diffuse and specular differentiable rendering.
Lin_ERM-KTP_Knowledge-Level_Machine_Unlearning_via_Knowledge_Transfer_CVPR_2023	Abstract Machine unlearning can fortify the privacy and security of machine learning applications. Unfortunately, the ex- act unlearning approaches are inefficient, and the approxi- mate unlearning approaches are unsuitable for complicated CNNs. Moreover, the approximate approaches have serious security flaws because even unlearning completely different data points can produce the same contribution estimation as unlearning the target data points. To address the above problems, we try to define machine unlearning from the knowledge perspective, and we propose a knowledge-level machine unlearning method, namely ERM-KTP . Specifi- cally, we propose an entanglement-reduced mask (ERM) structure to reduce the knowledge entanglement among classes during the training phase. When receiving the un- learning requests, we transfer the knowledge of the non- target data points from the original model to the unlearned model and meanwhile prohibit the knowledge of the tar- get data points via our proposed knowledge transfer and prohibition (KTP) method. Finally, we will get the un- learned model as the result and delete the original model to accomplish the unlearning process. Especially, our pro- posed ERM-KTP is an interpretable unlearning method be- cause the ERM structure and the crafted masks in KTP can explicitly explain the operation and the effect of un- learning data points. Extensive experiments demonstrate the effectiveness, efficiency, high fidelity, and scalability of the ERM-KTP unlearning method. Code is available at https://github.com/RUIYUN-ML/ERM-KTP	1. Introduction In recent years, many countries have raised concerns about protecting personal privacy. The privacy legislation, e.g., the well-known European Union’s GDPR [19], have been promulgated to oblige information service providers *Both authors contributed equally †Corresponding authorto remove personal data when receiving a request from the data owner, i.e., the right-to-be-forgotten . Besides, the GDPR stipulates that the service providers should remove the corresponding impact of the data requested by the data owner, in which the machine learning models are the most representative. Many kinds of research demonstrate that the machine learning models can memorize knowledge of the data points, e.g., the membership inference attack [7,14,15] can infer whether a data point is in the training set or not. A naive approach is retraining the model after remov- ing the target data points from the training set, but the hu- man resources and materials consumed are costly. Thus, aiming to efficiently remove data as well as their gener- ated contribution to the model, a new ML privacy protec- tion research direction emerged, called machine unlearning . A good deal of related work attempted to solve the data- removing challenge of inefficient retraining, and there are two representative research fields, including exact unlearn- ing[1] and approximate unlearning [3–5]. Unfortunately, these approaches usually require huge computational, stor- age overhead, and memory requirements for class-specific machine unlearning tasks on the complicated convolutional neural networks (CNNs). Furthermore, they may compro- mise the model’s performance and even cause disastrous forgetting. Most significantly, Thudi et al. [18] believed that such approximate unlearning approaches have serious secu- rity flaws because they usually define machine unlearning as the distribution difference between the unlearned model and the retrained model. According to this definition, even unlearning completely different data points can produce the same contribution estimation as unlearning the target data points. Unlearning algorithms are difficult to implement on deep learning models because these models are often seen as a black box and lack interpretability, which makes data points’ contributions challenging to estimate. Though many related works [2, 10, 16, 21] attempted to improve the inter- pretability of deep learning models, they can not apply to machine unlearning directly. For example, Zhou et al. [22] 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. 20147 leveraged the global average pooling (GAP) in CNNs to generate a class activation mapping (CAM) to indicate the discriminative image regions used by the CNNs to identify the class. Liang et al. [12] proposed class-specific filters to transform
Li_Lift3D_Synthesize_3D_Training_Data_by_Lifting_2D_GAN_to_CVPR_2023	Abstract This work explores the use of 3D generative models to synthesize training data for 3D vision tasks. The key require- ments of the generative models are that the generated data should be photorealistic to match the real-world scenar- ios, and the corresponding 3D attributes should be aligned with given sampling labels. However, we find that the recent NeRF-based 3D GANs hardly meet the above requirements due to their designed generation pipeline and the lack of explicit 3D supervision. In this work, we propose Lift3D, an inverted 2D-to-3D generation framework to achieve the data generation objectives. Lift3D has several merits compared to prior methods: (1) Unlike previous 3D GANs that the output resolution is fixed after training, Lift3D can generalize to any camera intrinsic with higher resolution and photorealistic output. (2) By lifting well-disentangled 2D GAN to 3D object NeRF , Lift3D provides explicit 3D information of generated objects, thus offering accurate 3D annotations for down- stream tasks. We evaluate the effectiveness of our framework by augmenting autonomous driving datasets. Experimental results demonstrate that our data generation framework can effectively improve the performance of 3D object detectors. Code: len-li.github.io/lift3d-web	1. Introduction It is well known that the training of current deep learning models requires a large amount of labeled data. However, col- lecting and labeling the training data is often expensive and time-consuming. This problem is especially critical when the data is hard to annotate. For example, it is difficult for humans to annotate 3D bounding boxes using a 2D image due to the inherent ill-posed 3D-2D projection (3D bounding boxes are usually annotated using LiDAR point clouds). To alleviate this problem, a promising direction is to use synthetic data to train our models. For example, data genera- tion that can be conveniently performed using 3D graphics *Work done during an internship at NIO Autonomous Driving. †Corresponding author. Noise z ∈𝑍𝑍 Pose p∈𝑃𝑃Low Reso.High Reso. Image 2D Upsample Pose p∈𝑃𝑃2D 3D3D 2D a: Previous 3D GANs b: Our approach 2D GAN Noise z ∈𝑍𝑍 NeRFNeRFDownstream Tasks (3D)Add novel objectsGoal: Synthesize downstream task dataset Noise z ∈𝑍𝑍 Pose p∈𝑃𝑃Train 3D Detectors GAN Downstream Tasks (3D)Figure 1. Our goal is to generate novel objects and use them to augment existing datasets. (a)Previous 3D GANs ( e.g., [28, 45]) rely on a 2D upsampler to ease the training of 3D generative ra- diance field, while struggle a trade-off between high-resolution synthesis and 3D consistency. (b)Our 2D-to-3D lifting process disentangles the 3D generation from generative image synthesis, leading to arbitrary rendering resolution and object pose sampling for downstream tasks. engines offers incredible convenience for visual perception tasks. Several such simulated datasets have been created in recent years [3, 14, 30, 31, 42]. These datasets have been used successfully to train networks for perception tasks such as semantic segmentation and object detection. However, these datasets are expensive to generate, requiring specialists to model specific objects and environments in detail. Such datasets also tend to have a large domain gap from real-world ones. With the development of Generative Adversarial Net- works (GAN) [18], researchers have paid increasing atten- tion to utilize GANs to replace graphics engines for syn- thesizing training data. For example, BigDatasetGAN [23] utilizes condition GAN to generate classification datasets via 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. 332 conditioning the generation process on the category labels. SSOD [27] designs a GAN-based generator to synthesize images with 2D bounding boxes annotation for the object detection task. In this paper, we explore the use of 3D GANs to synthesize datasets with 3D-related annotations, which is valuable but rarely explored. Neural radiance field (NeRF) [26] based 3D GANs [6,28], which display photorealistic synthesis and 3D controllable property, is a natural choice to synthesize 3D-related train- ing data. However, our experimental results show that they struggle to keep high-resolution outputs and geometry- consistency results by relying on a 2D upsampler. Further- more, the generated images are not well aligned with the given 3D pose, due to the lack of explicit 3D consistency reg- ularization. This misalignment would introduce large label noise in the dataset, limiting the performance in downstream tasks. In addition to these findings, the camera parameters are fixed after training, making them challenging to align the output resolution with arbitrary downstream data. In this paper, we propose Lift3D, a new paradigm for syn- thesizing 3D training data by lifting pretrained 2D GAN to 3D generative radiance field. Compared with the 3D GANs that rely on a 2D upsampler, we invert the gener- ation pipeline into 2D-to-3D rather than 3D-to-2D to achieve higher-resolution synthesis. As depicted in Fig. 1, we first take advantage of a well-disentangled 2D GAN to generate multi-view images with corresponding pseudo pose annota- tion. The multi-view images are then lifted to 3D represen- tation with NeRF reconstruction. In particular, by distilling from pretrained 2D GAN, lift3D achieves high-quality syn- thesis that is comparable to SOTA 2D generative models. By decoupling the 3D generation from generative image synthe- sis, Lift3D can generate images that are tightly aligned with the sampling label. Finally, getting rid of 2D upsamplers, Lift3D can synthesize images in any resolution by accumu- lating single-ray evaluation. With these properties, we can leverage the generated objects to augment existing dataset with enhanced quantity and diversity. To validate the effectiveness of our data generation frame- work, we conduct experiments on image-based 3D object detection tasks with KITTI [16] and nuScenes [4] datasets. Our framework outperforms the best prior data augmen- tation method [24] with significantly better 3D detection accuracy. Furthermore, even without any labeled data, it achieves promising results in an unsupervised manner. Our contributions are summarized as follows: •We provide the first exploration of using 3D GAN to synthesize 3D training data, which opens up a new possibility that adapts NeRF’s powerful capabilities of novel view synthesis to benefit downstream tasks in 3D. •To synthesize datasets with high-resolution images and accurate 3D labels, we propose Lift3D, an inverted 2D- GIRAFFE HD Ours Position Rotation Position RotationFigure 2. We compare our generation result with GIRAFFE HD [45]. We zoom in or rotate the sampled 3D box to control the generation of models. The rotation of the 3D box introduces artifacts to images generated by GIRAFFE HD. All images are plotted with sampled 3D bounding boxes. to-3D data generation framework that disentangles 3D generation from generative image synthesis. •Our experimental results demonstrate that the synthe- sized training data can improve image-based 3D detec- tors across different settings and datasets.
Lan_Self-Supervised_Geometry-Aware_Encoder_for_Style-Based_3D_GAN_Inversion_CVPR_2023	Abstract StyleGAN has achieved great progress in 2D face recon- struction and semantic editing via image inversion and la- tent editing. While studies over extending 2D StyleGAN to 3D faces have emerged, a corresponding generic 3D GAN inversion framework is still missing, limiting the applica- tions of 3D face reconstruction and semantic editing. In this paper, we study the challenging problem of 3D GAN inversion where a latent code is predicted given a single face image to faithfully recover its 3D shapes and detailed textures. The problem is ill-posed: innumerable composi- tions of shape and texture could be rendered to the cur- rent image. Furthermore, with the limited capacity of a global latent code, 2D inversion methods cannot preserve faithful shape and texture at the same time when applied to 3D models. To solve this problem, we devise an effective self-training scheme to constrain the learning of inversion. The learning is done efficiently without any real-world 2D- 3D training pairs but proxy samples generated from a 3D GAN. In addition, apart from a global latent code that cap- tures the coarse shape and texture information, we augment the generation network with a local branch, where pixel- aligned features are added to faithfully reconstruct face de- tails. We further consider a new pipeline to perform 3D view-consistent editing. Extensive experiments show that our method outperforms state-of-the-art inversion methods in both shape and texture reconstruction quality.	1. Introduction This work aims to devise an effective approach for encoder-based 3D Generative Adversarial Network (GAN) inversion. In particular, we focus on the reconstruction of 3D face, requiring just a single 2D face image as the input. In the inversion process, we wish to map a given image to the latent space and obtain an editable latent code with an encoder. The latent code will be further fed to a generator to reconstruct the corresponding 3D shape with high-quality shape and texture. Further to the learning of an inversion encoder, we also wish to develop an approach to synthesize 3D view-consistent editing results, e.g., changing a neutral expression to smiling, by altering the estimated latent code. GAN inversion [50] has been extensively studied for 2D images but remains underexplored in the 3D world. Inver-sion can be achieved via optimization [1,2,41], which typi- cally provides a precise image-to-latent mapping but can be time-consuming, or encoder-based techniques [40, 47, 49], which explicitly learn an encoding network that maps an image into the latent space. Encoder-based techniques en- joy faster inversion, but the mapping is typically inferior to optimization. In this study, we extend the notion of encoder- based inversion from 2D images to 3D shapes. Adding the additional dimension makes inversion more challenging beyond the goal of reconstructing an editable shape with detail preservation. In particular, 1)Recovering 3D shapes from 2D images is an ill-posed problem, where innumerable compositions of shape and texture could gen- erate identical rendering results. 3D supervisions are cru- cial to alleviate the ambiguity of shape inversion from im- ages. Though high-quality 2D datasets are easily accessi- ble, owing to the expensive cost of scans there is currently a lack of large-scale labeled 3D datasets. 2)The global latent code, due to its compact and low-dimensional nature, only captures the coarse shape and texture information. With- out high-frequency spatial details, we cannot generate high- fidelity outputs. 3)Compared with 2D inversion methods where the editing view mostly aligns with the input view , in 3D editing we expect the editing results to perform well over the novel views with large pose variations. There- fore, 3D GAN inversion is non-trivial task and could not be achieved by directly applying existing approaches. To this end, we propose a novel Encoder-based 3D G AN invErsion framework, E3DGE, which addresses the afore- mentioned three challenges. Our framework has three novel components with a delicate model design. Specifically: Learning Inversion with Self-supervised Learning - The first component focuses on the training of the inversion en- coder. To address the shape collapse of single-view 3D re- construction without external 3D datasets, we retrofit the generator of a 3D GAN model to provide us with diverse pseudo training samples, which can then be used to train our inversion encoder in a self-supervised manner. Specif- ically, we generate 3D shapes from the latent space Wof a 3D GAN, and then render diverse 2D views from each 3D shape given different camera poses. In this way, we can generate many pseudo 2D-3D pairs together with the corre- sponding latent codes. Since the pseudo pairs are generated from a smooth latent space that learns to approximate a nat- 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. 20940 ural shape manifold, they serve as effective surrogate data to train the encoder, avoiding potential shape collapse. Local Features for High-Fidelity Inversion - The second component learns to reconstruct accurate texture details. Our novelty here is to leverage local features to enhance the representation capacity, beyond just the global latent code generated by the inversion encoder. Specifically, in addi- tion to inferring an editable global latent code to represent the overall shape of the face, we further devise an hour-glass model to extract local features over the residuals details that the global latent code fails to capture. The local features, with proper projection to the 3D space, serve as conditions to modulate the 2D image rendering. Through this effective learning scheme, we marry the benefits of both global and local priors and achieve high-fidelity reconstruction. Synthesizing View-consistent Edited Output - The third component addresses the problem of novel view synthesis, a problem unique to 3D shape editing. Specifically, though we achieve high-fidelity reconstruction through aforemen- tioned designs, the local residual features may not fully align with the scene when being semantically edited. More- over, the occlusion issue further degrades the fusion perfor- mance when rendering from novel views with large pose variations. To this end, we propose a 2D-3D hybrid align- ment module for high-quality editing. Specifically, a 2D alignment module and a 3D projection scheme are intro- duced to jointly align the local features with edited images and inpaint occluded local features in novel view synthesis. Extensive experiments show that our method achieves 3D GAN inversion with plausible shapes and high-fidelity image reconstruction without affecting editability. Owing to the self-supervised training strategy with delicate global- local design, our approach performs well on real-world 2D and 3D benchmarks without resorting to any real-world 3D dataset for training. To summarize, our main contributions are as follows: • We propose an early attempt at learning an encoder- based 3D GAN inversion framework for high-quality shape and texture inversion. We show that, with care- ful design, samples synthesized by a GAN could serve as proxy data for self-supervised training in inversion. • We present an effective framework that uses local fea- tures to complement the global latent code for high- fidelity inversion. • We propose an effective approach to synthesize view- consistent output with a 2D-3D hybrid alignment.
Li_Ego-Body_Pose_Estimation_via_Ego-Head_Pose_Estimation_CVPR_2023	Abstract Estimating 3D human motion from an egocentric video sequence plays a critical role in human behavior understand- ing and has various applications in VR/AR. However, naively learning a mapping between egocentric videos and human motions is challenging, because the user’s body is often un- observed by the front-facing camera placed on the head of the user. In addition, collecting large-scale, high-quality datasets with paired egocentric videos and 3D human mo- tions requires accurate motion capture devices, which often limit the variety of scenes in the videos to lab-like environ- ments. To eliminate the need for paired egocentric video and human motions, we propose a new method, Ego-Body Pose Estimation via Ego-Head Pose Estimation (EgoEgo), which decomposes the problem into two stages, connected by the head motion as an intermediate representation. EgoEgo first integrates SLAM and a learning approach to estimate accu- rate head motion. Subsequently, leveraging the estimated head pose as input, EgoEgo utilizes conditional diffusion †indicates equal contribution.to generate multiple plausible full-body motions. This dis- entanglement of head and body pose eliminates the need for training datasets with paired egocentric videos and 3D human motion, enabling us to leverage large-scale egocen- tric video datasets and motion capture datasets separately. Moreover, for systematic benchmarking, we develop a syn- thetic dataset, AMASS-Replica-Ego-Syn (ARES), with paired egocentric videos and human motion. On both ARES and real data, our EgoEgo model performs significantly better than the current state-of-the-art methods.	1. Introduction Estimating 3D human motion from an egocentric video, which records the environment viewed from the first-person perspective with a front-facing monocular camera, is criti- cal to applications in VR/AR. However, naively learning a mapping between egocentric videos and full-body human motions is challenging for two reasons. First, modeling this complex relationship is difficult; unlike reconstruction motion from third-person videos, the human body is often out of view of an egocentric video. Second, learning this 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. 17142 mapping requires a large-scale, diverse dataset containing paired egocentric videos and the corresponding 3D human poses. Creating such a dataset requires meticulous instru- mentation for data acquisition, and unfortunately, such a dataset does not currently exist. As such, existing works have only worked on small-scale datasets with limited mo- tion and scene diversity [22, 47, 48]. We introduce a generalized and robust method, EgoEgo , to estimate full-body human motions from only egocentric video for diverse scenarios. Our key idea is to use head motion as an intermediate representation to decompose the problem into two stages: head motion estimation from the input egocentric video and full-body motion estimation from the estimated head motion. For most day-to-day activities, humans have an extraordinary ability to stabilize the head such that it aligns with the center of mass of the body [13], which makes head motion an excellent feature for full-body motion estimation. More importantly, the decomposition of our method removes the need to learn from paired egocentric videos and human poses, enabling learning from a combina- tion of large-scale, single-modality datasets (e.g., datasets with egocentric videos or 3D human poses only), which are commonly and readily available. The first stage, estimating the head pose from an ego- centric video, resembles the localization problem. However, directly applying the state-of-the-art monocular SLAM meth- ods [33] yields unsatisfactory results, due to the unknown gravity direction and the scaling difference between the es- timated space and the real 3D world. We propose a hybrid solution that leverages SLAM and learned transformer-based models to achieve significantly more accurate head motion estimation from egocentric video. In the second stage, we generate the full-body motion based on a diffusion model conditioned on the predicted head pose. Finally, to evaluate our method and train other baselines, we build a large-scale synthetic dataset with paired egocentric videos and 3D hu- man motions, which can also be useful for future work on visuomotor skill learning and sim-to-real transfer. Our work makes four main contributions. First, we pro- pose a decomposition paradigm, EgoEgo , to decouple the problem of motion estimation from egocentric video into two stages: ego-head pose estimation, and ego-body pose estimation conditioned on the head pose. The decomposi- tion lets us learn each component separately, eliminating the need for a large-scale dataset with two paired modalities. Second, we develop a hybrid approach for ego-head pose estimation, integrating the results of monocular SLAM and learning. Third, we propose a conditional diffusion model to generate full-body poses conditioned on the head pose. Finally, we contribute a large-scale synthetic dataset with both egocentric videos and 3D human motions as a test bed to benchmark different approaches and showcase that our method outperforms the baselines by a large margin.
Lee_Learning_Geometry-Aware_Representations_by_Sketching_CVPR_2023	Abstract Understanding geometric concepts, such as distance and shape, is essential for understanding the real world and also for many vision tasks. To incorporate such information into a visual representation of a scene, we propose learning to represent the scene by sketching, inspired by human be- havior. Our method, coined Learning by Sketching (LBS), learns to convert an image into a set of colored strokes that explicitly incorporate the geometric information of the scene in a single inference step without requiring a sketch dataset. A sketch is then generated from the strokes where CLIP-based perceptual loss maintains a semantic similar- ity between the sketch and the image. We show theoreti- cally that sketching is equivariant with respect to arbitrary affine transformations and thus provably preserves geomet- ric information. Experimental results show that LBS sub- stantially improves the performance of object attribute clas- sification on the unlabeled CLEVR dataset, domain trans- fer between CLEVR and STL-10 datasets, and for diverse downstream tasks, confirming that LBS provides rich geo- metric information.	1. Introduction Since geometric principles form the bedrock of our phys- ical world, many real-world scenarios involve geometric concepts such as position, shape, distance, and orienta- tion. For example, grabbing an object requires estimating its shape and relative distance. Understanding geometric concepts is also essential for numerous vision tasks such as image segmentation, visual reasoning, and pose estima- tion [27]. Thus, it is crucial to learn a visual representation of the image that can preserve such information [76], which we call geometry-aware representation . However, there is still a challenge in learning geometry- aware representations in a compact way that can be useful for various downstream tasks. Previous approaches have focused on capturing geometric features of an image in a 2D grid structure, using methods such as handcrafted fea- ture extraction [2, 4, 14, 31], segmentation maps [21, 57], or Sketch Image Downstream TasksStroke 1Stroke1 1 Stroke n 1Stroke Generator Task Task Task Geometric Information Inference Inference Inference Stroke Representation Figure 1. Overview of LBS, which aims to generate sketches that accurately reflect the geometric information of an image. A sketch consists of a set of strokes represented by a parameterized vec- tor that specifies curve, color, and thickness. We leverage it as a geometry-aware representation for various downstream tasks. convolution features [36, 55]. Although these methods are widely applicable to various domains, they often lack com- pactness based on a high-level understanding of the scene and tend to prioritize nuisance features such as the back- ground. Another line of work proposes architectures that guarantee to preserve geometric structure [12,13,23,58,73] or disentangle prominent factors in the data [8, 28, 39, 51]. Although these methods can represent geometric concepts in a compact latent space, they are typically designed to learn features that are specific to a particular domain and often face challenges in generalizing to other domains [64]. In this study, we present a novel approach to learning geometry-aware representations via sketching . Sketching, which is the process of converting the salient features of an image into a set of color-coded strokes, as illustrated in Figure 1, is the primary means by which humans represent images while preserving their geometry. Our key idea is that sketches can be a compact, high-level representation of an image that accurately reflects geometric information. Sketching requires a high-level understanding of the scene, as it aims to capture the most salient features of the image and abstract them into a limited number of strokes. In ad- dition, a sketch can be represented as a set of parameters by replacing each stroke with parametric curves. Sketch- 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. 23315 Image Sketch Stroke Image Sketch Stroke Original Transformed TransformedImage Sketch Stroke Figure 2. Examples of how sketches capture essential geometric concepts such as shape, size, orientation, curvature, and local distortion. The control points of each stroke compactly represent the local geometric information of the corresponding region in the sketch. The strokes as a whole maintain the geometric structure of the entire image under various transformations in the image domain. ing has also been linked to how people learn geometric concepts [66]. Based on these properties, we directly use strokes as a geometry-aware representation and utilize them for downstream tasks. Under the theoretical framework of geometric deep learning [3], we conduct theoretical analy- sis to validate the effectiveness of strokes as a representation and prove their ability to preserve affine transformations. To validate our hypothesis, we introduce Learning by Sketching (LBS), a method that generates abstract sketches coherent with the geometry of an input image. Our model is distinct from existing sketch generation methods as it does not require a sketch dataset for training, which often has limited abstraction or fails to accurately reflect geometric information. Instead, LBS learns to convert an image into a set of colored Bézier curves that explicitly represent the geometric concepts of the input image. To teach the style of sketching, we use perceptual loss based on the CLIP model [59, 68], which measures both semantic and geo- metric similarities between images and generated sketches. We propose a progressive optimization process that predicts how strokes will be optimized from their initial positions through CLIP-based perceptual loss to generate abstract sketches in a single inference step. As a result, LBS gen- erates a representation that reflects visual understanding in a single inference step, without requiring a sketch dataset. This produces highly explainable representations through sketches, as illustrated in Figure 2. We conduct experimental analyses to evaluate the ef- fectiveness of our approach, learning by sketching, by ad- dressing multiple research questions in various downstream tasks, including: (i)describing the relationships between geometric primitives, (ii)demonstrating simple spatial rea- soning ability by conveying global and local geometric in- formation, (iii) containing general geometric information shared across different domains, and (iv)improving perfor- mance on FG-SBIR, a traditional sketch task.
Li_Class_Balanced_Adaptive_Pseudo_Labeling_for_Federated_Semi-Supervised_Learning_CVPR_2023	Abstract This paper focuses on federated semi-supervised learn- ing (FSSL), assuming that few clients have fully labeled data (labeled clients) and the training datasets in other clients are fully unlabeled (unlabeled clients). Existing methods attempt to deal with the challenges caused by not independent and identically distributed data (Non-IID) set- ting. Though methods such as sub-consensus models have been proposed, they usually adopt standard pseudo label- ing or consistency regularization on unlabeled clients which can be easily influenced by imbalanced class distribution. Thus, problems in FSSL are still yet to be solved. To seek for a fundamental solution to this problem, we present Class Balanced Adaptive Pseudo Labeling (CBAFed), to study FSSL from the perspective of pseudo labeling. In CBAFed, the first key element is a fixed pseudo labeling strategy to handle the catastrophic forgetting problem, where we keep a fixed set by letting pass information of unlabeled data at the beginning of the unlabeled client training in each communication round. The second key element is that we design class balanced adaptive thresholds via consider- ing the empirical distribution of all training data in local clients, to encourage a balanced training process. To make the model reach a better optimum, we further propose a residual weight connection in local supervised training and global model aggregation. Extensive experiments on five datasets demonstrate the superiority of CBAFed. Code will be available at https://github.com/minglllli/ CBAFed .	1. Introduction Federated learning (FL) aims to train machine learning models on a decentralized manner while preserving data privacy, i.e., separate local models are trained on separate local training datasets independently. In recent years, FL has received much attention for privacy protection reasons *Corresponding Author.[32]. However, most of FL works focused on supervised learning with fully labeled data. But in practice, labeling process of large-scale training data is laborious and expen- sive. Due to the lack of funds or experts, large labeled train- ing dataset is difficult for many companies and institutions to obtain. These may hinder applicability of FL. To handle this problem, federated semi-supervised learn- ing (FSSL) has been explored by many researchers recently [8, 14, 15]. There are broadly three lines of FSSL methods by considering different places and status of labeled data. The first two lines consider that there are only limited la- beled data in the central server [8] or each client has par- tially labeled data [15]. The third line assumes that few clients have fully labeled data and the training datasets in other clients are fully unlabeled [14, 18, 31]. Our paper mainly focuses on the third line of FSSL, and we call lo- cal clients with fully labeled data as labeled clients and the other clients as unlabeled clients. The main difficulties to train a third line FSSL model lie in three folds: 1) There are no labeled data in unlabeled clients. Thus, the training can be easily biased without label guidance. 2) Due to the divergent class distribution of labeled and unlabeled clients, namely the not indepen- dent and identically distributed data (Non-IID) setting, in- accurate supervisory signals may be generated in unlabeled clients via employing the model trained in labeled clients by either pseudo labeling or consistency regularization frame- work. 3) Due to the catastrophic forgetting problems in CNNs, with the training process of unlabeled clients going on, models may forget the knowledge learned on labeled clients and so decrease the prediction accuracy drastically. RSCFed [14], the state-of-the-art method significantly boosts the FSSL performance by first distilling sub- consensus models, and then aggregating the sub-consensus models to the global model. The sub-consensus models can handle the Non-IID setting to some extent, but the mean- teacher based consistency regularization framework in un- labeled clients inevitably causes the accuracy degradation when the classes are imbalanced distributed. RSCFed at- 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. 16292 tempts to address the catastrophic forgetting problem by adjusting aggregation weights for labeled and unlabeled clients. But, it seems just alleviates the negative impact on the unlabeled clients and the problem is still yet to be tackled. Besides, the sub-consensus models may occur un- stable training problems, and increase the communication cost. Other methods like FedConsist [31] and FedIRM [18] achieve promising results by utilizing pseudo labeling methods to produce artificial labels, but they do not consider the Non-IID setting among local clients. Pseudo labeling methods are shown to be effective in semi-supervised learning (SSL) [13, 24, 33, 37], which gen- erate pseudo-labels for unlabeled images and propose sev- eral strategies to ensure the quality of pseudo-labels. Ex- isting FSSL works only adopt standard pseudo labeling or consistency regularization on FSSL. Since the mentioned difficulties hamper the usage of these methods, other rem- edy is usually proposed to alleviate the difficulties, which is palliative. Thus, we propose to study FSSL from the perspective of pseudo labeling, seeking for a fundamental solution to this problem. Concretely, we present Class Balanced Adaptive Pseudo Labeling, namely CBAFed, by rethinking standard pseudo labeling methods in SSL. To handle the catastrophic forget- ting problem, we propose a fixed pseudo labeling strategy, which builds a fixed set by letting pass informative unla- beled data and their pseudo labels at the beginning of the unlabeled client training. Due to the Non-IID and heteroge- neous data partition problems in FL, training distribution of unlabeled data can be highly imbalanced, so existing thresh- olds are not suitable in FSSL. We design class balanced adaptive thresholds via considering the empirical distribu- tion of all training data in local clients at the previous com- munication round. Analysis proves that our method sets a reasonably high threshold for extremely scarce classes and encourages a balanced training process. To enhance the learning ability and discover unlabeled data from tail classes, we propose to leverage information from so-called “not informative” unlabeled data. Besides, we also explore a novel training strategy for labeled clients and the central server, termed as residual weight connection, skip connect- ing weights from previous epochs (for labeled clients) or previous global models (for the central server). It can help the model reach better optimum, when the training distribu- tion is imbalanced and training amount is small. We con- duct extensive experiments on five datasets to show the ef- fectiveness of CBAFed. Overall, our main contributions can be summarized as follows: • We present a CBAFed method to deal with the catas- trophic forgetting problems in federated learning . Un- like existing FSSL frameworks that directly adopts pseudo labeling or consistency regularization methods, CBAFed explores a fundamental solution to FSSL viadesigning a novel pseudo labeling method. • We introduce a residual weight connection method, to improve the robustness of the models in labeled clients and the central server, which skip connects weights from previous epoch or communication round to fi- nally reach better optimum. • Experiments are conducted on five datasets: four nat- ural datasets CIFAR-10/100, SVHN, fashion MNIST and one medical dataset ISIC 2018 Skin. CBAFed outperforms state-of-the-art FSSL methods by a large margin, with 11.3% improvements on SVHN dataset.
Lin_Catch_Missing_Details_Image_Reconstruction_With_Frequency_Augmented_Variational_Autoencoder_CVPR_2023	Abstract The popular VQ-VAE models reconstruct images through learning a discrete codebook but suffer from a sig- nificant issue in the rapid quality degradation of image re- construction as the compression rate rises. One major rea- son is that a higher compression rate induces more loss of visual signals on the higher frequency spectrum which re- flect the details on pixel space. In this paper, a Frequency Complement Module (FCM) architecture is proposed to capture the missing frequency information for enhancing reconstruction quality. The FCM can be easily incorpo- rated into the VQ-VAE structure, and we refer to the new model as Frequancy Augmented VAE (FA-VAE ). In ad- dition, a Dynamic Spectrum Loss (DSL) is introduced to guide the FCMs to balance between various frequencies dynamically for optimal reconstruction. FA-VAE is further extended to the text-to-image synthesis task, and a Cross- attention Autoregressive Transformer (CAT) is proposed to obtain more precise semantic attributes in texts. Extensive reconstruction experiments with different compression rates are conducted on several benchmark datasets, and the re- sults demonstrate that the proposed FA-VAE is able to re- store more faithfully the details compared to SOTA meth- ods. CAT also shows improved generation quality with bet- ter image-text semantic alignment.	1. Introduction VQ-V AE models [6,11,25,29,39,46] reconstruct images through learning a discrete codebook of latent embeddings. They gained wide popularity due to the scalable and versa- tile codebook, which can be broadly applied to many visual tasks such as image synthesis [11,49] and inpainting [4,31]. A higher compression rate is typically preferable in VQ- VQ-GAN [11]𝑓:16lpips: 0.30, rFID: 9.49FA-VAE(Ours) 𝑓:16lpips: 0.21, rFID: 5.90FA-VAE (Ours) 𝑓:4lpips: 0.07, rFID: 2.25originalimageimage spectrum Figure 1. Images and their frequency maps. Row 1: original and reconstructed images. Row 2: the frequency maps of images, fre- quency increases in any direction away from the center. fis the compression rate. With a greater compression rate, more details are lost during reconstruction, i.e. eyes and mouth shape, and hair texture (pointed with red arrows) which align with the loss of high-frequency features. All frequency figures in this paper use the same colormap. rFID [14] and lpips [16] are lower the bet- ter, and frequency values increase from red to green, zoom in for better visualization. V AE models since it provides memory efficiency and better learning of coherent semantics structures [11, 40]. One main challenge quickly arises for a higher com- pression rate, which severely compromises reconstruction accuracy. Figure 1 row 1 shows that although the recon- structed images at higher compression rates appear consis- tent with the original image, details inconsistencies such as the color and contour of the lips become apparent upon closer scrutiny. Figure 1 row 2 reveals that similar degra- dation also manifests on the frequency domains where fea- tures towards the middle and higher frequency spectrum are the least recoverable with greater compression rate. Several causes stand behind this gap between pixel and frequency space. The convolutional nature of autoen- coders is prone to spectral bias , which favors learning low- frequency features [22, 36]. This challenge is further ag- 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. 1736 gravated when current methods exclusively design losses or improve model architecture for better semantics resem- blance [11,16,25] but often neglect the alignment on the fre- quency domain [12,15]. On top of that, it is intuitively more challenging for a decoder to reconstruct an image patch from a single codebook embedding (high compression) than multiple embeddings (less compression). The reason is that the former mixes up features of incomplete and diverse fre- quencies, while the latter could preserve more fine-grained and complete features at various frequencies. Inspired by these insights, the Frequency Augmented VAE (FA-VAE )model is proposed, which aims to improve reconstruction quality by achieving better alignment on the frequency spectrums between the original and reconstructed images. More specifically, new modules named Frequency Complement Modules (FCM) are crafted and embedded at multiple layers of FA-V AE’s decoder to learn to comple- ment the decoder’s features with missing frequencies. We observe that valuable middle and high frequencies are mingled with the encoder’s feature maps during the compression via an encoder, shown in Figure 3 row 4. Therefore, a new loss termed Spectrum Loss (SL) is pro- posed to guide FCMs to generate missing features that align with the same level’s encoder features on the fre- quency domain. Since most image semantics reside on the low-frequency spectrum [48], SL prioritizes learning lower- frequency features with diminishing weights as frequencies increase. Interestingly, we discover that checkerboard patterns ap- pear in the complemented decoder’s features with SL, al- though better reconstruction performance is achieved (Fig- ure 3 column 4). We speculate that because SL sets a de- terministic range for the low-frequency spectrum when ap- plying weights on the frequencies without considering that the importance of a frequency can vary from layer to layer. Thus, an improved loss function Dynamic Spectrum Loss (DSL) is crafted on top of SL with a learnable component to adjust the range of low-frequency spectrum dynamically for optimal reconstruction. DSL can improve reconstruction quality even further than SL without the unnatural checker- board artifacts in the features (Figure 3 column 5). We further extend FA-V AE to the text-to-image gener- ation task and propose the Cross-attention Autoregressive Transformer (CAT) model. We first observe that only using one or a few token embeddings is a coarse representation of lengthy texts [8, 27, 33]. Thus CAT uses all token embed- dings as a condition for more precise guidance. Moreover, existing works typically use self-attention, and the text con- dition is embedded merely at the beginning of the genera- tion [11, 49]. This mechanism becomes problematic in the autoregressive generation because one image token is gen- erated at a time, thus the text condition gradually loosens its connection with the generated tokens. To circumvent thisissue, CAT embeds a cross-attention mechanism that allows the text condition to guide each step generation. To summarize, our work includes the following contri- butions: • We propose a new type of architecture called Fre- quency Augmented VAE (FA-VAE) for improving im- age reconstruction through achieving more accurate details reconstruction. • We propose a new loss called Spectrum Loss (SL) and its enhanced version Dynamic Spectrum Loss (D SL), which guides the Frequency Complement Mod- ules (FCM) in FA-V AE to adaptively learn different low/high frequency mixtures for optimal reconstruc- tion. • We propose a new Cross-attention Autoregressive Transformer (CAT) for text-to-image generation using more fine-grained textual embeddings as a condition with a cross-attention mechanism for better image-text semantic alignment.
Li_An_In-Depth_Exploration_of_Person_Re-Identification_and_Gait_Recognition_in_CVPR_2023	Abstract The target of person re-identification (ReID) and gait recognition is consistent, that is to match the target pedes- trian under surveillance cameras. For the cloth-changing problem, video-based ReID is rarely studied due to the lack of a suitable cloth-changing benchmark, and gait recognition is often researched under controlled condi- tions. To tackle this problem, we propose a Cloth-Changing benchmark for Person re-identification and Gait recogni- tion (CCPG). It is a cloth-changing dataset, and there are several highlights in CCPG, (1) it provides 200 identities and over 16K sequences are captured indoors and out- doors, (2) each identity has seven different cloth-changing statuses, which is hardly seen in previous datasets, (3) RGB and silhouettes version data are both available for research purposes. Moreover, aiming to investigate the cloth-changing problem systematically, comprehensive ex- periments are conducted on video-based ReID and gait recognition methods. The experimental results demon- strate the superiority of ReID and gait recognition sepa- rately in different cloth-changing conditions and suggest that gait recognition is a potential solution for addressing the cloth-changing problem. Our dataset will be available athttps://github.com/BNU-IVC/CCPG .	1. Introduction With the rapid development of video devices, more and more surveillance systems are taken into real applications and play an essential role in protecting the security of our society. However, along with the significantly growing *Equal contribution. †Corresponding authors. Person Re-identification Gait RecognitionQuery QueryGallery Gallery Query Person Retrieval Method Gallery Figure 1. The targets of person re-identification and gait recogni- tion are consistent, to search the target person from the gallery. The main difference is that gait recognition normally requires pedestrian segmentation. amount of data, traditional social security is facing two dilemmas: data storage and inefficient monitoring. Due to artificial intelligence progress and diverse demand in so- cial security, many security technologies have come into our sights, such as face recognition, person re-identification (ReID), and gait recognition. These technologies reduce data size vastly and improve efficiency in social security. Video-based ReID plays a vital role in surveillance video analysis, and it intends to match the probe sequence of the specific person from gallery sequences in surveillance sys- tems by learning features of multiple frames [5, 27, 43]. Compared with image-based ReID [10,21,36], video-based ReID provides both appearance and rich temporal informa- tion. Previous video-based ReID methods [9, 12, 19, 20, 33, 36, 38] focus on the cloth-consistent setting, where people do not change their clothes in the short term. However, in the real world, clothes-changing situations are everywhere. Due to its practical application in social security, cloth- 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. 13824 changing ReID has gotten more attention in these years. Another technology is gait recognition. Gait is a unique biometric characteristic that describes the walking patterns of one person [23] and contains spatial statics and temporal dynamics in the walking process. Compared with other bio- metric features, gait is hard to disguise and can work in long distances [13], so it has a potential for social security. Popu- lar gait recognition datasets [28,37] are collected under con- trol conditions, and some real-world benchmarks [41, 44] come into our sights in these years. Many gait recogni- tion methods [4, 7, 17, 18, 24] are mainly developed based on these datasets. Several challenges are studied explicitly, such as carrying bags and cloth-changing. However, the re- search on gait recognition in real scenarios is insufficient. So gait recognition in the wild has attracted increasing at- tention. In real scenarios, the final targets of ReID and gait recog- nition are consistent, that is, to recognize the target person across cameras, as shown in Fig. 1. Previous works [3, 39] conduct experiments to compare the performance of video- based ReID and gait recognition on video-based ReID datasets, which are cloth-unchanged. However, dressing variation in practical applications is a significant problem, especially the finer cloth-changing problem, e.g., changing a whole fit, changing only tops, and changing only pants. Little work specializes in these cloth-changing conditions, and no comparison experiment has been conducted between video-based ReID and gait recognition. The main reason is the lack of such a cloth-changing benchmark for com- parison of cloth-changing conditions. So it is time to build such a cloth-changing dataset for video-based ReID and gait recognition. First of all, we should take three aspects into consideration. (1) RGB data . A benchmark should pro- vide RGB data for comparison between video-based ReID and gait recognition at least, because RGB data is the basic research data for them. (2) Clothes statuses . People usu- ally change their clothes daily, and a dataset should contain different clothing statuses close to daily life. (3) Collec- tion environment . The raw data should not be collected under strict restrictions, which makes the dataset similar to real-world applications as much as possible since our final target is to satisfy social security demands in the real world. To our best knowledge, no such public dataset could sat- isfy all the requirements mentioned above. CCVID [8] only contains the front views of the pedestrians, rather than di- verse views. GREW [44] lacks RGB data and finer clothes changing. Therefore, in this paper, we propose a benchmark, Cloth- Changing benchmark for Person re-identification and Gait recognition ( CCPG ). Especially, the CCPG dataset has sev- eral important features. (1)It contains 200 subjects wearing many different clothes and over 16,000 sequences, and the RGB data is available. (2)Subjects contain seven differentclothes statuses, whose cloth-changing ways include chang- ing a whole fit, changing tops, changing pants, and carrying bags, and these abundant types of dressing are supported for cloth-changing comparison. In addition, more accu- rate tracking results are provided and reduce the noise in the dataset, because precise human detection guarantees the following recognition mission. (3)Our raw data is collected indoors and outdoors, and many challenges are considered, including viewing angles, occlusions, illumination changes, etc., which make it more similar to the real world. Taking advantages of our new cloth-changing bench- mark, we conduct a series of systematic experiments be- tween video-based ReID and gait recognition, aiming to compare their performance in different dressing settings. First, leading video-based ReID methods are performed on the CCPG dataset in different cloth-changing settings, which indicates that they are sensitive to the appearance features and still have room to improve on cloth-changing problems. Second, popular gait recognition methods are implemented on the CCPG dataset and achieve higher per- formance in some cloth-changing settings. In addition, gait recognition methods are as good as ReID in partial cloth-changing situations, e.g., only changing tops and only changing pants. Third, we compare the performance of these pedestrian retrieval methods and analyze experimental results in different cloth-changing settings. The key to ad- dressing the cloth-changing problem is to exploit the cloth- invariant features, and these experimental results demon- strate the SOTA gait recognition methods have more poten- tial for cloth-changing problems. In summary, our main contributions are summarized as follows: • First, we present a brand new cloth-changing bench- mark, named CCPG. Compared with others, our benchmark provides finer cloth-changing conditions. We hope it could help study the cloth-changing prob- lem for video-based ReID and gait recognition. Impor- tantly, it should be emphasized that our data collection obtains permission from authorities and adult volun- teers for research purposes. • Second, with the motivation to study the performance of video-based ReID and gait recognition methods un- der different cloth-changing conditions, comprehen- sive experiments are conducted on them, and the re- sults show the better performance of gait recognition on CCPG, and suggest that gait recognition is a poten- tial solution for solving cloth-changing problems.
Liu_Progressive_Neighbor_Consistency_Mining_for_Correspondence_Pruning_CVPR_2023	Abstract The goal of correspondence pruning is to recognize cor- rect correspondences (inliers) from initial ones, with appli- cations to various feature matching based tasks. Seeking neighbors in the coordinate and feature spaces is a com- mon strategy in many previous methods. However, it is difficult to ensure that these neighbors are always consis- tent, since the distribution of false correspondences is ex- tremely irregular. For addressing this problem, we propose a novel global-graph space to search for consistent neigh- bors based on a weighted global graph that can explicitly explore long-range dependencies among correspondences. On top of that, we progressively construct three neighbor embeddings according to different neighbor search spaces, and design a Neighbor Consistency block to extract neigh- bor context and explore their interactions sequentially. In the end, we develop a Neighbor Consistency Mining Net- work (NCMNet) for accurately recovering camera poses and identifying inliers. Experimental results indicate that our NCMNet achieves a significant performance advantage over state-of-the-art competitors on challenging outdoor and indoor matching scenes. The source code can be found athttps://github.com/xinliu29/NCMNet .	1. Introduction Estimating high-quality feature correspondences be- tween two images is of crucial significance to numerous computer vision tasks, such as visual simultaneous local- ization and mapping (SLAM) [33], structure from mo- tion (SfM) [41, 49], image fusion [31], and image reg- istration [48, 51]. Off-the-shelf feature extraction meth- ods [4, 29, 52] can be employed to establish initial corre- spondences. Due to complex matching situations ( e.g., se- vere viewpoint variations, illumination changes, occlusions, blurs, and repetitive structures), a great number of false cor- respondences, called outliers, are inevitable [19, 30]. To mitigate the negative impact of outliers for downstream tasks, correspondence pruning [5, 24, 53] can be imple- ∗Corresponding Author Input(a) neighbors incoordinatespace 𝑘𝑛𝑛search (b) neighborsin featurespace 𝑘𝑛𝑛search (c) neighbors inglobal-graphspace GlobalGraph𝑘𝑛𝑛searchSGC OutlierInlierSampledinlierSearchregionNeuralNetworkFigure 1. The acquisition process and visual comparison of (a) spatial neighbors, (b) feature-space neighbors, and (c) global- graph neighbors. SGC : the spectral graph convolution operation. mented to further identify correct correspondences, also known as inliers, from initial ones. However, unlike images that contain sufficient information, e.g., texture and RGB information, correspondence pruning is extremely challeng- ing since the spatial positions of initial correspondences are discrete and irregular [55]. Intuitively, inliers commonly conform to consistent con- straints ( e.g., lengths, angles, and motion) under the 2D rigid transformation, while outliers are randomly dis- tributed, see the top left of Fig. 1. Therefore, the con- sistency of correspondences as a vital priori knowledge has been studied extensively to distinguish inliers and out- liers [9, 13, 14], in which the neighbor consistency has re- ceived widespread attention [25, 26, 57]. For well-defined neighbors, previous approaches [5, 8, 27] employ k-nearest neighbor ( knn) search in the coordinate space of raw corre- spondences to seek spatially consistent neighbors, denoted 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. 9527 as spatial neighbors. Subsequently, several works, such as CLNet [57] and MS2DG-Net [11], search for feature- consistent neighbors ( i.e., feature-space neighbors) by per- forming knn search in the feature space learned from the neural network. They devise various strategies to exploit the neighbor consistency of correspondences, and show sat- isfactory progress. Nevertheless, there may exist numerous outliers in the vicinity of an inlier due to the random dis- tribution of outliers, especially for the challenging match- ing scenes. As shown in Fig. 1, in the coordinate and fea- ture spaces, the searched neighbors of a sampled inlier (blue line) always contain some unexpected outliers (red line). To tackle this issue, we propose a new global-graph space to seek consistent neighbors for each correspondence. Inspired by the fact that inliers have strong consistency at a global level [13, 22, 23], we first construct a weighted global graph, in which nodes denote all correspondences and edges represent their pairwise affinities calculated by the preliminary inlier scores. The dependence between two correspondences is determined to be tight if they have high scores simultaneously. Next, we use a spectral graph con- volution operation [21,57] to further explore long-range de- pendencies among correspondences and increase the dis- crimination between inliers and outliers. We finally adopt knn search in the global-graph space to search for globally consistent neighbors, called global-graph neighbors as il- lustrated in Fig. 1(c). Noteworthily, the positions of global- graph neighbors are not required to be spatially close to the sampled inlier. In other words, this kind of neighbor has a large search region (see the ablation for quantitative results) due to our global operation. Moreover, a single type of neighbor is inadequate for all complex matching situations. Therefore, we present a new Neighbor Consistency (NC) block to take full advan- tage of three types of neighbors and improve the robust- ness. Specifically, we progressively construct three neigh- bor embeddings according to the spatial, feature-space, and global-graph neighbors. To extract corresponding neigh- bor context and explore their interactions, we design two successive layers, i.e., Self-Context Extraction (SCE) layer and Cross-Context Interaction(CCI) layer. The SCE layer is responsible for dynamically capturing intra-neighbor re- lations and aggregating their context, while the CCI layer fuses and modulates inter-neighbor interactive information. Finally, an effective Neighbor Consistency Mining Network (NCMNet) is developed to achieve correspondence pruning. Our contributions are three-fold: (1) Based on the fact that inliers have strong consistency at a global level, we pro- pose a novel global-graph space to seek consistent neigh- bors for each correspondence. (2) We present a new NC block to progressively mine the consistency of three types of neighbors by extracting intra-neighbor context and explor- ing inter-neighbor interactions in a sequential manner. (3)We develop an effective NCMNet for correspondence prun- ing, obtaining considerable performance gains when com- pared to state-of-the-art works.
Lin_Neural_Scene_Chronology_CVPR_2023	Abstract In this work, we aim to reconstruct a time-varying 3D model, capable of rendering photo-realistic renderings with independent control of viewpoint, illumination, and time, from Internet photos of large-scale landmarks. The core challenges are twofold. First, different types of temporal changes, such as illumination and changes to the under- lying scene itself (such as replacing one graffiti artwork with another) are entangled together in the imagery. Sec- ond, scene-level temporal changes are often discrete and sporadic over time, rather than continuous. To tackle these problems, we propose a new scene representation equipped with a novel temporal step function encoding method that can model discrete scene-level content changes as piece- wise constant functions over time. Specifically, we represent the scene as a space-time radiance field with a per-image The authors from Zhejiang University are affiliated with the State Key Lab of CAD&CG.∗This work was done when Haotong Lin was in a remote internship at Cornell University.†Corresponding author: Xiaowei Zhou.illumination embedding, where temporally-varying scene changes are encoded using a set of learned step functions. To facilitate our task of chronology reconstruction from In- ternet imagery, we also collect a new dataset of four scenes that exhibit various changes over time. We demonstrate that our method exhibits state-of-the-art view synthesis results on this dataset, while achieving independent control of view- point, time, and illumination. Code and data are available athttps://zju3dv.github.io/NeuSC/ .	1. Introduction If we revisit a space we once knew during our childhood, it might not be as we remembered it. The buildings may have weathered, or have been newly painted, or may have been replaced entirely. Accordingly, there is no such thing as a single, authoritative 3D model of a scene—only a model of how it existed at a given instant in time. For a famous landmark, Internet photos can serve as a kind of chronicle of that landmark’s state over time, if we could organize 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. 20752 information in those photos in a coherent way. For instance, if we could reconstruct a time-varying 3D model, then we could revisit the scene at any desired point in time. In this work, we explore this problem of chronology re- construction , revisiting the work on Scene Chronology from nearly a decade ago [27]. As in that work, we seek to use In- ternet photos to build a 4D model of a scene, from which we can dial in any desired time (within the time interval where we have photos). However, the original Scene Chronol- ogy work was confined to reconstructing planar, rectangular scene elements, leading to limited photo-realism. We can now revisit this problem with powerful neural scene represen- tations, inspired by methods such as NeRF in the Wild [26]. However, recent neural reconstruction methods designed for Internet photos assume that the underlying scene is static, which works well for landmarks with a high degree of per- manence, but fails for other scenes, like New York’s Times Square, that feature more ephemeral elements like billboards and advertisements. However, we find that adapting neural reconstruction methods [26] to the chronology reconstruction problem has many challenges, and that straightforward extensions do not work well. For instance, augmenting a neural radiance field (NeRF) model with an additional time input t, and fit- ting the resulting 4D radiance field to a set of images with timestamps yields temporally oversmoothed models, where different scene appearances over time are blended together, forming ghosted content; such a model underfits the tempo- ral signal. On the other hand, applying standard positional encoding [31] to the time input overfits the temporal sig- nal, conflating transient appearance changes due to factors like illumination with longer-term, sporadic changes to the underlying scene itself. Instead, we seek a model that can disentangle transient, per-image changes from longer-term, scene-level changes, and that allows for independent control of viewpoint, time, and illumination at render-time. Based on the observation that scene-level content changes are often sudden, abrupt “step function”-like changes (e.g., a billboard changing from one advertisement to another), we introduce a novel encod- ing method for time inputs that can effectively model piece- wise constant scene content over time, and pair this method with a per-image illumination code that models transient appearance changes. Accordingly, we represent 4D scene content as a multi-layer perceptron (MLP) that stores density and radiance at each space-time (x, y, z, t )scene point, and takes an illumination code as a side input. The time input tto this MLP is encoded with our proposed step function encoding that models piecewise constant temporal changes. When fit to a set of input images, we find that our representa- tion can effectively factor different kinds of temporal effects, and can produce high-quality renderings of scenes over time. To evaluate our method, we collect a dataset of imagesfrom Flickr and calibrate them using COLMAP, resulting in 52K successfully registered images. These photos are sourced from four different scenes, including dense tourist ar- eas, graffiti meccas, and museums, building upon the datasets used in Scene Chronology. These scenes feature a variety of elements that change over time, including billboards, graffiti art, and banners. Experiments on these scenes show that our method outperforms current state-of-the-art methods and their extensions to space-time view synthesis [6, 26]. We also present a detailed ablation and analysis of our proposed time encoding method. In summary, our work makes the following contributions: •To the best of our knowledge, ours is the first work to achieve photo-realistic chronology reconstruction, allowing for high-quality renderings of scenes with controllable viewpoint, time, and illumination. •We propose a novel encoding method that can model abrupt content changes without overfitting to transient factors. This leads to a fitting procedure that can ef- fectively disentangle illumination effects from content changes in the underlying scene. •We benchmark the task of chronology reconstruction from Internet photos and make our dataset and code available to the research community.
Karim_MED-VT_Multiscale_Encoder-Decoder_Video_Transformer_With_Application_To_Object_Segmentation_CVPR_2023	Abstract Multiscale video transformers have been explored in a wide variety of vision tasks. To date, however, the multiscale processing has been confined to the encoder or decoder alone. We present a unified multiscale encoder-decoder transformer that is focused on dense prediction tasks in videos. Multiscale representation at both encoder and de- coder yields key benefits of implicit extraction of spatiotem- poral features ( i.e. without reliance on input optical flow) as well as temporal consistency at encoding and coarse- to-fine detection for high-level ( e.g. object) semantics to guide precise localization at decoding. Moreover, we pro- pose a transductive learning scheme through many-to-many label propagation to provide temporally consistent predic- tions. We showcase our Multiscale Encoder-Decoder Video Transformer (MED-VT) on Automatic Video Object Seg- mentation (AVOS) and actor/action segmentation, where we outperform state-of-the-art approaches on multiple bench- marks using only raw images, without using optical flow.	1. Introduction Transformers have been applied to a wide range of image and video understanding tasks as well as other areas [13]. The ability of such architectures to establish data relation- ships across space and time without the local biases in- herent in convolutional and other similarly constrained ap- proaches arguably is key to the success. Multiscale process- ing has potential to enhance further the learning abilities of transformers through cross-scale learning [4, 6, 8, 19, 43]. A gap exists, however, as no approach has emerged that makes full use of multiscale processing during both encod- ing and decoding in video transformers. Recent work has focused on multiscale transformer encoding [8,19], yet does not incorporate multiscale processing in the transformer de- coder. Other work has proposed multiscale transformer de- coding [6], yet was designed mainly for single images and VisTRMviTMask2 FormerMEDVT (Ours)Input ClipInput ClipInput ClipInput ClipCoarse FeaturesFeature PyramidFeature PyramidMultiscale Transformer EncoderTransformer EncoderMultiscale Transformer EncoderMultiscale Transformer DecoderTransformer DecoderMultiscale Transformer DecoderTask HeadInductive LearningTransductive Learning Figure 1. Comparison of state-of-the-art multiscale video trans- formers and our approach. Video transformers take an input clip and feed features to a transformer encoder-decoder, with alterna- tive approaches using multiscale processing only in the encoder or decoder. We present a unified multiscale encoder-decoder video transformer (MED-VT), while predicting temporally consistent segmentations through transductive learning. We showcase MED- VT on two tasks, video object and actor/action segmentation. did not consider the structured prediction nature inherent to video tasks, i.e. the importance of temporal consistency. In response, we present the first Multiscale Encoder Decoder Video Transformer (MED-VT). At encoding, its within and between-scale attention mechanisms allow it to capture both spatial, temporal and integrated spatiotemporal information. At decoding, it introduces learnable coarse-to- fine queries that allow for precise target delineation, while enforcing temporal consistency among the predicted masks through transductive learning [38, 52]. We primarily illustrate the utility of MED-VT on the task 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. 6323 of Automatic Video Object Segmentation (A VOS). A VOS separates primary foreground object(s) from background in a video without any supervision, i.e. without information on the objects of interest [53]. This task is important and chal- lenging, as it is a key enabler of many subsequent visually- guided operations, e.g., autonomous driving and augmented reality. A VOS shares challenges common to any VOS task (e.g., object deformation and clutter). Notably, however, the requirement of complete automaticity imposes extra chal- lenges to A VOS, as it does not benefit from any per video initialization. Lacking prior information, solutions must ex- ploit appearance ( e.g., colour and shape) as well as motion to garner as much information as possible. MED-VT responds to these challenges. Its within and between scale attention mechanisms capture both appear- ance and motion information as well as yield temporal con- sistency. Its learnable coarse-to-fine queries allow seman- tically laden information at deeper layers to guide finer scale features for precise object delineation. Its transductive learning through many-to-many label propagation ensures temporally consistent predictions. To showcase our model beyond A VOS, we also apply it to actor/action segmenta- tion. Figure 1 overviews MED-VT compared to others.
Lin_Actionlet-Dependent_Contrastive_Learning_for_Unsupervised_Skeleton-Based_Action_Recognition_CVPR_2023	Abstract The self-supervised pretraining paradigm has achieved great success in skeleton-based action recognition. How- ever, these methods treat the motion and static parts equally, and lack an adaptive design for different parts, which has a negative impact on the accuracy of action recognition. To realize the adaptive action modeling of both parts, we pro- pose an Actionlet-Dependent Contrastive Learning method (ActCLR). The actionlet, defined as the discriminative sub- set of the human skeleton, effectively decomposes motion regions for better action modeling. In detail, by con- trasting with the static anchor without motion, we extract the motion region of the skeleton data, which serves as the actionlet, in an unsupervised manner. Then, center- ing on actionlet, a motion-adaptive data transformation method is built. Different data transformations are ap- plied to actionlet and non-actionlet regions to introduce more diversity while maintaining their own characteristics. Meanwhile, we propose a semantic-aware feature pooling method to build feature representations among motion and static regions in a distinguished manner. Extensive experi- ments on NTU RGB+D and PKUMMD show that the pro- posed method achieves remarkable action recognition per- formance. More visualization and quantitative experiments demonstrate the effectiveness of our method. Our project website is available at https://langlandslin. github.io/projects/ActCLR/	1. Introduction Skeletons represent human joints using 3D coordinate locations. Compared with RGB videos and depth data, skeletons are lightweight, privacy-preserving, and compact to represent human motion. On account of being easier and more discriminative for analysis, skeletons have been widely used in action recognition task [19,23,31,32,46,48]. Supervised skeleton-based action recognition meth- *Corresponding author. This work is supported by the National Natural Science Foundation of China under contract No.62172020. 1Data Actionlet Positive Negative Attract Push AwayFigure 1. Our proposed approach (ActCLR) locates the motion regions as actionlet to guide contrastive learning. ods [3,27,28] have achieved impressive performance. How- ever, their success highly depends on a large amount of la- beled training data, which is expensive to obtain. To get rid of the reliance on full supervision, self-supervised learn- ing [16, 32, 34, 49] has been introduced into skeleton-based action recognition. It adopts a two-stage paradigm, i.e.first applying pretext tasks for unsupervised pretraining and then employing downstream tasks for finetuning. According to learning paradigms, all methods can be classified into two categories: reconstruction-based [14, 32, 41] and contrastive learning-based. Reconstruction-based methods capture the spatial-temporal correlation by pre- dicting masked skeleton data. Zheng et al. [49] first pro- posed reconstructing masked skeletons for long-term global motion dynamics. Besides, the contrastive learning-based methods have shown remarkable potential recently. These methods employ skeleton transformation to generate pos- itive/negative samples. Rao et al. [24] applied Shear and Crop as data augmentation. Guo et al. [8] further proposed to use more augmentations, i.e.rotation, masking, and flip- pling, to improve the consistency of contrastive learning. These contrastive learning works treat different regions of the skeleton sequences uniformly. However, the motion regions contain richer action information and contribute more to action modeling. Therefore, it is sub-optimal to di- rectly apply data transformations to all regions in the previ- ous works, which may degrade the motion-correlated infor- 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. 2363 mation too much. For example, if the mask transformation is applied to the hand joints in the hand raising action, the motion information of the hand raising is totally impaired. It will give rise to the false positive problem, i.e., the seman- tic inconsistency due to the information loss between pos- itive pairs. Thus, it is necessary to adopt a distinguishable design for motion and static regions in the data sequences. To tackle these problems, we propose a new actionlet- dependent contrastive learning method (ActCLR) by treat- ing motion and static regions differently, as shown in Fig. 1. Anactionlet [38] is defined as a conjunctive structure of skeleton joints. It is expected to be highly representative of one action and highly discriminative to distinguish the ac- tion from others. The actionlet in previous works is defined in a supervised way, which relies on action labels and has a gap with the self-supervised pretext tasks. To this end, in the unsupervised learning context, we propose to obtain actionlet by comparing the action sequence with the aver- age motion to guide contrastive learning. In detail, the av- erage motion is defined as the average of all the series in the dataset. Therefore, this average motion is employed as the static anchor without motion. We contrast the action sequence with the average motion to get the area with the largest difference. This region is considered to be the re- gion where the motion takes place, i.e., actionlet. Based on this actionlet, we design a motion-adaptive transformation strategy. The actionlet region is transformed by performing the proposed semantically preserving data transformation. Specifically, we only apply stronger data transformations to non-actionlet regions. With less inter- ference in the motion regions, this motion-adaptive trans- formation strategy makes the model learn better seman- tic consistency and obtain stronger generalization perfor- mance. Similarly, we utilize a semantic-aware feature pool- ing method. By extracting the features in the actionlet re- gion, the features can be more representative of the motion without the interference of the semantics in static regions. We provide thorough experiments and detailed analysis on NTU RGB+D [17, 26] and PKUMMD [18] datasets to prove the superiority of our method. Compared to the state- of-the-art methods, our model achieves remarkable results with self-supervised learning. In summary, our contributions are summarized as fol- lows: • We propose a novel unsupervised actionlet-based con- trastive learning method. Unsupervised actionlets are mined as skeletal regions that are the most discrimina- tive compared with the static anchor, i.e., the average motion of all training data. • A motion-adaptive transformation strategy is designed for contrastive learning. In the actionlet region, we employ semantics-preserving data transformations tolearn semantic consistency. And in non-actionlet re- gions, we apply stronger data transformations to obtain stronger generalization performance. • We utilize semantic-aware feature pooling to extract motion features of the actionlet regions. It makes fea- tures to be more focused on motion joints without be- ing distracted by motionless joints.
Li_Deep_Random_Projector_Accelerated_Deep_Image_Prior_CVPR_2023	Abstract Deep image prior (DIP) has shown great promise in tackling a variety of image restoration (IR) and general vi- sual inverse problems, needing no training data . However, the resulting optimization process is often very slow, in- evitably hindering DIP’s practical usage for time-sensitive scenarios. In this paper, we focus on IR, and propose two crucial modifications to DIP that help achieve sub- stantial speedup: 1) optimizing the DIP seed while freez- ing randomly-initialized network weights, and 2) reduc- ing the network depth. In addition, we reintroduce ex- plicit priors, such as sparse gradient prior—encoded by total-variation regularization, to preserve the DIP peak per- formance. We evaluate the proposed method on three IR tasks, including image denoising, image super-resolution, and image inpainting, against the original DIP and vari- ants, as well as the competing metaDIP that uses meta- learning to learn good initializers with extra data . Our method is a clear winner in obtaining competitive restora- tion quality in a minimal amount of time. Our code is avail- able at https://github.com/sun-umn/Deep- Random-Projector .	1. Introduction Deep image prior as an emerging “prior” for visual inverse problems Deep image prior (DIP) [29] and its variants [9, 21] have recently claimed numerous successes in solving image restoration (IR) (e.g., image denoising, super-resolution, and image inpainting) and general visual inverse problems [22]. The idea is to parameterize a visual object of interest, say x, as the output of a structured deep neural network (DNN), i.e., x=Gθ(z), where Gθis a DNN and zis a seed that is often drawn randomly and then frozen. Now given a visual inverse problem of the form y≈f(x)—yis the observation, fmodels the observation process, and ≈allows modeling observational noise—and the typical maximum-a posterior (MAP)-inspired regular-ized data-fitting formulation ( λ: regularization parameter): min xℓ(y, f(x))\|{z} data-fitting loss+λ R(x)\|{z} regularizer, (1) one can plug in the reparametrization x=Gθ(z)to obtain min θℓ(y, f◦Gθ(z)) +λR◦Gθ(z), (2) where ◦denotes functional composition. A salient feature of DIP for IR is that they are training-free despite the pres- ence of the DNN Gθ, i.e., no extra data other than yand fare needed for problem-solving. DIP is not a standalone prior, unlike traditional priors such as sparse gradient [24], dark channel [8], and self- similarity [7]: favorable results from DIP are obtained as a result of the tight integration of architecture, over- parametrization of Gθ, first-order optimization methods, and appropriate early stopping (ES) . The Gθin prac- tice is often a structured convolutional neural network and significantly overparameterized. Due to the heavy over- parametrization, in principle, f◦Gθ(z)could perfectly fity, especially when no extra regularization R◦Gθ(z) is present. When such overfitting occurs, the recovered bx=Gθ(z)accounts for the potential noise in yalso in ad- dition to the desired visual content. What comes to the res- cue is the hallmark “early-learning-then-overfitting” phe- nomenon: the learning process picks up mostly the desired visual content first before starting to fit the noise [16, 30], which is believed to be a combined implicit regulariza- tion effect of overparametrization, convolutional structures inGθ, and first-order optimization methods [10, 12]. So, if one can locate the peak-performance point and stop the fitting process sharp there, i.e., performing appropriate ES, they could get a good estimate for x. Practical issues: overfitting and slowness Despite the remarkable empirical success of DIP in solving IRs, there 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. 18176 Figure 1. Comparison of DIP and our DRP-DIP for image denoising. Left: immediate reconstructions generated by DIP and DRP-DIP respectively at different time cutoffs. DRP-DIP can restore skeleton of the image in 1second, whereas DIP cannot do so even after 10 seconds. Right : PSNR trajectories over time. Our DRP-DIP reaches the performance peak in much shorter time than DIP. are a couple of pressing issues (see Fig. 1) impeding the practical deployment of DIP: •Overfitting : Most previous works reporting the suc- cesses of DIP set the number of iterations directly, based on visual inspection and trial-and-error. Visual inspec- tion precludes large-scale batch processing or deployment into unfamiliar (e.g., scientific imaging of unknown mi- nuscule objects) or unobservable scenarios (e.g., multi- dimensional objects that are hard to visualize), whereas trial-and-error likely performs the tuning with reference to the unknown groundtruth object. So, strictly speaking, previous successes mostly only show the potential of DIP , without providing an operational ES strategy [16,22,30] . Ideas that try to mitigate overfitting, including controlling the capacity of Gθ, explicit regularization, and explicit noise modeling, tend to prolong the iteration process and push the peak performance to final iterations. •Slowness : For DIP to reach the performance peak from random initialization, it typically takes thousands of steps. Although the number is comparable to that of typ- ical iterative methods for solving Eq. (1), here each step entails a forward and a backward pass through the Gθ and hence is much more expensive. In fact, on a state-of- the-art GPU card such as Nvidia V100, the whole pro- cess can take up to tens of minutes for simple IR and up to hours for advanced IR tasks. The optimization slowness inevitably hinders DIP’s applicability to time- sensitive problems. These two issues are intertwined: mitigating overfitting es- calates the slowness. Thus an ideal solution is to speed up the process of climbing to the performance peak and then stop around the peak . Our focus and contributions Our previous works [16, 30] have proposed effective ES methods for DIP. In this paper, we tackle the slowness issue. We propose an effec- tive and efficient DIP variant, called deep random projec-tor(DRP), that requires substantial less computation time than DIP to obtain a comparable level of peak performance. DRP consists of three judiciously chosen modifications to DIP: (1) optimizing the DIP seed while freezing randomly- initialized network weights, (2) reducing the network depth, and (3) including additional explicit prior(s) for regulariza- tion, such as total variation (TV) regularization that encodes the sparse-gradient prior [24]. One can quickly see the su- periority of our method from Fig. 1. Our main contributions include: • proposing deep random projector (DRP) that integrates three essential modifications to DIP for speedup (Sec. 3); • validating that DRP achieves peak performance compa- rable to that of DIP in much less time on three IR tasks (Sec. 4.2, Sec. 4.3, and Sec. 4.4), and showing that DRP is much more efficient than meta-learning-based meta-DIP for speedup (Sec. 4.6); • demonstrating that our ES method in [30] can be directly integrated, without modification, to detect near-peak per- formance for DRP. Hence, we have a complete solution to both overfitting and slowness issues in DIP (Sec. 4.5).
Li_A_Whac-a-Mole_Dilemma_Shortcuts_Come_in_Multiples_Where_Mitigating_One_CVPR_2023	Abstract Machine learning models have been found to learn shortcuts—unintended decision rules that are unable to generalize—undermining models’ reliability. Previous works address this problem under the tenuous assumption that only a single shortcut exists in the training data. Real-world im- ages are rife with multiple visual cues from background to texture. Key to advancing the reliability of vision systems is understanding whether existing methods can overcome multiple shortcuts or struggle in a Whac-A-Mole game, i.e., where mitigating one shortcut amplifies reliance on others. To address this shortcoming, we propose two benchmarks: 1) UrbanCars, a dataset with precisely controlled spuri- ous cues, and 2) ImageNet-W, an evaluation set based on ImageNet for watermark, a shortcut we discovered affects nearly every modern vision model. Along with texture and background, ImageNet-W allows us to study multiple short- cuts emerging from training on natural images. We find computer vision models, including large foundation models— regardless of training set, architecture, and supervision— struggle when multiple shortcuts are present. Even methods explicitly designed to combat shortcuts struggle in a Whac- A-Mole dilemma. To tackle this challenge, we propose Last Layer Ensemble, a simple-yet-effective method to mitigate multiple shortcuts without Whac-A-Mole behavior. Our re- sults surface multi-shortcut mitigation as an overlooked chal- lenge critical to advancing the reliability of vision systems. The datasets and code are released: https://github. com/facebookresearch/Whac-A-Mole .	1. Introduction Machine learning often achieves good average perfor- mance by exploiting unintended cues in the data [ 26]. For instance, when backgrounds are spuriously correlated with objects, image classifiers learn background as a rule for ob- ject recognition [ 93]. This phenomenon—called “shortcut learning”—at best suggests average metrics overstate model performance and at worst renders predictions unreliable as models are prone to costly mistakes on out-of-distribution (OOD) data where the shortcut is absent. For example, †Work done during the internship at Meta AI.∗Equal Contribution.COVID diagnosis models degraded significantly when spuri- ous visual cues ( e.g., hospital tags) were removed [17]. Most existing works design and evaluate methods under the tenuous assumption that a single shortcut is present in the data [ 33,61,74]. For instance, Waterbirds [ 74], the most widely-used dataset, only benchmarks the mitigation of the background shortcut [ 7,15,59]. While this is a useful simpli- fied setting, real-world images contain multiple visual cues; models learn multiple shortcuts. From ImageNet [ 18,82] to facial attribute classification [ 51] and COVID-19 chest radiographs [ 17], multiple shortcuts are pervasive. Whether existing methods can overcome multiple shortcuts or strug- gle in a Whac-A-Mole game—where mitigating one shortcut amplifies others—remains a critical open question. We directly address this limitation by proposing two datasets to study multi-shortcut learning: UrbanCars and ImageNet-W . In UrbanCars (Fig. 1a), we precisely inject two spurious cues—background and co-occurring object. Ur- banCars allows us to conduct controlled experiments probing multi-shortcut learning in standard training as well as short- cut mitigation methods, including those requiring shortcut labels. In ImageNet-W (IN-W) (Fig. 1b), we surface a new watermark shortcut in the popular ImageNet dataset (IN- 1k). By adding a transparent watermark to IN-1k validation set images, ImageNet-W, as a new test set, reveals vision models ranging from ResNet-50 [ 31] to large foundation models [ 10]universally rely on watermark as a spurious cue for the “carton” class ( cf. cardboard box in Fig. 1b). When a watermark is added, ImageNet top-1 accuracy drops by 10.7% on average across models. Some, such as ResNet-50, suffer a catastrophic 26.7% drop (from 76.1% on IN-1k to 49.4% on IN-W) (Sec. 2.2)). Along with texture [ 27,34] and background [ 93] benchmarks, ImageNet-W allows us to study multiple shortcuts emerging in natural images. We find that across a range of supervised/self-supervised methods, network architectures, foundation models, and shortcut mitigation methods, vision models struggle when multiple shortcuts are present. Benchmarks on UrbanCars and multiple shortcuts in ImageNet (including ImageNet-W) reveal an overlooked challenge in the shortcut learning prob- lem: multi-shortcut mitigation resembles a Whac-A-Mole game, i.e., mitigating one sho
Liu_MarS3D_A_Plug-and-Play_Motion-Aware_Model_for_Semantic_Segmentation_on_Multi-Scan_CVPR_2023	Abstract 3D semantic segmentation on multi-scan large-scale point clouds plays an important role in autonomous sys- tems. Unlike the single-scan-based semantic segmentation task, this task requires distinguishing the motion states of points in addition to their semantic categories. However, methods designed for single-scan-based segmentation tasks perform poorly on the multi-scan task due to the lacking of an effective way to integrate temporal information. We pro- pose MarS3D, a plug-and-play motion-aware module for semantic segmentation on multi-scan 3D point clouds. This module can be flexibly combined with single-scan models to allow them to have multi-scan perception abilities. The model encompasses two key designs: the Cross-Frame Fea- ture Embedding module for enriching representation learn- ing and the Motion-Aware Feature Learning module for en- hancing motion awareness. Extensive experiments show that MarS3D can improve the performance of the base- line model by a large margin. The code is available at https://github.com/CVMI-Lab/MarS3D .	1. Introduction 3D semantic segmentation on multi-scan large-scale point clouds is a fundamental computer vision task that ben- efits many downstream problems in autonomous systems, such as decision-making, motion planning, and 3D recon- struction, to name just a few. Compared with the single- scan semantic segmentation task, this task requires under- standing not only the semantic categories but also the mo- tion states ( e.g., moving or static) of points based on multi- scan point cloud data. In the past few years, extensive research has been con- ducted on single-scan semantic segmentation with signifi- cant research advancements [4, 5, 12, 25, 31, 33, 36]. These approaches are also applied to process multi-scan point *Equal contribution. †Corresponding author. non-movingnon-movingnon-movingcarcarcar carcarcar non-movingmovingmovingBaselineMarS3DGTFigure 1. Comparison of our proposed method, MarS3D, with baseline method using SPVCNN [25] as the backbone on Se- manticKITTI [1] dataset. MarS3D achieves excellent results in the classification of semantic categories and motion states, while the baseline method can not distinguish motion well from static. clouds, wherein multiple point clouds are fused to form a single point cloud before being fed to the network for processing. Albeit simple, this strategy may lose tempo- ral information and make distinguishing motion states a challenging problem. As a result, they perform poorly in classifying the motion states of objects. As shown in Figure 1, the simple point cloud fusion strategy can- not effectively enable the model to distinguish the motion states of cars even with a state-of-the-art backbone network SPVCNN [25]. Recently, there have been some early at- tempts [7, 23, 24, 28] to employ attention modules [24] and recurrent networks [7,23,28] to fuse information across dif- ferent temporal frames. However, these approaches do not perform well on the multi-scan task due to the insufficiency of temporal representations and the limited feature extrac- tion ability of the model. In sum, a systematic investigation of utilizing the rich spatial-temporal information from multiple-point cloud scans is still lacking. This requires answering two critical questions: (1) how can we leverage the multi-scan informa- tion to improve representation learning on point clouds 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. 9372 better semantic understanding? and (2) how can the tem- poral information be effectively extracted and learned for classifying the motion states of objects? In this paper, we propose a simple plug-and-play Motion- awareSegmentation module for 3Dmulti-scan analysis ( MarS3D ), which can seamlessly integrate with existing single-scan semantic segmentation models and en- dow them with the ability to perform accurate multi-scan 3D point cloud semantic segmentation with negligible com- putational costs. Specifically, our method incorporates two core designs: First, to enrich representation learning of multi-frame point clouds, we propose a Cross-Frame Fea- ture Embedding (CFFE) module which embeds time-step information into features to facilitate inter-frame fusion and representation learning. Second, inspired by the observa- tion that objects primarily move along the horizontal ground plane ( i.e.,xy-plane) in large-scale outdoor scenes, i.e., min- imal motion along the z-axis, we propose a Motion-Aware Feature Learning (MAFL) module based on Bird’s Eye View (BEV), which learns the motion patterns of objects between frames to facilitate effectively discriminating the motion states of objects. We extensively evaluate our approach upon several mainstream baseline frameworks on SemanticKITTI [1] and nuScenes [3] dataset. It consistently improves the per- formance of the baseline approaches, e.g., MinkUnet [5], by 6.24% in mIoU on SemanticKITTI with a negligible in- crease in model parameters, i.e., about 0.2%. The main con- tributions are summarized as follows: • We are the first to propose a plug-and-play mod- ule for large-scale multi-scan 3D semantic segmenta- tion, which can be flexibly integrated with mainstream single-scan segmentation models without incurring too much cost. • We devise a Cross-Frame Feature Embedding module to fuse multiple point clouds while preserving their temporal information, thereby enriching representa- tion learning for multi-scan point clouds. • We introduce a BEV-based Motion-Aware Feature Learning module to exploit temporal information and enhance the model’s motion awareness, facilitating the prediction of motion states. • We conduct extensive experiments and comprehensive analyses of our approach with different backbone mod- els. The proposed model performs favorably compared to the baseline methods while introducing negligible extra parameters and inference time.
Kong_Understanding_Masked_Image_Modeling_via_Learning_Occlusion_Invariant_Feature_CVPR_2023	Abstract Recently, Masked Image Modeling (MIM) achieves great success in self-supervised visual recognition. However, as a reconstruction-based framework, it is still an open question to understand how MIM works, since MIM appears very dif- ferent from previous well-studied siamese approaches such as contrastive learning. In this paper, we propose a new viewpoint: MIM implicitly learns occlusion-invariant fea- tures, which is analogous to other siamese methods while the latter learns other invariance. By relaxing MIM formulation into an equivalent siamese form, MIM methods can be in- terpreted in a unified framework with conventional methods, among which only a) data transformations, i.e. what invari- ance to learn, and b) similarity measurements are different. Furthermore, taking MAE [29] as a representative example of MIM, we empirically find the success of MIM models relates a little to the choice of similarity functions, but the learned occlusion invariant feature introduced by masked image – it turns out to be a favored initialization for vision transformers, even though the learned feature could be less semantic. We hope our findings could inspire researchers to develop more powerful self-supervised methods in computer vision community.	1. Introduction Invariance matters in science [33]. In self-supervised learning, invariance is particularly important: since ground truth labels are not provided, one could expect the favored learned feature to be invariant (or more generally, equiv- ariant [17]) to a certain group of transformations on the inputs. Recent years, in visual recognition one of the most successful self-supervised frameworks – contrastive learn- ing[22, 42, 47] – benefits a lot from learning invariance . The key insight of contrastive learning is, because recog- nition results are typically insensitive to the deformations (e.g. cropping, resizing, color jittering) on the input images, †Corresponding author. This work is supported by The National Key Research and Development Program of China (No. 2017YFA0700800) and Beijing Academy of Artificial Intelligence (BAAI).a good feature should also be invariant to the transforma- tions. Therefore, contrastive learning suggests minimizing the distance between two (or more [10]) feature maps from the augmented copies of the same data, which is formulated as follows: min θE x∼DM(z1, z2), z 1=fθ(T1(x)), z 2=fθ(T2(x)), (1) where Dis the data distribution; fθ(·)means the encoder network parameterized by θ;T1(·)andT2(·)are two trans- formations on the input data, which defines what invariance to learn; M(·,·)is the distance function1(orsimilarity mea- surement ) to measure the similarity between two feature maps z1andz2. Clearly, the choices of TandMare es- sential in contrastive learning algorithms. Researchers have come up with a variety of alternatives. For example, for the transformation T, popular methods include random crop- ping [3, 12, 28, 30], color jittering [12], rotation [25, 44], jigsaw puzzle [41], colorization [56] and etc. For the simi- larity measurement M,InfoMax principle [3] (which can be implemented with MINE [7] or InfoNCE loss [12,14,30,42]), feature de-correlation [6, 55], asymmetric teacher [15, 28], triplet loss [36] and etc., are proposed. Apart from contrastive learning, very recently Masked Image Modeling (MIM ,e.g.[5]) quickly becomes a new trend in visual self-supervised learning. Inspired by Masked Language Modeling [18] in Natural Language Processing , MIM learns feature via a form of denoising autoencoder [48]: images which are occluded with random patch masks are fed into the encoder, then the decoder predicts the original embeddings of the masked patches: min θ,ϕE x∼DM(dϕ(z), x⊙(1−M)), z =fθ(x⊙M), (2) where “ ⊙” means element-wise product; Mispatch mask2; fθ(·)anddϕ(·)areencoder anddecoder respectively; zis 1Following the viewpoint in [15], we suppose distance functions could contain parameters which are jointly optimized with Eq. (1). For example, weights in project head [12] or predict head [15, 28] are regarded as a part of distance function M(·). 2So “x⊙M” represents “unmasked patches” and vice versa. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Excep
Lei_A_Hierarchical_Representation_Network_for_Accurate_and_Detailed_Face_Reconstruction_CVPR_2023	Abstract Limited by the nature of the low-dimensional represen- tational capacity of 3DMM, most of the 3DMM-based face reconstruction (FR) methods fail to recover high-frequency facial details, such as wrinkles, dimples, etc. Some attempt to solve the problem by introducing detail maps or non- linear operations, however, the results are still not vivid. To this end, we in this paper present a novel hierarchical repre- sentation network (HRN) to achieve accurate and detailed face reconstruction from a single image. Specifically, we implement the geometry disentanglement and introduce the hierarchical representation to fulfill detailed face modeling. Meanwhile, 3D priors of facial details are incorporated to enhance the accuracy and authenticity of the reconstruction results. We also propose a de-retouching module to achieve better decoupling of the geometry and appearance. It is noteworthy that our framework can be extended to a multi- view fashion by considering detail consistency of different views. Extensive experiments on two single-view and two multi-view FR benchmarks demonstrate that our method outperforms the existing methods in both reconstruction ac- curacy and visual effects. Finally, we introduce a high- quality 3D face dataset FaceHD-100 to boost the research of high-fidelity face reconstruction. The project homepage is at https://younglbw.github.io/HRN-homepage/.	1. Introduction High-fidelity 3D face reconstruction finds a wide range of applications in many scenarios, such as AR/VR, medicaltreatment, film production, etc. While extensive works al- ready achieved excellent reconstruction performance using specialized hardware like LightStage [2, 11, 35], estimating highly detailed face models from single or sparse-view im- ages is still a challenging problem. Based on 3DMM [8], a statistical model learned from a collection of face scans, many works [16, 22, 23, 32] attempt to reconstruct the 3D face from a single image and achieve impressive results. However, limited by the nature of the low dimensional rep- resentational ability of the 3DMM, these methods can not recover the detailed facial geometry. Recently, some methods [13, 24, 38] devote to capturing high-frequency facial details such as wrinkles by predicting a displacement map. They achieve realistic results, how- ever, fail to model the mid-frequency details, such as the detailed contour of the jaw, cheeck, etc. To this end, some works try to capture the overall details by introducing latent encoding of details [19] or non-linear operations [20, 44]. Nevertheless, it is hard to make a trade-off when handling the mid- and high-frequency details simultaneously. An- other challenge is how to obtain accurate shapes and de- tailed 3D facial priors considering multifarious lightings and skins for different images. [10, 13] resort to the wrin- kle statistics computed from 3D face scans to fulfill realis- tic high-frequency details, but still fail to model the mid- frequency details. Based on the observations above, we introduce a hierar- chical representation network (HRN) for accurate and de- tailed face reconstruction from single image, as shown 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. 394 Fig. 2. Firstly, we decouple the facial geometry into low- frequency geometry, mid-frequency (MF) details, and high- frequency (HF) details. Then, in a hierarchical fashion, we model these parts with face-wise blendshape coefficients, vertex-wise deformation map, and pixel-wise displacement map, respectively (shown in Fig. 1). Concretely, we employ two image translation networks [27] to estimate the corre- sponding detail maps (deformation and displacement map), and further employ them to generate the detailed face model in a coarse-to-fine manner. Moreover, we introduce the 3D priors of MF and HF details by fitting face scans with our hierarchical representation to facilitate accurate and faith- ful modeling. Inspired by [33], we propose a de-retouching module to adaptively refine the base texture to overcome the ambiguities between skin blemishes and illuminations. Ex- tensive experiments show that our method outperforms the existing methods on two large-scale benchmarks, exhibiting excellent performance in terms of detail capturing and accu- rate shape modeling. Thanks to the detail disentanglement strategy and the guidance of detail priors, we extend HRN to a multi-view fashion and achieve accurate FR from only a few views. Finally, to boost the research of sparse-view and high-fidelity FR, we introduce a high-quality 3D face dataset named FaceHD-100. Our main contributions in this work are as follows: (A)We present a hierarchical modeling strategy and pro- pose a novel framework HRN for single-view FR task. Our HRN produces accurate and highly detailed FR results and outperforms the existing state-of-the-art methods on two large-scale single-view FR benchmarks. (B)We introduce detail priors to guide the faithful modeling of hierarchical details and design a de-retouching module to facilitate the decoupling of geometry and appearance. (C)We extend HRN to a multi-view fashion to form MV- HRN, which enables accurate face modeling from sparse- view images and outperforms the existing methods on two large-scale multi-view FR benchmarks. (D)To boost the research on sparse-view and high-fidelity FR tasks, we introduce a high-quality 3D face dataset FaceHD-100, containing 2,000 detailed 3D face models and corresponding high-definition multi-view images.
Le_GamutMLP_A_Lightweight_MLP_for_Color_Loss_Recovery_CVPR_2023	Abstract Cameras and image-editing software often process im- ages in the wide-gamut ProPhoto color space, encompass- ing 90% of all visible colors. However, when images are encoded for sharing, this color-rich representation is trans- formed and clipped to ﬁt within the small-gamut standard RGB (sRGB) color space, representing only 30% of visi- ble colors. Recovering the lost color information is chal- lenging due to the clipping procedure. Inspired by neu- ral implicit representations for 2D images, we propose a method that optimizes a lightweight multi-layer-perceptron (MLP) model during the gamut reduction step to predict the clipped values. GamutMLP takes approximately 2 sec- onds to optimize and requires only 23 KB of storage. The small memory footprint allows our GamutMLP model to be saved as metadata in the sRGB image—the model can be extracted when needed to restore wide-gamut color values. We demonstrate the effectiveness of our approach for color recovery and compare it with alternative strategies, includ- ing pre-trained DNN-based gamut expansion networks and other implicit neural representation methods. As part of this effort, we introduce a new color gamut dataset of 2200 wide-gamut/small-gamut images for training and testing.	1. Introduction The RGB values of our color images do not represent the entire range of visible colors. The span of visible colors that can be reproduced by a particular color space’s RGB primaries is called a gamut. Currently, the vast majority of color images are encoded using the standard RGB (sRGB) color space [ 7]. The sRGB gamut is capable of reproducing approximately 30% of the visible colors and was optimized for the display hardware of the 1990s. Close to 30 years later, this small-gamut color space still dominates how im- ages are saved, even though modern display hardware is ca- pable of much wider gamuts. Interestingly, most modern DSLR and smartphone cam- eras internally encode images using the ProPhoto color space [ 12]. ProPhoto RGB primaries deﬁne a wide gamut capable of representing 90% of all visible colors [ 33]. 0.01 0.00 Original wide-gamut ProPhoto imagesRGB image with color clipping Conversion back to ProPhotoError map w.r.t. original ProPhoto PSNR: 40.11 PSNR: 54.77(A) Input ProPhoto image and saved sRGB image with MLP embedded (requires 23 KB).ProPhoto gamutsRGB gamut Original ProPhoto RGB values remain clipped inside the sRGB gamut. (B) Current approach converting sRGB back to ProPhoto.(C) Our approach converting sRGB back to ProPhoto using the embedded MLP.Error map w.r.t. original ProPhotoConversion back to ProPhoto with MLP Recovered colors using the embedded MLP to restore clipped wide-gamut RGB values.Clipped color values Predict original colors Lightweight MLP Embedded MLP as metadata in sRGB image𝑥 𝑦 𝑅′ G′ B′෠𝑅 ෠𝐺 ෠𝐵Figure 1. (A) shows a wide-gamut (ProPhoto) image that has been converted and saved as a small-gamut (sRGB) image; color clipping is required to ﬁt the smaller sRGB gamut (as shown in the chromaticity diagrams). (B) Standard color conversion back to the wide-gamut color space is not able to recover the clipped colors. (C) Conversion back to the wide-gamut RGB using our lightweight GamutMLP (23 KB) can recover the clipped color val- ues back to their original values. Image-processing software, such as Adobe Photoshop, also uses this color-rich space to manipulate images, especially when processing camera RAW-DNG ﬁles. By process- ing images in the wide-gamut ProPhoto space, cameras and editing software allow users the option to save an image in other color spaces—such as AdobeRGB, UHD, and Display-P3—that have much wider color gamuts than sRGB. However, these color spaces are still rare, and most images are ultimately saved in sRGB. To convert color val- ues between ProPhoto and sRGB, a gamut reduction step is applied that clips the wide-gamut color values to ﬁt the smaller sRGB color gamut. Once gamut reduction is ap- plied, it is challenging to recover the original wide-gamut values. As a result, when images are converted back to a wide-gamut color space for editing or display, much of the color ﬁdelity is lost, 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. 18268 Optimize MLP to predict original ProPhoto RGB valuesConvert to sRGB (with clipping) Input GT ProPhoto image (wide-gamut) Prediction MLP … … … L2 LosssRGB image (small-gamut) Convert to ProPhoto (values remained clipped) MLP model bitstream 01101011101110111011101110 11100011010100011101101010 1100110.. 1 2 34 Target ProPhoto RGB values Input spatial coordinate ( x,y) and "clipped" ProPhoto RGB values ( R', G', B' ) Clipped (wide-gamut) ProPhoto Embed MLP in sRGB image as metadata (23 KB) Clipped values shown in "red" (𝑅, 𝐺, 𝐵) (𝑥, 𝑦, 𝑅′, 𝐺′, 𝐵′)Figure 2. An overview of the gamut reduction stage in our framework. This phase shows the gamut reduction step, where the wide-gamut ProPhoto is converted to the small-gamut sRGB. While saving the sRGB image, an MLP is optimized based on the original and clipped ProPhoto color values. The MLP is embedded in the sRGB image as metadata. Contribution We address the problem of recovering the RGB colors in sRGB images back to their original wide- gamut RGB representation. Our work is inspired by coordinate-based implicit neural image representations that use multilayer perceptrons (MLPs) as a differentiable im- age representation. We propose to optimize a lightweight (23 KB) MLP model that takes the gamut-reduced RGB values and their spatial coordinates as input and predicts the original wide-gamut RGB values. The idea is to opti- mize the MLP model when the ProPhoto image is saved to sRGB and embed the MLP model parameters in the sRGB image as a comment ﬁeld. The lightweight MLP model is extracted and used to recover the wide-gamut color values when needed. We describe an optimization process for the MLP that requires ∼2 seconds per full-sized image. We demonstrate the effectiveness of our method against several different approaches, including other neural image repre- sentations and pre-trained deep-learning-based models. As part of this work, we have created a dataset of 2200 wide- gamut/small-gamut image pairs for training and testing.
Li_Improving_Vision-and-Language_Navigation_by_Generating_Future-View_Image_Semantics_CVPR_2023	Abstract Vision-and-Language Navigation (VLN) is the task that requires an agent to navigate through the environment based on natural language instructions. At each step, the agent takes the next action by selecting from a set of naviga- ble locations. In this paper, we aim to take one step further and explore whether the agent can benefit from generating the potential future view during navigation. Intuitively, hu- mans will have an expectation of how the future environ- ment will look like, based on the natural language instruc- tions and surrounding views, which will aid correct naviga- tion. Hence, to equip the agent with this ability to generate the semantics of future navigation views, we first propose three proxy tasks during the agent’s in-domain pre-training: Masked Panorama Modeling (MPM), Masked Trajectory Modeling (MTM), and Action Prediction with Image Gen- eration (APIG). These three objectives teach the model to predict missing views in a panorama (MPM), predict miss- ing steps in the full trajectory (MTM), and generate the next view based on the full instruction and navigation his- tory (APIG), respectively. We then fine-tune the agent on the VLN task with an auxiliary loss that minimizes the dif- ference between the view semantics generated by the agent and the ground truth view semantics of the next step. Em- pirically, our VLN-SIG achieves the new state-of-the-art on both Room-to-Room dataset and CVDN dataset. We fur- ther show that our agent learns to fill in missing patches in future views qualitatively, which brings more interpretabil- ity over agents’ predicted actions. Lastly, we demonstrate that learning to predict future view semantics also enables the agent to have better performance on longer paths.1	1. Introduction In Vision-and-Language Navigation, the agent needs to navigate through the environment based on natural lan- 1Code is available at https://github.com/jialuli-luka/ VLN-SIG.git . :DONWKURXJKWKHEHGURRPLQWRWKHKDOOZD\7XUQULJKWDWWKHVHFRQGGRRU« «&URVV0RGDOLW\7UDQVIRUPHU « 0DVNHG7UDMHFWRU\0RGHOLQJ070 0DVNHG3DQRUDPD0RGHOLQJ030 $FWLRQ3UHGLFWLRQZLWK,PDJHHQHUDWLRQ$3, ,PDJH7RNHQL]HU3DWFK6HPDQWLF&DOFXODWLRQ,PDJH6HPDQWLFV&RGHERRN6HOHFWLRQ&DOFXODWH,PDJH6HPDQWLFV7UDLQWR*HQHUDWH,PDJH6HPDQWLFV 7H[W(QFRGLQJ+LVWRU\(QFRGLQJ2EVHUYDWLRQ(QFRGLQJFigure 1. Overview of our proposed method VLN-SIG. We obtain the semantics of an image with a pre-trained image tokenizer, and use codebook selection (Sec. 3.4) and patch semantic calculation (Sec. 3.3) to adapt it to efficient in-domain VLN learning. We pre- train the agent on three proxy tasks (Sec. 3.5) and fine-tune the agent using Action Prediction with Image Generation (APIG) as the additional auxiliary task (Sec. 3.6). guage instructions. Many datasets have been proposed to solve this challenging task [2, 5, 22, 33, 40]. Based on these datasets, previous works aim to strengthen the navigation model or agent from several aspects: understanding long and detailed instructions in different languages, inferring and locating target objects based on common knowledge, navigating in diverse and potentially unseen environments, and learning better alignment between the environment and the instructions. However, most previous work simplifies the navigation process as an action selection process, where at each time step, the agent picks the action from a set of pre-defined candidates. This simplification does not take advantage of human’s ability to expect what scene will be more likely to happen next during navigation. For exam- ple, given the instruction “Walk through the bedroom into the hallway. Turn right and wait in the kitchen doorway.”, before walking out of the bedroom, humans will expect the hallway to be a long passage with doors, and the kitchen probably contains objects like a kitchen island and sink. Humans have these expectations based on common sense knowledge and use them to select candidates during navi- gation. Thus, in this paper, we explore whether AI agents 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. 10803 could also benefit from the ability to generate future scenes for action selection during navigation. [21] first explores the useful idea of generating future scenes with high quality, based on history observations on an indoor Vision-and-Language Navigation dataset. Fur- thermore, they adapt their synthetic future scenes for VLN by replacing the original observations with the generated observations. However, their replacement method did not enhance the agents’ performance on the VLN task. The po- tential of using semantic information from generated future observations is still underexplored. Thus, in this paper, we aim to equip the agent with both the ability to predict future scenes and also benefit from learning semantics in future generated observations. We propose VLN-SIG: Vision-and-Language Naviga- tion with Image Semantics Generation. As shown in Fig- ure 1, to first calculate the overall image semantics, we tokenize the view images into visual tokens with a pre- trained discrete variational autoencoder (dV AE). While the pre-trained dV AE has a large vocabulary of 8192, we pro- pose and compare static and dynamic codebook selection processes to optimize a subset of the codebook at one time during training. This helps the agent to learn the more im- portant tokens and focus on optimizing more difficult to- kens. Then, we propose and compare three ways based on mean value, block-wise weighted value, and sampling to represent the semantics of all patches for efficient image semantics learning. In the pre-training stage, we propose three tasks: Masked Panorama Modeling (MPM), Masked Trajectory Modeling (MTM) and Action Prediction with Image Generation (APIG). In Masked Panorama Modeling, the agent learns to predict the semantics of multiple miss- ing views in a full panorama. In Masked Trajectory Mod- eling, the agent learns to predict the semantics of multiple steps in the full trajectory. MPM and MTM together give the agent the ability to understand the semantics in each vi- sual token and learn to recognize the semantics contained in each view. In Action Prediction with Image Genera- tion, the agent mimics the navigation process to predict the next step by generating the visual tokens contained in the next navigation view. This task enables the agent to imag- ine the next step view semantics before making actions. We then fine-tune the agent on the step-by-step Vision-and- Language Navigation task. We further enhance agents’ abil- ity to predict future views by optimizing an auxiliary task during navigation, which minimizes the difference between predicted observations and the target observation. Though this task does not help the agent make navigation decision directly, the future visual semantics injected by this task helps the agent understand the environment better. We conduct experiments on Room-to-Room (R2R) datasets [2] and Cooperative Vision-and-Dialog Navigation (CVDN) dataset [40]. Empirical results show that our pro-posed VLN-SIG outperforms the strong SotA agents [8,10] by a relative gain of 4.5% in goal progress (meters) on CVDN test leaderboard, and 3% absolute gain in success rate on Room-to-Room test leaderboard. We further demon- strate that our proposed codebook selection methods and patch semantic calculation methods are crucial for learn- ing to generate image semantics with ablation studies. Be- sides, we show that our agent achieves better performance for longer paths. Lastly, we show that our agent learns to fill in missing patches in the future views, which brings more interpretability over agents’ predictions.
Kim_Generalizable_Implicit_Neural_Representations_via_Instance_Pattern_Composers_CVPR_2023	Abstract Despite recent advances in implicit neural representa- tions (INRs), it remains challenging for a coordinate-based multi-layer perceptron (MLP) of INRs to learn a common representation across data instances and generalize it for unseen instances. In this work, we introduce a simple yet effective framework for generalizable INRs that enables a coordinate-based MLP to represent complex data instances by modulating only a small set of weights in an early MLP layer as an instance pattern composer; the remaining MLP weights learn pattern composition rules for common repre- sentations across instances. Our generalizable INR frame- work is fully compatible with existing meta-learning and hy- pernetworks in learning to predict the modulated weight for unseen instances. Extensive experiments demonstrate that our method achieves high performance on a wide range of domains such as an audio, image, and 3D object, while the ablation study validates our weight modulation.	1. Introduction Implicit neural representations (INR) have shown the po- tential to represent complex data as continuous functions. Assuming that a data instance comprises the pairs of a co- ordinate and its output features, INRs adopt a parameterized neural network as a mapping function from an input coor- dinate into its output features. For example, a coordinate- based MLP [23] predicts RGB values at each 2D coordinate as an INR of an image. Despite the popularity of INRs, a trained MLP cannot be generalized to represent other in- stances, since each MLP learns to memorize each data in- stance. Thus, INRs necessitate separate training of MLPs to represent a lot of data instances as continuous functions. *Equal contribution †Corresponding author Figure 1. The reconstructed images of 178 ×178 ImageNette by TransINR [4] (left) and our generalizable INRs (right). Generalizable INRs aim to learn common representa- tions of a MLP across instances, while modulating fea- tures or weights of the coordinate-based MLP to adapt unseen data instances [4, 7, 24]. The feature-modulation method exploits the latent vector of an instance to con- dition the activations in MLP layers through concatena- tion [17] or affine-transform [8, 18]. Despite the computa- tional efficiency of feature-modulation, the modulated INRs have unsatisfactory results to represent complex data due to their limited modulation capacity. On the other hand, the weight-modulation method learns to update the whole MLP weights to increase the modulation capacity for high perfor- mance. However, modulating whole MLP weights leads to unstable and expensive training [4, 7, 9, 24]. In this study, we propose a simple yet effective frame- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11808 work for generalizable INRs via Instance Pattern Com- posers to modulate only a small set of MLP weights. We postulate that a complex data instance can be represented by composing low-level patterns in the instance [23]. Thus, we rethink and categorize the weights of MLP into i) instance pattern composers and ii) pattern composition rule . The in- stance pattern composer is a weight matrix in the early layer of our coordinate-based MLP to extract the instance content patterns of each data instance as a low-level feature. The re- maining weights of MLP is defined as a pattern composition rule, which composes the instance content patterns in an instance-agnostic manner. In addition, our framework can adopt both optimization-based meta-learning and hypernet- works to predict the instance pattern composer for an INR of unseen instance. In experiments, we demonstrate the ef- fectiveness of our generalizable INRs via instance pattern composers on various domains and tasks. Our main contributions are summarized as follows. 1) Instance pattern composers enable a coordinate-based MLP to represent complex data by modulating only one weight, while pattern composition rule learns the common repre- sentation across data instances. 2) Our instance pattern composers are compatible with optimization-based meta- learning and hypernetwork to predict modulated wights of unseen data during training. 3) We conduct extensive exper- iments to demonstrate the effectiveness of our framework through quantitative and qualitative analysis.
Liu_PoseExaminer_Automated_Testing_of_Out-of-Distribution_Robustness_in_Human_Pose_and_CVPR_2023	Abstract Human pose and shape (HPS) estimation methods achieve remarkable results. However, current HPS bench- marks are mostly designed to test models in scenarios that are similar to the training data. This can lead to criti- cal situations in real-world applications when the observed data differs significantly from the training data and hence is out-of-distribution (OOD). It is therefore important to test and improve the OOD robustness of HPS methods. To address this fundamental problem, we develop a simula- tor that can be controlled in a fine-grained manner us- ing interpretable parameters to explore the manifold of im- ages of human pose, e.g. by varying poses, shapes, and clothes. We introduce a learning-based testing method, termed PoseExaminer, that automatically diagnoses HPS algorithms by searching over the parameter space of hu- man pose images to find the failure modes. Our strat- egy for exploring this high-dimensional parameter space is a multi-agent reinforcement learning system, in which the agents collaborate to explore different parts of the pa- rameter space. We show that our PoseExaminer discov- ers a variety of limitations in current state-of-the-art mod- els that are relevant in real-world scenarios but are missed by current benchmarks. For example, it finds large regions of realistic human poses that are not predicted correctly, as well as reduced performance for humans with skinny and corpulent body shapes. In addition, we show that fine-tuning HPS methods by exploiting the failure modes found by PoseExaminer improve their robustness and even their performance on standard benchmarks by a significant margin. The code are available for research purposes at https://github.com/qihao067/PoseExaminer.	1. Introduction In recent years, the computer vision community has made significant advances in 3D human pose and shape (HPS) estimation. But despite the high performance on standard benchmarks, current methods fail to give reliable predictions for configurations that have not been trained on Figure 1. PoseExaminer is an automatic testing tool used to study the performance and robustness of HPS methods in terms of ar- ticulated pose, shape, global rotation, occlusion, etc. It system- atically explores the parameter space and discovers a variety of failure modes. (a) illustrates three failure modes in PARE [18] and (b) shows the efficacy of training with PoseExaminer. or in difficult viewing conditions such as when humans are significantly occluded [18] or have unusual poses or cloth- ing [34]. This lack of robustness in such out-of-distribution (OOD) situations, which typically would not fool a human observer, is generally acknowledged and is a fundamental open problem for HPS methods that should be addressed. The main obstacle, however, to testing the robustness of HPS methods is that test data is limited, because it is ex- pensive to collect and annotate. One way to address this problem is to diagnose HPS systems by generating large synthetic datasets that randomly generate images of humans in varying poses, clothing, and background [4, 34]. These studies suggest that the gap between synthetic and real do- 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. 672 mains in HPS is sufficiently small for testing on synthetic to be a reasonable strategy, as we confirm in our experiments. These methods, however, only measure the average per- formance on the space of images. In sensitive domains that use HPS ( e.g. autonomous driving), the failure modes , where performance is low, can matter more. Most impor- tantly, despite large scale, these datasets lack diversity in some dimensions. For example, they mostly study pose for common actions such as standing, sitting, walking, etc. Inspired by the literature on adversarial machine learn- ing [5, 11], recent approaches aim to systematically ex- plore the parameter space of simulators for weaknesses in the model performance. This has proven to be an effec- tive approach for diagnosing limitations in image classifi- cation [45], face recognition [43], and path planning [39]. However, they are not directly applicable to testing the ro- bustness of the high-dimensional regression task of pose es- timation. For example, [43, 45] were designed for simple binary classification tasks and can only optimize a limited number of parameters, and [39] only perturbs an initial real- world scene to generate adversarial examples, which does not enable the full exploration of the parameter space. In this work, we introduce PoseExaminer (Fig. 1), a learning-based algorithm to automatically diagnose the ro- bustness of HPS methods. It efficiently searches through the high-dimensional continuous space of a simulator of human images to find failure modes. The strategy in Pose- Examiner is a multi-agent reinforcement learning approach, in which the agents collaborate with one another to search the latent parameter space for model weaknesses. More specifically, each agent starts at different random initialized seeds and explores the latent parameter space to find failure cases, while at the same time avoiding exploring the regions close to the other agents in the latent space. This strategy enables a highly parallelizable way of exploring the high- dimensional continuous latent space of the simulator. After converging to a local optimum, each agent explores the lo- cal parameter space to find a connected failure region that defines a whole subspace of images where the pose is in- correctly predicted ( i.e. failure mode). We demonstrate that very large subspaces of failures exist even in the best HPS models. We use the size of these failure subspaces together with the success rate of the agents as a new measure of out- of-distribution robustness. Our experiments on four state-of-the-art HPS models show that PoseExaminer successfully discovers a variety of failure modes that provide new insights about their real- world performance. For example, it finds large subspaces of realistic human poses that are not predicted correctly, and reduced performance for humans with skinny and corpulent body shapes. Notably, we find that the failure modes found in synthetic data generalize well to real images: In addi- tion to using PoseExaminer as a new benchmark, we alsofind that fine-tuning SOTA methods using the failure modes discovered by PoseExaminer enhances their robustness and even the performance on 3DPW [52] and AIST++ [25, 50] benchmarks. We also note that computer graphics rendering pipelines become increasingly realistic, which will directly benefit the quality of our automated testing approach. In short, our contributions are three-fold: • We propose PoseExaminer, a learning-based algorithm to automatically diagnose the robustness of human pose and shape estimation methods. Compared to prior work, PoseExaminer is the first to efficiently search for a variety of failure modes in the high-dimensional con- tinuous parameter space of a simulator using a multi- agent reinforcement learning framework. • We introduce new metrics, which have not been pos- sible to measure before, for quantifying the robust- ness of HPS methods based on our automated testing framework. We perform an in-depth analysis of cur- rent SOTA methods, revealing a variety of diverse fail- ure modes that generalize well to real images. • We show that the failure modes discovered by PoseEx- aminer can be used to significantly improve the real- world performance and robustness of current methods.
Kim_Achieving_a_Better_Stability-Plasticity_Trade-Off_via_Auxiliary_Networks_in_Continual_CVPR_2023	Abstract In contrast to the natural capabilities of humans to learn new tasks in a sequential fashion, neural networks are known to suffer from catastrophic forgetting , where the model’s performances on old tasks drop dramatically after being optimized for a new task. Since then, the continual learning (CL) community has proposed several solutions aiming to equip the neural network with the ability to learn the current task ( plasticity ) while still achieving high accu- racy on the previous tasks ( stability ). Despite remarkable improvements, the plasticity-stability trade-off is still far from being solved and its underlying mechanism is poorly understood. In this work, we propose Auxiliary Network Continual Learning (ANCL), a novel method that applies an additional auxiliary network which promotes plasticity to the continually learned model which mainly focuses on stability. More concretely, the proposed framework mate- rializes in a regularizer that naturally interpolates between plasticity and stability, surpassing strong baselines on task incremental and class incremental scenarios. Through extensive analyses on ANCL solutions, we identify some essential principles beneath the stability-plasticity trade- off. The code implementation of our work is available at https://github.com/kim-sanghwan/ANCL .	1. Introduction The continual learning (CL) model aims to learn from current data while still maintaining the information from previous training data. The naive approach of continuously fine-tuning the model on sequential tasks, however, suffers from catastrophic forgetting [8, 21]. Catastrophic forget- ting occurs in a gradient-based neural network because the updates made with the current task are likely to override the model weights that have been changed by the gradients from the old tasks. Catastrophic forgetting can be understood in terms ofstability-plasticity dilemma [22], one of the well-known challenges in continual learning. Specifically, the model not only has to generalize well on past data ( stability ) but also learn new concepts ( plasticity ). Focusing on stability will hinder the neural network from learning the new data, whereas too much plasticity will induce more forgetting of the previously learned weights. Therefore, CL model should strike a balance between stability and plasticity. There are various ways to define the problem of CL. Generally speaking, it can be categorized into three sce- narios [27] : Task Incremental Learning (TIL), Domain In- cremental Learning (DIL), and Class Incremental Learning (CIL). In TIL, the model is informed about the task that needs to be solved; the task identity is given to the model during the training session and the test time. In DIL, the model is required to solve only one task at hands without the task identity. In CIL, the model should solve the task itself and infer the task identity. Since the model should discriminate all classes that have been seen so far, it is usu- ally regarded as the hardest continual learning scenario. Our study performs extensive evaluations on TIL and CIL set- ting which will be further explained in Sec. 4. Recently, several papers [19, 20, 28, 31] proposed the usage of an auxiliary network or an extra module that is solely trained on the current dataset, with the purpose of combining this additional structure with the previous net- work or module that has been continuously trained on the old datasets. For example, Active Forgetting with synap- tic Expansion-Convergence (AFEC) [28] regularizes the weights relevant to the current task through a new set of network parameters called the expanded parameters based on weight regularization methods. The expanded param- eters are solely optimized on the current task and are al- lowed to forget the previous ones. As a result, AFEC can reduce potential negative transfer by selectively merging the old parameters with the expanded parameters. The stability- plasticity balance in AFEC is adjusted via hyperparameters which scale the regularization terms for remembering the old tasks and learning the new tasks. 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. 11930 The authors of the above papers propose to mitigate the stability-plasticity dilemma by infusing plasticity through the auxiliary network or module (detailed explanation in Appendix A). However, a precise characterization of the in- teractive mechanism between the previous model and the auxiliary model is still missing in the literature. Therefore, in this paper, we first formalize the framework of CL that adopts the auxiliary network called Auxiliary Network Con- tinual Learning (ANCL). Given this environment, we then investigate the stability-plasticity trade-off through various analyses from both a theoretical and empirical point of view. Our main contributions can be summarized as follows: • We propose the framework of Auxiliary Network Con- tinual Learning (ANCL) that can naturally incorporate the auxiliary network into a variety of CL approaches as a plug-in method (Sec. 3.1). • We empirically show that ANCL outperforms existing CL baselines on both CIFAR-100 [16] and Tiny Ima- geNet [17] (Sec. 4). • Furthermore, we perform three analyses to investigate the stability-plasticity trade-off within ANCL (Sec. 5): Weight Distance ,Centered Kernel Alignment , and Mean Accuracy Landscape .
Koley_Picture_That_Sketch_Photorealistic_Image_Generation_From_Abstract_Sketches_CVPR_2023	Abstract Sketches Subhadeep Koley1,2Ayan Kumar Bhunia1Aneeshan Sain1,2Pinaki Nath Chowdhury1,2 Tao Xiang1,2Yi-Zhe Song1,2 1SketchX, CVSSP, University of Surrey, United Kingdom. 2iFlyTek-Surrey Joint Research Centre on Artificial Intelligence. {s.koley, a.bhunia, a.sain, p.chowdhury, t.xiang, y.song }@surrey.ac.uk (a) (b) Edgemap Sketch Sketch Sketch Sketch Sketch Existing Methods Proposed Method Figure 1. (a) Set of photos generated by the proposed method. (b) While existing methods can generate faithful photos from perfectly pixel-aligned edgemaps , they fall short drastically in case of highly deformed and sparse free-hand sketches . In contrast, our autoregressive sketch-to-photo generation model produces highly photorealistic outputs from highly abstract sketches . Abstract Given an abstract, deformed, ordinary sketch from un- trained amateurs like you and me, this paper turns it into a photorealistic image – just like those shown in Fig. 1(a), all non-cherry-picked. We differ significantly from prior art in that we do not dictate an edgemap-like sketch to start with, but aim to work with abstract free-hand human sketches. In doing so, we essentially democratise the sketch- to-photo pipeline, “picturing” a sketch regardless of how good you sketch. Our contribution at the outset is a de- coupled encoder-decoder training paradigm, where the de- coder is a StyleGAN trained on photos only. This impor- tantly ensures that generated results are always photoreal- istic. The rest is then all centred around how best to deal with the abstraction gap between sketch and photo. For that, we propose an autoregressive sketch mapper trained on sketch-photo pairs that maps a sketch to the StyleGAN latent space. We further introduce specific designs to tackle the abstract nature of human sketches, including a fine- grained discriminative loss on the back of a trained sketch- photo retrieval model, and a partial-aware sketch augmen- tation strategy. Finally, we showcase a few downstream tasks our generation model enables, amongst them is show- ing how fine-grained sketch-based image retrieval, a well- studied problem in the sketch community, can be reduced to an image (generated) to image retrieval task, surpass- ing state-of-the-arts. We put forward generated results in the supplementary for everyone to scrutinise. Project page: https://subhadeepkoley.github.io/PictureThatSketch	1. Introduction People sketch, some better than others. Given a shoe image like ones shown in Fig. 1(a), everyone can scribble a few lines to depict the photo, again mileage may vary – top left sketch arguably lesser than that at bottom left. The opposite, i.e., hallucinating a photo based on even a very ab- stract sketch, is however something humans are very good at having evolved on the task over millions of years. This seemingly easy task for humans, is exactly one that this pa- per attempts to tackle, and apparently does fairly well at – given an abstract sketch from untrained amateurs like us, our paper turns it into a photorealistic image (see Fig. 1). This problem falls into the general image-to-image trans- lation literature [41, 64]. Indeed, some might recall prior arts ( e.g., pix2pix [41], CycleGAN [105], MUNIT [38], Bi- cycleGAN [106]), and sketch-specific variants [33, 86] pri- marily based on pix2pix [41] claiming to have tackled the exact problem. We are strongly inspired by these works, but significantly differ on one key aspect – we aim to generate from abstract human sketches, not accurate photo edgemaps which are already “photorealistic”. This is apparent in Fig. 1(b), where when edgemaps are used prior works can hallucinate high-quality photorealis- tic photos, whereas rather “peculiar” looking results are ob- tained when faced with amateur human sketches. This is be- cause all prior arts assume pixel-alignment during transla- tion – so your drawing skill (or lack of it), got accurately re- flected in the generated result. As a result, chance is you and me will not fetch far on existing systems if not art-trained 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. 6850 sketch photorealistic edgemaps – we, in essence, democra- tise the sketch-to-photo generation technology, “picturing” a sketch regardless of how good you sketch. Our key innovation comes after a pilot study where we discovered that the pixel-aligned artefact [69] in prior art is a direct result of the typical encoder-decoder [41] archi- tecture being trained end-to-end – this enforces the gener- ated results to strictly follow boundaries defined in the in- put sketch (edgemap). Our first contribution is therefore a decoupled encoder-decoder training, where the decoder is pre-trained StyleGAN [46] trained on photos only, and is frozen once trained. This importantly ensures generated re- sults are sampled from the StyleGAN [46] manifold there- fore of photorealistic quality. The second, perhaps more important innovation lies with how we bridge the abstraction gap [20, 21, 36] between sketch and photo. For that, we propose to train an encoder that performs a mapping from abstract sketch representa- tion to the latent space of the learned latent space of Style- GAN [46] ( i.e., not actual photos as per the norm). To train this encoder, we use ground-truth sketch-photo pairs, and impose a novel fine-grained discriminative loss between the input sketch and the generated photo, together with a con- ventional reconstruction loss [102] between the input sketch and the ground-truth photo, to ensure the accuracy of this mapping process. To double down on dealing with the ab- stract nature of sketches, we further propose a partial-aware augmentation strategy where we render partial versions of a full sketch and allocate latent vectors accordingly (the more partial the input, the lesser vectors assigned). Our autoregressive generative model enjoys a few in- teresting properties once trained: (i)abstraction level ( i.e., how well the fine-grained features in a sketch are reflected in the generated photo) can be easily controlled by alter- ing the number of latent vectors predicted and padding the rest with Gaussian noise, (ii)robustness towards noisy and partial sketches, thanks to our partial-aware sketch aug- mentation strategy, and (iii) good generalisation on input sketches across different abstraction levels (from edgemaps, to sketches across two datasets). We also briefly show- case two potential downstream tasks our generation model enables: fine-grained sketch-based image retrieval (FG- SBIR), and precise semantic editing. On the former, we show how FG-SBIR, a well-studied task in the sketch com- munity [10,71–73], can be reduced to an image (generated) to image retrieval task, and that a simple nearest-neighbour model based on VGG-16 [78] features can already surpass state-of-the-art. On the latter, we demonstrate how pre- cise local editing can be done that is more fine-grained than those possible with text and attributes. We evaluate using conventional metrics (FID, LPIPS), plus a new retrieval-informed metric to demonstrate supe- rior performance. But, as there is no better way to convincethe jury other than presenting allfacts, we offer allgener- ated results in the supplementary for everyone to scrutinise.
Kong_Indescribable_Multi-Modal_Spatial_Evaluator_CVPR_2023	Abstract Multi-modal image registration spatially aligns two im- ages with different distributions. One of its major challenges is that images acquired from different imaging machines have different imaging distributions, making it difficult to focus only on the spatial aspect of the images and ignore differences in distributions. In this study, we developed a self-supervised approach, Indescribable Multi-model Spatial Evaluator (IMSE), to address multi-modal image registration. IMSE creates an accurate multi-modal spatial evaluator to measure spatial differences between two images, and then op- timizes registration by minimizing the error predicted of the evaluator. To optimize IMSE performance, we also proposed a new style enhancement method called Shuffle Remap which randomizes the image distribution into multiple segments, and then randomly disorders and remaps these segments, so that the distribution of the original image is changed. Shuffle Remap can help IMSE to predict the difference in spatial location from unseen target distributions. Our results show that IMSE outperformed the existing methods for registration using T1-T2 and CT-MRI datasets. IMSE also can be easily integrated into the traditional registration process, and can provide a convenient way to evaluate and visualize registra- tion results. IMSE also has the potential to be used as a new paradigm for image-to-image translation. Our code is avail- able at https://github.com/Kid-Liet/IMSE . *Corresponding author. X Y IMSE Instance mapping X YGAN Distribution mapping Moving x Target y Corresponding Non-corresponding Figure 1. The GAN based methods can only ensure that the distri- bution of the Xdomain is mapped that of the Ydomain. Ideally, we want to achieve instance registration in which moving and target images are one-to-one corresponded.	1. Introduction The purpose of multi-modal image registration is to align two images with different distributions (Moving ( M) and Target ( T) images) by warping the space through the defor- mation field ϕ. A major challenge in multi-modal image registration is that images from different modalities may dif- fer in multiple aspects given the fact that images are acquired using different imaging machines, or different acquisition parameters. Due to dramatically different reconstruction and acquisition methods, there is no simple one-to-one map- ping between different imaging modalities. From the per- spective of measuring similarity, mainstream unsupervised multi-modal registration methods can be divided into two categories: similarity operator based registration and image- to-image translation based registration. Similarity operator based registration uses multi-modal similarity operators as loss functions for registration, for ex- 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. 9853 ample, normalized cross-correlation (NCC) [15, 22, 31, 32], mutual information (MI) [5, 23, 27, 35], and modality- independent neighborhood descriptor (MIND) [3, 8, 13, 39]. Similarity operators are based on a prior mathematical knowl- edge. These are carefully designed and improved over time. They can be applied to both traditional registration process (Eq. 1) and neural network registration (Eq. 2): ˆϕ= arg min ϕLsim(M(ϕ), T).(1) Or ˆθ= arg min θ E(M,T )[Lsim(M, T, g θ(M, T ))] .(2) Similarity operators have several limitations. 1)It is un- likely to design similarity operators that can maintain high accuracy for all data from various imaging modalities. 2) It is not possible to estimate the upper limit these operators can achieve and hence it is difficult to find the improvement directions. Image-to-image translation [1, 16, 17, 21, 29, 38, 42] based multi-modal registration first translations multi-modal images into single-modal images (Eq 3) using a generative adversarial network (GAN [10]), and then use Mean Abso- lute Error (MAE) or Mean Squared Error (MSE) to evaluate the error at each pixel in space (Eq 4). min Gmax DLAdv(G, D ) =ET[log(D(T))] + EM[log(1−D(G(M)))].(3) And ˆθ= arg min θ E(M,T )[∥G(M), T, g θ(G(M), T)∥1] . (4) Image-to-image translation based registration cleverly avoids the complex multi-modal problem and reduces the difficulty of registration to a certain extent. However, it has obvious drawbacks. 1)The methods based on GAN require training a generator using existing multi-modal data. The trained model will not work if it encounters unseen data, which greatly limits its applicable scenarios. 2)More importantly, registration is an instance problem. However, the method based on GAN is to remap the data distribution between different modal. As shown in Figure 1, the distribution has bias and variance. We cannot guarantee that the translated image corresponds exactly to the instance target image at the pixel level. For example, Figure 2 shows that there is still residual distribution difference between the target image and translated image. Therefore, even if they are well aligned in space, there is still a large error in the same organ. To address these challenges, we propose a novel idea based on self-supervision, namely, Indescribable Multi- modal Spatial Evaluator, or IMSE for short. The IMSE 1 0 CT - MR Moving Translated TargetCycleGAN ErrorIMSE Error T1 - T2Figure 2. The distributions of the Moving images translated by Cy- cleGAN still have large residual differences from the Target images. IMSE gives smaller error assessment values in the overlapping regions (converging to blue). approach creates an accurate multi-modal spatial evaluator to metric spatial differences between two images, and then optimizes registration by minimizing the error predicted by the evaluator. In Figure 2, we provide a visual demonstration of IMSE in terms of spatial location for T1-T2 and MRI-CT, respectively. Even though distribution differences between the Moving and Target images are still significant, IMSE is low in overlapping regions of the same organs. The main contributions of this study can be summarized as follows: •We introduce the concepts of relative single-modal and absolute single-modal as an extension of the current definition of single-modality. •Based on relative single-modal and absolute single- modal, we propose a self-supervised IMSE method to evaluate spatial differences in multi-modal image registration. The main advantage of IMSE is that it focuses only on differences in spatial location while ignoring differences in multi-modal distribution caused by different image acquisition mechanisms. Our results show that IMSE outperformed the existing metrics for registration using T1-T2 and CT-MRI datasets. •We propose a new style enhancement method named Shuffle Remap. Shuffle Remap can help IMSE to ac- curate predict the difference in spatial location from an unseen target distribution. As a enhancement method, Shuffle Remap can be impactful in the field of domain generalization. •We develop some additional functions for IMSE. 1) As a measure, IMSE can be integrated into both the registration based on neural network and the traditional registration. 2)IMSE can also be used to establish a new image-to-image translation paradigm. 3)IMSE can provide provide a convenient way to evaluate and visualize registration results. 9854
Lim_BiasAdv_Bias-Adversarial_Augmentation_for_Model_Debiasing_CVPR_2023	Abstract Neural networks are often prone to bias toward spurious correlations inherent in a dataset, thus failing to generalize unbiased test criteria. A key challenge to resolving the issue is the significant lack of bias-conflicting training data (i.e., samples without spurious correlations). In this paper, we propose a novel data augmentation approach termed Bias- Adversarial augmentation (BiasAdv) that supplements bias- conflicting samples with adversarial images. Our key idea is that an adversarial attack on a biased model that makes decisions based on spurious correlations may generate syn- thetic bias-conflicting samples, which can then be used as augmented training data for learning a debiased model. Specifically, we formulate an optimization problem for gen- erating adversarial images that attack the predictions of an auxiliary biased model without ruining the predictions of the desired debiased model. Despite its simplicity, we find that BiasAdv can generate surprisingly useful synthetic bias-conflicting samples, allowing the debiased model to learn generalizable representations. Furthermore, BiasAdv does not require any bias annotations or prior knowledge of the bias type, which enables its broad applicability to exist- ing debiasing methods to improve their performances. Our extensive experimental results demonstrate the superiority of BiasAdv, achieving state-of-the-art performance on four popular benchmark datasets across various bias domains.	1. Introduction Real-world datasets are often inherently biased [2, 34], where certain visual attributes are spuriously correlated with class labels. For example, let us consider a binary clas- sification task between cats and dogs. Unbeknownst to us, our dataset could consist of most cats indoors and most dogs outdoors, as illustrated in Figure 1. When trained on such a biased dataset, neural networks often learn unintended shortcuts [2, 8, 34, 39] ( e.g., making predictions based on *Corresponding author: [email protected] (Cat, Indoor) ( Dog, Outdoor) Debiased model (Cat, Outdoor) ( Dog, Indoor) Synthetic bias-conflicting samples by BiasAdv Cat or Dog Auxiliary model (Biased) Adversarial Adversari a AttackObjective : Classify Catvs. Dog Bias attribute : Indoor vs. Outdoor (Unknown) Outdoor) ( Dog ( Ind Bias-conflicting ndoor) ( Dog ( O Bias-guidingFigure 1. An overview of BiasAdv. In MetaShift [26], the bias attribute {Indoor, Outdoor }is spuriously correlated to the class label{Cat, Dog }. In this work, we refer to data with such spurious correlations as bias-guiding samples and without such correlations asbias-conflicting samples, respectively. Using the biased dataset, we train an auxiliary model to be biased, and BiasAdv supple- ments bias-conflicting samples using adversarial images which at- tack the biased predictions of the auxiliary model while preserving the predictions of the debiased model. By leveraging the diversi- fied bias-conflicting data, BiasAdv allows the debiased model to learn generalizable representations for unbiased classification. the background) and fail to generalize in a new unbiased test environment. To tackle the problem, conventional methods have utilized explicit bias annotations [1,19,39,43] or prior knowledge of the bias type [2, 3, 5, 9, 46]. However, bias annotations are expensive and laborious to obtain, and pre- suming certain bias types in advance limits the capability to be universally applicable to various bias types. To train a debiased model without bias annotations, the main line of recent research [6, 22, 30, 34, 42] has com- monly utilized an intentionally biased model as an auxiliary model under the idea that bias attributes are easy-to-learn . In essence, these methods identify bias-conflicting samples based on the auxiliary model and train the debiased model in a way that focuses more on the identified samples ( i.e., re- 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. 3832 weighting based on the auxiliary model). Although recent re-weighting methods have achieved remarkable success in debiasing without bias annotations, they have inherent limi- tation; since the number of bias-conflicting samples is often too small for a model to learn generalizable representations, the model is prone to over-fitting [25]. Consequently, re- weighting methods suffer from the degraded performance on bias-guiding samples [20, 44], which raises the question of whether these methods truly make models debiased or simply deflect models in unintended directions. To resolve the aforementioned issues, data augmentation methods have recently been proposed to supplement bias- conflicting samples. For example, BiaSwap [20] conducts image-to-image translation to synthesize bias-conflicting samples. However, it requires delicate training of complex and expensive image translation models [36], limiting its applicability. On the other hand, DFA [25] utilizes feature- level swapping based on disentangled representations be- tween bias-guiding and bias-conflicting features. Learning disentangled representations, however, is often challenging on real-world datasets [27, 28, 31]. In this paper, we devise a much simpler yet more effec- tive approach to generate bias-conflicting samples, coined Bias-Adversarial augmentation (BiasAdv). Figure 1 shows an overview of BiasAdv. We utilize an auxiliary model that intentionally learns biased shortcuts, likewise [30, 34]. The key idea of BiasAdv is that an adversarial attack on the biased auxiliary model may generate adversarial images that alter the bias cue from the input images ( i.e., bias- conflicting samples). Concretely, we formulate an optimiza- tion problem to generate adversarial images that attack the predictions of the biased auxiliary model without ruining the predictions of the desired debiased model. Then, the generated adversarial images are used as additional training data to train the debiased model. It is noteworthy that, un- like previous data augmentation methods [20, 25], BiasAdv does not require complex image translation models or dis- entangled representations, so it can be seamlessly applied to any debiasing method based on the biased model. Fur- thermore, we show that BiasAdv, despite its simplicity, can generate surprisingly useful synthetic bias-conflicting sam- ples, which significantly improves debiasing quality. The main contributions of our work are three-fold: • We propose BiasAdv, a simple and effective data aug- mentation method for model debiasing, which utilizes adversarially attacked images as additional training data. Our method does not require any bias annotations or prior knowledge of the bias type during training. • BiasAdv can be easily applied to existing re-weighting methods without architectural or algorithmic changes. We confirm that BiasAdv significantly improves the performance, achieving up to 22.8%, 13.4%, 7.9%,and 8.0% better performance than the state-of-the-art results on CIFAR-10C [25], BFFHQ [25], BAR [34], and MetaShift [26], respectively. • We demonstrate the effectiveness of BiasAdv through extensive ablation studies and analyses. Our key find- ing is that BiasAdv helps to learn generalizable repre- sentations and prevents over-fitting; it does not degrade the performance of bias-guiding samples and improves model robustness against input corruptions.
Li_Causally-Aware_Intraoperative_Imputation_for_Overall_Survival_Time_Prediction_CVPR_2023	Abstract Previous efforts in vision community are mostly made on learning good representations from visual patterns. Beyond this, this paper emphasizes the high-level ability of causal reasoning. We thus present a case study of solving the chal- lenging task of Overall Survival (OS) time in primary liver cancers. Critically, the prediction of OS time at the early stage remains challenging, due to the unobvious image pat- terns of reflecting the OS. To this end, we propose a causal inference system by leveraging the intraoperative attributes and the correlation among them, as an intermediate super- vision to bridge the gap between the images and the final OS. Particularly, we build a causal graph, and train the im- ages to estimate the intraoperative attributes for final OS prediction. We present a novel Causally-aware Intraop- erative Imputation Model (CAWIM) that can sequentially predict each attribute using its parent nodes in the esti- mated causal graph. To determine the causal directions, *Equal contribution †Corresponding authorwe propose a splitting-voting mechanism, which votes for the direction for each pair of adjacent nodes among multi- ple predictions obtained via causal discovery from hetero- geneity. The practicability and effectiveness of our method are demonstrated by the promising results on liver cancer dataset of 361 patients with long-term observations.	1. Introduction The success of recent deep learning model is largely at- tributed to learning the good representations for visual pat- terns. Such representations essentially facilitate various vi- sion task, such as recognition and synthesis [15, 25,33]. Nevertheless, one important goal for the vision commu- nity is to model and summarize the relationships of ob- served variables of a system, in order to enable well pre- dictions on similar data. Essentially, it is desirable to un- derstand how the system is changed if one modifies these relationships under certain conditions, e.g., the effects of a treatment in healthcare. Thus this demands the high- 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. 15681 level ability of causal reasoning beyond the previous ef- forts of only learning good representations for visual pat- terns [1, 5, 16,21,35,37]. This naturally leads into our task of causal inference. This paper presents a case study of solving the challeng- ing task of Overall Survival (OS) time estimation in Pri- mary Liver Cancers (PLC). Generally, the liver cancer re- mains one of the most common malignancies worldwide in the 21st century, as there are about one million new cases every year [36]. The five-year survival rate for advanced PLC is only about 5% [7]. Therefore, early and accurate prediction of OS time estimation can provide informative guidance for individualized treatment planning and reduc- ing burden of medical resources [6, 27,32,34]. On the other hand, one shall easily notice that with the renaissance of deep learning, great achievements have been made on medical imaging analysis, such as diagnosis, segmentation, and classification [16, 21,35,37]. Unfortunately, it still re- mains challenging for experienced clinicians to predict OS time at early stage, even with advanced modern diagnostic tools such as Magnetic Resonance Imaging (MRI) or tumor marker tests [3, 11], and the deep learning tools [1, 5]. Some studies propose to leverage deep neural networks for OS time prediction. They focus on fusion learning of multi-modal image features and some basic information (i.e., age and gender) [5, 8,14,22,24,30,39]. Neverthe- less, the accurate prediction based on only preoperative in- formation (such as image and tumor marker indexes) is still challenging, possibly due to missing information from early diagnosis stage to the final stage. This missing informa- tion includes the texture and pathological attributes of the liver, health level of the patient, and post-operative treat- ment [9, 17,20]. For example, as in Fig. 1, patient A and B with different OS time have almost identical preopera- tive indicators cause, history of disease, etc,, making the preoperative-based model hard for discrimination. To amend this problem, we present a causal inference system that can well utilize the intraoperative information, which is pretty easy to be accessed in training data. Accord- ing to medical priories, such information records patholog- ical attributes, which can be more reflective about the pa- tient’s health level and the postoperative recovery. Inspired by this, we propose to leverage this auxiliary information to help build our causal inference system. Again, we take the example in Fig. 1. Although preoperative information cannot differentiate patient A from B, their intraoperative index shows great difference in distribution, which can thus be employed for the discrimination. Additionally, there are many indicators from medical experts that the intraoperative indexes are related to each other. For instance, the clinico- pathologic hepatocirrhosis is dependent on the hepatocir- rhosis; the sum of tumor diameter is affected by the num- ber of tumors; the fibrosis (reflected on S-Score) can prob-ably deteriorate to cirrhosis [4], etc. By leveraging these relationships, we can better understand the causal inference concrete from both observed data modelling and medical expert-level knowledge of these variables. To this end, we encapsulate these priors and the inspired proposals into a new method, dubbed as Causally-Aware In- traoperative Imputation Model (CAWIM). It incorporates causal discovery module to sequentially estimate intraop- erative indexes as an intermediate stage towards final OS time prediction. Specifically, our model is composed of two key steps: i)estimating the intraoperative indexes us- ing preoperative information, i.e., image and indexes; ii) followed by OS time prediction using estimated intraopera- tive indexes and preoperative information. To achieve more accurate prediction of intraoperative indexes that is determi- nant to the prediction power of the whole method, we pro- pose a Causally-aware Directed Acyclic Graph (CaDAG) module. It learns the causal structure represented as a DAG over intraoperative features. To identify the causal rela- tions beyond the traditional PC algorithm [26], we propose asplitting-voting mechanism, which is inspired by the re- cent work [18] that learn the causal structure with the as- sistance of an auxiliary domain index variable. Our pro- posed mechanism can not only identify the causal relations even when this domain index variable is not available, but also can be theoretically guaranteed that the learned graph is not acyclic. During test stage, we sequentially estimate each intraoperative index with preoperative information and additionally, its parent set among other intraoperative in- dexes. The utility of our method can be demonstrated by a significant improvement of OS time prediction on an in- house liver cancer dataset, as well as better interpretabil- ity of learned causal structure, more accurate estimation of intraoperative indexes and more interpretable visualization results. In a nutshell, we for the first time present a case study of building a causally-aware intraoperative imputation sys- tem for the challenging task of overall survival time pre- diction. The proposed method of building the casual in- ference system, can be naturally extended to other similar medical tasks. Our key contributions are listed as follows. (1) New Paradigm for OS time Prediction. We propose to leverage intraoperative indexes as an intermediate stage during training. The leverage of this information can sig- nificantly alleviate the “missing information” issue. To the best of our knowledge, we are the first to leverage auxiliary information (in addition to preopearative features) for OS time prediction. (2) Causal Structure Learning. We pro- pose a novel splitting-voting mechanism that can identify the causal structure even when the domain index variable is missing. (3) Better Prediction Power. Our method can sig- nificantly improve the prediction power for liver cancer over the competitors. The methods are evaluated on the medical 15682 dataset, which, to the best of our knowledge, is the largest primary live cancer dataset.
Li_Hard_Sample_Matters_a_Lot_in_Zero-Shot_Quantization_CVPR_2023	Abstract Zero-shot quantization (ZSQ) is promising for compress- ing and accelerating deep neural networks when the data for training full-precision models are inaccessible. In ZSQ, network quantization is performed using synthetic samples, thus, the performance of quantized models depends heavily on the quality of synthetic samples. Nonetheless, we find that the synthetic samples constructed in existing ZSQ meth- ods can be easily fitted by models. Accordingly, quantized models obtained by these methods suffer from significant performance degradation on hard samples. To address this issue, we propose HArd sample Synthesizing and Training (HAST). Specifically, HAST pays more attention to hard sam- ples when synthesizing samples and makes synthetic samples hard to fit when training quantized models. HAST aligns features extracted by full-precision and quantized models to ensure the similarity between features extracted by these two models. Extensive experiments show that HAST significantly outperforms existing ZSQ methods, achieving performance comparable to models that are quantized with real data.	1. Introduction Deep neural networks (DNNs) achieve great success in many domains, such as image classification [28, 43, 44], ob- ject detection [15, 16, 39], semantic segmentation [13, 53], and embodied AI [3,4,11]. These achievements are typically paired with the rapid growth of parameters and computa- tional complexity, making it challenging to deploy DNNs on resource-constrained edge devices. In response to the challenge, network quantization proposes to represent the full-precision models, i.e., floating-point parameters and ac- tivations, using low-bit integers, resulting in a high compres- sion rate and an inference-acceleration rate [24]. These meth- *Equal contribution. Email: [email protected] †Corresponding author. Email: [email protected] CIFAR-10 CIFAR-100 ImageNet406080100T op-1 Accuracy(%)IntraQ HAST(Ours) IntraQ(Real Data)(a) 3-bit quantization. CIFAR-10 CIFAR-100 ImageNet6080100T op-1 Accuracy(%)IntraQ HAST(Ours) IntraQ(Real Data) (b) 4-bit quantization. Figure 1. Performance of the proposed HAST on three datasets compared with the state-of-the-art method IntraQ [50] and the method fine-tuning with real data [50]. HAST quantizes ResNet-20 on CIFAR-10/CIFAR-100 and ResNet-18 on ImageNet to 3-bit (left) and 4-bit (right), achieving performance comparable to the method fine-tuning with real data. ods implicitly assume that the training data of full-precision models are available for the process of network quantization. However, the training data, e.g., medical records, can be inaccessible due to privacy and security issues. Such a prac- tical and challenging scenario has led to the development ofzero-shot quantization (ZSQ) [2], quantizing networks without accessing the training data. Many efforts have been devoted to ZSQ [2, 19, 38, 46, 48, 50]. In ZSQ, some works perform network quantization by weight equalization [38], bias correction [1], or weight round- ing strategy [19], at the cost of some performance degrada- tion. To promote the performance of quantized models, ad- vanced works propose to leverage synthetic data for network quantization [2, 7, 46, 48, 50]. Specifically, they fine-tune quantized models with data synthesized using full-precision models, achieving promising improvement in performance. Much attention has been paid to the generation of syn- thetic sample, since high-quality synthetic samples lead to high-performance quantized models [2, 46]. Recent works employ generative models to synthesize data with fruitful approaches, considering generator design [51], boundary sample generation [6], adversarial training scheme [35], 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. 24417 IntraQ IntraQ(Real Data) HAST(Ours) 020406080100T op-1 Accuracy(%)Synthetic data Real training data Real test data(a) Performance of converged 3-bit ResNet-20. 0.4 0.6 0.8 1.0 Sample difficulty0.02.04.06.0T est error rate(%)IntraQ IntraQ(Real Data) HAST(Ours) (b) Variation of error rate with sample difficulty. 0.0 0.2 0.4 0.6 0.8 1.0 Sample difficulty103 101 FractionIntra IntraQ(Real Data) HAST(Ours) (c) Variation of fraction with sample difficulty. Figure 2. Analysis on synthetic data. (a) Performance of converged 3-bit ResNet-20. We quantize ResNet-20 to 3-bit using IntraQ [50], Real Data [50], and Our HAST, respectively, The top-1 accuracy on both training data (synthetic data for ZSQ methods) and test data is reported. (b) The error rate of test samples with different difficulties. (c) Distribution visualization of sample difficulty using GHM [30]. For each converged quantized model, we randomly sample 10,000 synthetic/real samples and count the fraction of samples based on difficulty. Note that the y-axis uses a log scale since the number of samples with different difficulties can differ by order of magnitude. effective training strategy [7]. Since the quality of synthetic samples is typically limited by the generator [36], advanced works treat synthesizing samples as a problem of noise op- timization [2]. Namely, the noise distribution is optimized to approximate some specified properties of real data dis- tributions, such as batch normalization statistics (BNS) and inception loss (IL) [20]. To promote model performance, IntraQ [50] focuses on the property of synthetic samples and endows samples with heterogeneity, achieving state-of-the- art performance as depicted in Figure 1. Although existing ZSQ methods achieve considerable performance gains by leveraging synthetic samples, there is still a significant performance gap between models trained with synthetic data and those trained with real data [50]. To reduce the performance gap, we investigate the difference between real and synthetic data. Specifically, we study the difference in generalization error between models trained with real data and those trained with synthetic data. Our experimental results show that synthetic data lead to larger generalization errors than real data, as illustrated in Figure 2a. Namely, synthetic sample lead to a more significant gap between training and test accuracy than real data. We conjecture that the performance gap stems from the misclassification of hard test samples. To verify the conjec- ture, we conduct experiments to study how model perfor- mance varies with sample difficulty, where GHM [30] is employed to measure the difficulty of samples quantitatively. The results shown in Figure 2b demonstrate that, on difficult samples, models trained with synthetic data perform worse than those trained with real data. This may result from that synthetic samples are easy to fit, which is consistent with the observation on inception loss of synthetic data [33]. We ver- ify the assumption through a series of experiments, where we count the fraction of samples of different difficulties using GHM [30]. The results are reported in Figure 2c, where we observe a severe missing of hard samples in synthetic sam- ples compared to real data. Consequently, quantized models fine-tuned with these synthetic data may fail to generalize well on hard samples in the test set.In light of conclusions drawn from Figure 2, the samples synthesized for fine-tuning quantized models in ZSQ should be hard to fit. To this end, we propose a novel HArd sample Synthesizing and Training (HAST) scheme. The insight of HAST has two folds: a) The samples constructed for fine- tuning models should not be easy for models to fit; b) The features extracted by full-precision and the quantized model should be similar. To this end, in the process of synthesizing samples, HAST pays more attention to hard samples in a re- weighting manner, where the weights are equal to the sample difficulty introduced in GHM [30]. Meanwhile, in the fine- tuning process, HAST further promotes the sample difficulty on the fly and aligns the features between the full-precision and quantized models. To verify the effectiveness of HAST, we conduct compre- hensive experiments on three datasets under two quantization precisions, following settings used in [50]. Our experimental results show that HAST using only 5,120 synthetic samples outperforms previous state-of-the-art method [50] and even achieves performance comparable with quantization with real data. Our main contributions can be summarized as follows: •We observe that the performance degradation in zero- shot quantization is attributed to the lack of hard sam- ples. Namely, the synthetic samples used in existing ZSQ methods are easily fitted by quantized models, dis- tinguishing models trained on synthetic samples from those trained on real data, as depicted in Figure 2. •Built upon our empirical observation, we propose a novel HArd sample Synthesizing and Training (HAST) scheme to promote the performance of ZSQ. Specifi- cally, HAST generates hard samples and further pro- motes the sample difficulty on the fly when training models, paired with a feature alignment constraint to ensure the similarity of features extracted by these two models, as summerized in Algorithm 1. •Extensive experiments demonstrate the superiority of HAST over existing ZSQ methods. More specifically, 24418 HAST using merely 5,120 synthetic samples outper- forms the previous state-of-the-art method and achieves performance comparable to models fine-tuned using real training data, as shown in Figure 1.
Levy_SeaThru-NeRF_Neural_Radiance_Fields_in_Scattering_Media_CVPR_2023	Abstract Research on neural radiance fields (NeRFs) for novel view generation is exploding with new models and exten- sions. However, a question that remains unanswered is what happens in underwater or foggy scenes where the medium strongly influences the appearance of objects. Thus far, NeRF and its variants have ignored these cases. However, since the NeRF framework is based on volumetric render- ing, it has inherent capability to account for the medium’s effects, once modeled appropriately. We develop a new ren- dering model for NeRFs in scattering media, which is based on the SeaThru image formation model, and suggest a suit- able architecture for learning both scene information and medium parameters. We demonstrate the strength of our method using simulated and real-world scenes, correctly rendering novel photorealistic views underwater. Even more excitingly, we can render clear views of these scenes, removing the medium between the camera and the scene and reconstructing the appearance and depth of far objects, which are severely occluded by the medium. Our code and unique datasets are available on the project’s website.	1. Introduction The pioneering work of Mildenhall et al. [25] on Neural Radiance Fields (NeRFs) has tremendously advanced the field of Neural Rendering, due to its flexibility and unprece- dented quality of synthesized images. Yet, the formulations of the original NeRF [25] and its followup variants assumethat images were acquired in clear air, i.e., in a medium that does not scatter or absorb light in a significant man- ner and that the acquired image is composed solely of the object radiance. The NeRF formulation is based on volu- metric rendering equations that take into account sampled points along 3D rays. Assuming a clear air environment, an implicit assumption, which is often enforced explicitly with dedicated loss components [5], is that a single opaque (high density) object is encountered per ray, with zero density be- tween the camera and the object. In stark contrast to clear air case, when the medium is absorbing and / or scattering (e.g., haze, fog, smog, and all aquatic habitats), the volume rendering equation has a true volumetric meaning, as the entire volume, and not only the object, contributes to image intensity. As the NeRF model estimates color and density at every point of a scene, it lends itself perfectly to general volumetric rendering, given that the appropriate rendering model is used. Here, we bridge this gap with SeaThru-NeRF , a framework that incorporates a rendering model that takes into account scattering media. This is achieved by assigning separate color and density parameters to the object (scene) and the medium, within the NeRF framework. Our approach adopts the SeaThru un- derwater image formation model [1, 3] to account for scat- tering media. SeaThru is a generalization of the standard wavelength-independent attenuation (e.g., fog) image for- mation model, where twodifferent wideband coefficients are used to represent the medium, which is more accurate when attenuation is wavelength-dependent (as in all wa- 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. 56 ter bodies and under some atmospheric conditions). In our model, the medium parameters are separate per color chan- nel, and are learned functions of the viewing angles, enforc- ing them to be constant only along 3D rays in the scene. Attempting to optimize existing NeRFs on scenes with scattering medium results in cloud-like objects floating in space, while our formulation enables the network to learn the correct representation of the entire 3D volume, that con- sists of both the scene and the medium. Our experiments demonstrate that SeaThru-NeRF produces state-of-the-art photorealistic novel view synthesis on simulated and chal- lenging real-world scenes (see Fig. 1) with scattering media, that include complex geometries and appearances. In addi- tion, it enables: 1.Color restoration of the scenes as if they were not im- aged through a medium, as our modeling allows full sepa- ration of object appearance from the medium effects. 2.Estimation of 3D scene structure which surpasses that of structure-from-motion (SFM) or current NeRFs, espe- cially in far areas of bad visibility, as we jointly reconstruct and reason for the geometry and medium. 3.Estimation of wideband medium parameters , which are informative properties of the captured environment, and potentially allowing simulation under different conditions.
Kumari_Multi-Concept_Customization_of_Text-to-Image_Diffusion_CVPR_2023	Abstract While generative models produce high-quality images of concepts learned from a large-scale database, a user often wishes to synthesize instantiations of their own concepts (for example, their family, pets, or items). Can we teach a model to quickly acquire a new concept, given a few examples? Fur- thermore, can we compose multiple new concepts together? We propose Custom Diffusion, an efficient method for aug- menting existing text-to-image models. We find that only op- timizing a few parameters in the text-to-image conditioning mechanism is sufficiently powerful to represent new concepts while enabling fast tuning ( ∼6minutes). Additionally, we can jointly train for multiple concepts or combine multi- ple fine-tuned models into one via closed-form constrained optimization. Our fine-tuned model generates variations of multiple new concepts and seamlessly composes them with existing concepts in novel settings. Our method outperforms or performs on par with several baselines and concurrent works in both qualitative and quantitative evaluations, while being memory and computationally efficient.	1. Introduction Recently released text-to-image models [53, 57, 60, 79] have represented a watershed year in image generation. By simply querying a text prompt, users are able to generate images of unprecedented quality. Such systems can generate a wide variety of objects, styles, and scenes – seemingly “anything and everything”. However, despite the diverse, general capability of such models, users often wish to synthesize specific concepts from their own personal lives. For example, loved ones such as family, friends, pets, or personal objects and places, such as a new sofa or a recently visited garden, make for intriguing concepts. As these concepts are by nature personal, they are unseen during large-scale model training. Describing these concepts after the fact, through text, is unwieldy and unable to produce the personal concept with sufficient fidelity. This motivates a need for model customization . Given the few user-provided images, can we augment existing text-to- image diffusion models with the new concept (for example, their pet dog or a “moongate” as shown in Figure 1)? The fine-tuned model should be able to generalize and compose them with existing concepts to generate new variations. This 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. 1931 poses a few challenges – first, the model tends to forget [12, 35, 52] or change [34, 41] the meanings of existing concepts: e.g., the meaning of “moon” being lost when adding the “moongate” concept. Secondly, the model is prone to overfit the few training samples and reduce sampling variations. Moreover, we study a more challenging problem, compo- sitional fine-tuning – the ability to extend beyond tuning for a single, individual concept and compose multiple concepts together, e.g., pet dog in front of moongate (Figure 1). Improving compositional generation has been studied in re- cent works [40]. But composing multiple new concepts poses additional challenges, such as mixing unseen concepts. In this work, we propose a fine-tuning technique, Custom Diffusion for text-to-image diffusion models. Our method is computationally and memory efficient. To overcome the above-mentioned challenges, we identify a small subset of model weights, namely the key and value mapping from text to latent features in the cross-attention layers [5, 70]. Fine-tuning these is sufficient to update the model with the new concept. To prevent model forgetting, we use a small set of real images with similar captions as the target images. We also introduce augmentation during fine-tuning, which leads to faster convergence and improved results. To inject multiple concepts, our method supports training on both simultaneously or training them separately and then merging. We build our method on Stable Diffusion [1] and experi- ment on various datasets with as few as four training images. For adding single concepts, our method shows better text alignment and visual similarity to the target images than con- current works and baselines. More importantly, our method can compose multiple new concepts efficiently, whereas con- current methods struggle and often omit one. Finally, our method only requires storing a small subset of parameters (3%of the model weights) and reduces the fine-tuning time (6 minutes on 2 A100 GPUs, 2−4×faster compared to concurrent works). Full version of the paper is available at https://arxiv.org/abs/2212.04488.
Kil_Towards_Unified_Scene_Text_Spotting_Based_on_Sequence_Generation_CVPR_2023	Abstract Sequence generation models have recently made signif- icant progress in unifying various vision tasks. Although some auto-regressive models have demonstrated promising results in end-to-end text spotting, they use specific detec- tion formats while ignoring various text shapes and are limited in the maximum number of text instances that can be detected. To overcome these limitations, we propose a UNIfied scene Text Spotter, called UNITS. Our model unifies various detection formats, including quadrilater- als and polygons, allowing it to detect text in arbitrary shapes. Additionally, we apply starting-point prompting to enable the model to extract texts from an arbitrary start- ing point, thereby extracting more texts beyond the num- ber of instances it was trained on. Experimental results demonstrate that our method achieves competitive perfor- mance compared to state-of-the-art methods. Further anal- ysis shows that UNITS can extract a larger number of texts than it was trained on. We provide the code for our method at https://github.com/clovaai/units.	1. Introduction End-to-end scene text spotting, which can jointly de- tect and recognize text from an image at once, has recently gained significant attention. This work has practical appli- cations in various fields such as visual navigation, visual question answering, and document image understanding. In this paper, we formulate scene text spotting as a se- quence generation task. This approach casts the vision task as a language modeling task conditioned on the image and text prompt [6,7,27,38]. By formulating the output as a se- quence of discrete tokens, the vision task can be performed by generating a sequence through an auto-regressive trans- former decoder. Recent studies have attempted to integrate various tasks into an auto-regressive model, including text spotting. SPTS [28] treated all detection formats as the sin- *Corresponding author. †This work was done while the authors were at Naver Cloud. (a) Single central point format. (b) Bounding box format. (c) Quadrilateral format. (d) Polygonal format. Figure 1. Various types of detection formats. The green line rep- resents the boundary shape of the detection format, and the red dot represents the points used for the corresponding format. gle central point and predicted the coordinate tokens of the central point and word tokens auto-regressively. However, there are several challenges associated with applying the se- quence generation approach directly to scene text spotting. As illustrated in Fig. 1, there are various methods to indi- cate the location and boundaries of text instances, and each method has its own trade-offs in terms of annotation costs and potential applicability. For instance, a single central point or bounding box annotation has a relatively low anno- tation cost and is appropriate when the detailed shape infor- mation is not necessary. However, in many fields where text spotting is applied, handling text location information as only a single point might be insufficient. As shown in Fig. 2, in the scene text editing [17, 32], which converts the text in the image to desired text while preserving the text style, de- tailed text shape extraction is required. Therefore, for such This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15223 (a) Original image. (b) Result of image translation. Figure 2. Example of image translation using scene text editing. tasks, outputting text location information in quadrilateral or polygonal formats beyond a single point is necessary. Similarly, in visual document understanding, which utilizes optical character recognition as a pre-processing step, more detailed detection information of texts could be useful. In summary, since each detection format has its own trade- offs, it is better to cover all detection formats instead of relying on only one. Also, conventional sequence generation models have a limitation in terms of generating sequences longer than the maximum length the model has been trained on. This limi- tation is particularly problematic for scene text spotting, as it requires generating more points than bounding boxes and generating text transcriptions in addition to object classes. For text-rich images such as documents, this limitation be- comes a bottleneck in extracting all the texts. To address these challenges in applying the sequence generation method to text spotting, we propose a novel end- to-end UNIfied scene Text Spotter, called UNITS, which aims to overcome these limitations. We unify various de- tection formats, such as quadrilateral and polygon, into the model, allowing it to detect arbitrary-shaped text areas. This approach enables the model to learn different annotation formats together using a single unified model, and it can extract all detection formats. We use prompts to unify var- ious detection formats, enabling a single model to predict the coarse or fine-grained location information of text in- stances. Moreover, we introduce starting-point prompting, which enables the model to extract texts from arbitrary starting points, allowing it to generate longer sequences than the maximum decoding length. This approach enhances the model’s efficiency and enables it to extract more text than the number of text boxes it has been trained on. In addition, we employ a multi-way transformer decoder from the mixture-of-experts (MoE) [4, 9, 37]. This decoder separates each time stamp into detection and recognition, allowing the model to converge faster by guiding the model to generate a detection or a recognition token and assigningan expert accordingly. Each block of the multi-way trans- former consists of shared attention modules and two task experts: detection and recognition experts. Experimental results demonstrate that our proposed method achieves competitive performance on text spotting benchmarks while extracting texts in various detection for- mats using a single unified model. Our main contributions are: • We propose a novel sequence generation-based scene text spotting method that can extract arbitrary-shaped text areas by unifying various detection formats. • Our model can extract more texts than the decoder length allows using the starting-point prompt, which generalizes the model to spot more texts than it has been trained on. • Experimental results demonstrate that our method achieves competitive performance on text spotting benchmarks and provides additional functionalities
Liu_Reducing_the_Label_Bias_for_Timestamp_Supervised_Temporal_Action_Segmentation_CVPR_2023	Abstract Timestamp supervised temporal action segmentation (TSTAS) is more cost-effective than fully supervised coun- terparts. However, previous approaches suffer from severe label bias due to over-reliance on sparse timestamp an- notations, resulting in unsatisfactory performance. In this paper, we propose the Debiasing-TSTAS (D-TSTAS) frame- work by exploiting unannotated frames to alleviate this bias from two phases: 1) Initialization. To reduce the depen- dencies on annotated frames, we propose masked times- tamp predictions (MTP) to ensure that initialized model captures more contextual information. 2) Refinement. To overcome the limitation of the expressiveness from sparsely annotated timestamps, we propose a center-oriented times- tamp expansion (CTE) approach to progressively expand pseudo-timestamp groups which contain semantic-rich mo- tion representation of action segments. Then, these pseudo- timestamp groups and the model output are used to iter- atively generate pseudo-labels for refining the model in a fully supervised setup. We further introduce segmental con- fidence loss to enable the model to have high confidence predictions within the pseudo-timestamp groups and more accurate action boundaries. Our D-TSTAS outperforms the state-of-the-art TSTAS method as well as achieves compet- itive results compared with fully supervised approaches on three benchmark datasets.	1. Introduction Analyzing and understanding human actions in videos is very important for many applications, such as human- robot interaction [14] and healthcare [33]. Recently, sev- eral approaches have been very successful in locating and analyzing activities in videos, including action localization [12,29,37,50], action segmentation [7,11,20,23], and action recognition [9, 27, 39, 41, 47]. *These authors contributed equally to this work. †Corresponding author. Figure 1. (a) The TSTAS task aims to segment actions in videos by timestamp annotations. (b) An example of focus bias: exist- ing initialization methods prefer to predict frames similar to the annotated timestamp, leading to incorrectly identifying dissimilar frames within the segment and similar frames outside the segment. (c) An example of representation bias: the frames of complex action share large semantic variance, e.g. the process of cutting cheese includes both taking out and cutting, resulting in the mod- els that rely on sparse timestamps to produce biased pseudo-labels. Despite the success of previous approaches, they rely on fully temporal supervision, where the start and end frames of each action are annotated. As a lightweight alternative, many researchers have started exploring timestamp super- vised temporal action segmentation (TSTAS), where each action segment is annotated with only one frame in the untrimmed video, as shown in Fig. 1 (a). Most previous approaches follow a two-step pipeline by initializing and then iteratively refining the model with gen- erated pseudo-labels. However, relying only on supervised signals of sparse single-timestamp annotations, these meth- ods fail to learn the semantics of entire action segments, which is referred to as label bias in this paper, including fo- cus bias andrepresentation bias . Specifically, to initialize the segmentation model, the previous methods [17, 28, 32] adopt the Naive approach proposed in [28] (referred to as Naive), which computes the loss at the annotated frames for training. However, utilizing only the annotated frames leads to focus bias , where the initialized model tends to fo- 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. 6503 cus on frames distributed over segments of various action categories similar to the annotated frames [51] (shown in Fig. 1 (b)). During refining the segmentation model, to utilize the unlabeled frames, typical methods generate hard [17,28,51] or soft weighted [32] pseudo-labels that are used to re- fine the model like fully supervised methods. To gener- ate pseudo-labels, the above methods depend on the times- tamps that contain semantic or relative position informa- tion. Despite the success of such refined models, these ap- proaches suffer from representation bias that single-frame fails to represent the entire action segment for complex ac- tions with the large semantic variance of various frames (il- lustrated in Fig. 1 (c)). Moreover, when refining the seg- mentation model by these biased pseudo-labels, represen- tation bias can be accumulated and expanded, resulting in unsatisfactory performance that is still a gap compared to fully supervised approaches. In this paper, we propose a novel framework called Debiasing-TSTAS (D-TSTAS) to reduce label bias, con- taining masked timestamp predictions and center-oriented timestamp expansion approaches. During the initialization phase, we aim to alleviate the focus bias problem by prop- agating timestamp supervision information to the contex- tual frames of the timestamps. In contrast to previous ap- proaches, we propose the Masked Timestamp Predictions (MTP) approach that masks the input features of times- tamps to remove the dependencies of the model on anno- tated frames. In this way, the initialized model is forced to reconstruct the annotated frames in corresponding out- put features and then predict their action categories by con- textual information of timestamps. Furthermore, to capture the semantics of both timestamps and contextual frames, we adopt the Naive after our MTP. While our initialized model can reduce focus bias by capturing more contextual information, it does not guar- antee that the generated pseudo-labels avoid represen- tation bias during refining. Inspired by query expan- sion, we propose a Center-oriented Timestamp Expansion (CTE) approach for obtaining potential trustworthy unla- beled frames, which we refer to as pseudo-timestamps, to progressively expand the pseudo-timestamp groups that contain more semantics than single annotated timestamps. More specifically, it consists of three steps: 1) In the gen- erating step, we generate pseudo-labels by current pseudo- timestamp groups and the model output. 2) In the updating step, we choose the semantically dissimilar center frames of each segment in the pseudo-labels as pseudo-timestamps to expand the pseudo-timestamp groups. 3) In the segmenting step, the model is refined by pseudo-labels and our segmen- tal confidence loss, which smooths the predicted probabil- ities in each action segment and maintains high confidence within pseudo-timestamp groups. The above steps are ex-ecuted several times alternately to improve the model pre- diction in a lazy manner instead of generating pseudo-labels per epoch in previous methods [28, 51]. Our contributions can be summarised as follows: • We study the label bias in the TSTAS task and propose a novel D-TSTAS framework to reduce both focus and representation bias. • Our masked timestamp predictions approach is the first attempt to alleviate the dependencies on timestamps, promoting the model to capture contextual informa- tion. Coupling MTP and Naive as a general solution is used to initialize the model in the TSTAS. • Compared to sparsely annotated timestamps, our center-oriented timestamp expansion approach pro- gressively expands pseudo-timestamp groups to con- tain semantic-rich motion representations of action segments. • The proposed D-TSTAS not only outperforms state-of- the-art TSTAS approaches but also achieves competi- tive results compared with fully supervised approaches on three benchmark datasets.
Kong_Efficient_Frequency_Domain-Based_Transformers_for_High-Quality_Image_Deblurring_CVPR_2023	Abstract We present an effective and efﬁcient method that explores the properties of Transformers in the frequency domain for high-quality image deblurring. Our method is motivated by the convolution theorem that the correlation or convo- lution of two signals in the spatial domain is equivalent to an element-wise product of them in the frequency domain. This inspires us to develop an efﬁcient frequency domain- based self-attention solver (FSAS) to estimate the scaled dot-product attention by an element-wise product operation instead of the matrix multiplication in the spatial domain. In addition, we note that simply using the naive feed-forward network (FFN) in Transformers does not generate good de- blurred results. To overcome this problem, we propose a simple yet effective discriminative frequency domain-based FFN (DFFN), where we introduce a gated mechanism in the FFN based on the Joint Photographic Experts Group (JPEG) compression algorithm to discriminatively determine which low- and high-frequency information of the features should be preserved for latent clear image restoration. We formulate the proposed FSAS and DFFN into an asymmetri- cal network based on an encoder and decoder architecture, where the FSAS is only used in the decoder module for better image deblurring. Experimental results show that the pro- posed method performs favorably against the state-of-the-art approaches.	1. Introduction Image deblurring aims to restore high-quality images from blurred ones. This problem has achieved signiﬁcant progress due to the development of various effective deep models with large-scale training datasets. Most state-of-the-art methods for image deblurring are mainly based on deep convolutional neural networks (CNNs). The main success of these methods is due to developing kind- s of network architectural designs, for example, the multi- Corresponding author: Jiangxin Dong and Jinshan Pan. /uni00000014/uni00000015/uni00000013 /uni00000014/uni00000019/uni00000013 /uni00000015/uni00000013/uni00000013 /uni00000015/uni00000017/uni00000013 /uni00000015/uni0000001b/uni00000013 /uni00000016/uni00000015/uni00000013 /uni00000016/uni00000019/uni00000013 /uni00000017/uni00000013/uni00000013 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 DeepRFT+ Stripformer Restormer MIMO-Unet+MAXIM-3S IPTFigure 1. Comparisons of the proposed method and state-of-the- art ones on the GoPro dataset [16] in terms of accuracy, ﬂoating point operations (FLOPs), and network parameters. The circle size indicates the number of the network parameter. scale [4,16,22] or multi-stage [31,32] network architectures, generative adversarial learning [11, 12], physics model in- spired network structures [6 –8, 17, 33], and so on. As the basic operation in these networks, the convolution opera- tion is a spatially-invariant local operation, which does not model the spatially variant properties of the image contents. Most of them use larger and deeper models to remedy the limitation of the convolution. However, simply increasing the capacity of deep models does not always lead to better performance as shown in [17, 33]. Different from the convolution operation that models the local connectivity, Transformers are able to model the global contexts by computing the correlations of one token to all other tokens. They have been shown to be an effective ap- proach in lots of high-level vision tasks and also have great potential to be the alternatives of deep CNN models. In im- age deblurring, the methods based on Transformers [27, 30] also achieve better performance than the CNN-based meth- ods. However, the computation of the scaled dot-product attention in Transformers leads to quadratic space and time complexity in terms of the number of tokens. Although using smaller and fewer tokens can reduce the space 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. 5886 time complexity, such strategy cannot model the long-range information of features well and usually leads to signiﬁcant artifacts when handling high-resolution images, which thus limits the performance improvement. To alleviate this problem, most approaches use the down- sampling strategy to reduce the spatial resolution of fea- tures [26]. However, reducing the spatial resolution of fea- tures will cause information loss and thus affect the image deblurring. Several methods reduce the computational cost by computing the scaled dot-product attention in terms of the number of features [29, 30]. Although the computational cost is reduced, the spatial information is well not explored, which may affect the deblurring performance. In this paper, we develop an effective and efﬁcient method that explores the properties of Transformers for high-quality image deblurring. We note that the scaled dot-product at- tention computation is actually to estimate the correlation of one token from the query and all the tokens from the key. This process can be achieved by a convolution operation when rearranging the permutations of tokens. Based on this observation and the convolution theorem that the convolu- tion in the spatial domain equals a point-wise multiplication in the frequency domain, we develop an efﬁcient frequen- cy domain-based self-attention solver (FSAS) to estimate the scaled dot-product attention by an element-wise product operation instead of the matrix multiplication. Therefore, the space and time complexity can be reduced to O(N) O(NlogN)for each feature channel, where Nis the num- ber of the pixels. In addition, we note that simply using the feed-forward network (FFN) by [30] does not generate good deblurred results. To generate better features for latent clear image restoration, we develop a simple yet effective discrimina- tive frequency domain-based FFN (DFFN). Our DFFN is motivated by the Joint Photographic Experts Group (JPEG) compression algorithm. It introduces a gated mechanism in the FFN to discriminatively determine which low- and high-frequency information should be preserved for latent clear image restoration. We formulate the proposed FSAS and DFFN into an end- to-end trainable network based on an encoder and decoder architecture to solve image deblurring. However, we ﬁnd that as features of shallow layers usually contain blur effects, applying the scaled dot-product attention to shallow features does not effectively explore global clear contents. As the features from deep layers are usually clearer than those from shallow layers, we develop an asymmetric network architec- ture, where the FSAS is only used in the decoder module for better image deblurring. We analyze that the exploring properties of Transformers in the frequency domain is able to facilitate blur removal. Experimental results demonstrate that the proposed method generates favorable results against state-of-the-art methods in terms of accuracy and efﬁciency(Figure 1). The main contributions are summarized as follows: We develop an efﬁcient frequency domain-based self- attention solver to estimate the scaled dot-product at- tention. Our analysis demonstrates that using the fre- quency domain-based solver reduces the space and time complexity and is much more effective and efﬁcient. We propose a simple yet effective discriminative fre- quency domain-based FFN based on the JPEG compres- sion algorithm to discriminatively determine which low and high-frequency information should be preserved for latent clear image restoration. We develop an asymmetric network architecture based on an encoder and decoder network, where the frequen- cy domain-based self-attention solver is only used in the decoder module for better image deblurring. We analyze that the exploring properties of Transform- ers in the frequency domain is able to facilitate blur removal and show that our approach performs favor- ably against state-of-the-art methods.
Liu_Hierarchical_Prompt_Learning_for_Multi-Task_Learning_CVPR_2023	Abstract Vision-language models (VLMs) can effectively transfer to various vision tasks via prompt learning. Real-world sce- narios often require adapting a model to multiple similar yet distinct tasks. Existing methods focus on learning a specific prompt for each task, limiting the ability to exploit poten- tially shared information from other tasks. Naively training a task-shared prompt using a combination of all tasks ig- nores fine-grained task correlations. Significant discrepan- cies across tasks could cause negative transferring. Consid- ering this, we present Hierarchical Prompt (HiPro) learn- ing, a simple and effective method for jointly adapting a pre-trained VLM to multiple downstream tasks. Our method quantifies inter-task affinity and subsequently constructs a hierarchical task tree. Task-shared prompts learned by in- ternal nodes explore the information within the correspond- ing task group, while task-individual prompts learned by leaf nodes obtain fine-grained information targeted at each task. The combination of hierarchical prompts provides high-quality content of different granularity. We evaluate HiPro on four multi-task learning datasets. The results demonstrate the effectiveness of our method.	1. Introduction Vision-language pre-training [23, 34, 49, 71, 74] has re- cently shown great potential to leverage human language for addressing a wide range of downstream recognition tasks. Vision-language models (VLMs), e.g., CLIP [49] and ALIGN [23], align embeddings of images and texts from massive web data, encouraging the matching image- text pair to be similar and pushing away the unmatched pair [6, 19]. During inference, the task-relevant content in text modality can be provided to query the latent knowledge of the pre-trained VLMs for facilitating visual recognition. *Equal contribution. †Corresponding author. −202468101202468wrandwindwshrwavg 0.0110.0290.0560.130.310.7725>5(a) Train loss −202468101202468wrandwindwshrwavg 2425252628303440>40 (b) Test error Figure 1. Task-shared prompt vs. task-individual prompt on multi-task learning. We visualize ( a) train loss and ( b) test er- ror surface [15] for classifier weights ( wrand,wind, and wshr), which synthesized from the random initialization prompt, task- individual prompt, and task-shared prompt, respectively, on one of the target tasks (i.e., the Art task of the Office-Home dataset [65]). The task-individual prompt is only trained on this task. The task- shared prompt is trained on the combination of all tasks. The av- erage weights ( wavg=1 2(wshr+wind)) can perform well to test samples. More details refer to the supplementary materials. The provided task-relevant texts, often constructed by theprompt template and category words, can significantly influence the recognition performance. Prompt engineer- ing[23, 49], i.e., manually designing prompts, is a straight- forward way to obtain meaningful prompts for adapting VLMs. However, it inevitably introduces artificial bias and relies on time-consuming attempts [49]. Recent advances on prompt learning [79,80] show an alternative way, which This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10888 Task: Satellite Images (EuroSAT)Task-IndividualPromptAcc(%)a photo of a {class}texture.40.54Task: DescribableTexture (DTD)Task-SharedPromptAcc(%)a photo with repeating {class}pattern.39.72Individual Prompt + SharedPromptAcc(%)a photo of a {class}texture.a photo with repeating {class}pattern.42.73Task-IndividualPromptAcc(%)a centered satellite photo of a {class}.36.68Task-SharedPromptAcc(%)a photo with repeating {class}pattern.36.75Individual Prompt + SharedPromptAcc(%)a centered satellite photo of a {class}.a photo with repeating {class}pattern.39.57Figure 2. The benefits of prompt learning with multiple tasks. Note that DTD [9] dataset and EuroSAT [45] dataset employ the same task-shared prompt. Task-individual prompt and task-shared prompt can represent different contents of recognition tasks. Ensembling their zero-shot classifiers can improve performance. aims to learn the appropriate soft prompt in a data-driven manner on each downstream task. With few training data, prompt learning has shown considerable improvement com- pared with the hand-crafted prompt. Despite substantial progress, existing approaches [79– 81] still focus on adapting VLMs to the individual task. However, challenges in realistic situations demand adapting a model to several similar but different tasks, also known as the problem of multi-task learning [20, 75]. More im- portantly, current methods learn the specific prompt corre- sponding to each task, which can not leverage information in other tasks to benefit individual tasks. Actually, the trans- ferred prompt can be reused for similar tasks. For exam- ple, “a photo of a {class}.” is a general prompt for most recognition tasks. Specifically, as shown in Figure 2, for two distinct tasks, i.e., texture images and satellite images, a well-designed prompt can leverage the potential connec- tions across them. This paper explores how to simultaneously adapt a pre- trained VLM to multiple target tasks through prompting. A straightforward way is to learn the same prompt for all tasks. However, this na ¨ıve approach ignores the charac- teristics of each task and fails to achieve the optimum on each task. Nevertheless, we found that the task-shared prompt can significantly complement the prompt designed (or learned) individually for each task . As shown in Fig- ure 2, the task-individual (hand-crafted) prompt captures the fine-grained content of each task. The task-shared (hand-crafted) prompt represents the general content across tasks. The combination of task-shared and task-individual prompts can embrace both general and fine-grained content to enhance recognition. Another perspective is provided for an in-depth explana- tion. In Figure 1a, we see that, the classifier weights synthe- sized from the task-individual prompt ( trained on the indi- vidual target task) have lower training loss than the weights from the task-shared prompt ( trained on the combination of all tasks). However, the performance of task-individual prompt on the test set is poor (Figure 1b), which implies that the task-individual prompt has the risk of over-fitting.Meanwhile, the task-shared prompt, generalizing on various tasks, can be considered as a regularization to avoid over- fitting. Averaging weights from the task-shared prompt and the task-individual prompt can improve the performance on test data (Figure 1b). Although similar tasks can facilitate each other by shar- ing knowledge, we can notassume all the offered tasks can benefit from training together. Significant discrepancies across tasks could lead to poor performance, also known as negative-transfer [73]. On the other hand, even for the ideal case, i.e., there exists the same beneficial prompt across all tasks, only learning the global (coarse-grained) task-shared prompt neglects the information transferred within some fine-grained task groups. To address this problem, we present Hierarchical Prompt (HiPro) learning to capture multi-grained shared informa- tion while mitigating negative transfer between dissimilar tasks. Our HiPro constructs a hierarchical task tree by agglomerative hierarchical clustering based on inter-task affinity. Specifically, the internal node of the tree repre- sents a task group containing a cluster of similar tasks (at descendant leaves). Meanwhile, dissimilar tasks would be divided into different sub-trees, mitigating conflict. For each node, HiPro learns a corresponding prompt to cap- ture the general information of the fine-grained task group. Our HiPro learns not only task-individual prompts (for leaf nodes) but also multi-grained task-share prompts (for non- leaf nodes). For inference, HiPro combines various weights generated from learned prompts, leveraging the information in all tasks to improve the performance of the individual task. Comprehensive experiments are constructed to validate the effectiveness of our method. HiPro works well on a large-scale multi-task learning benchmark consisting of di- verse visual recognition tasks. Compared with the existing prompt learning methods [40,79,80], HiPro has a significant improvement demonstrating the benefit of learning prompts with multiple tasks. Additional visualizations are also pro- vided for analysis. 10889
Litrico_Guiding_Pseudo-Labels_With_Uncertainty_Estimation_for_Source-Free_Unsupervised_Domain_Adaptation_CVPR_2023	Abstract Standard Unsupervised Domain Adaptation (UDA) meth- ods assume the availability of both source and target data during the adaptation. In this work, we investigate Source- free Unsupervised Domain Adaptation (SF-UDA), a specific case of UDA where a model is adapted to a target domain without access to source data. We propose a novel approach for the SF-UDA setting based on a loss reweighting strategy that brings robustness against the noise that inevitably af- fects the pseudo-labels. The classification loss is reweighted based on the reliability of the pseudo-labels that is measured by estimating their uncertainty. Guided by such reweight- ing strategy, the pseudo-labels are progressively refined by aggregating knowledge from neighbouring samples. Further- more, a self-supervised contrastive framework is leveraged as a target space regulariser to enhance such knowledge aggregation. A novel negative pairs exclusion strategy is pro- posed to identify and exclude negative pairs made of samples sharing the same class, even in presence of some noise in the pseudo-labels. Our method outperforms previous methods on three major benchmarks by a large margin. We set the new SF-UDA state-of-the-art on VisDA-C and DomainNet with a performance gain of +1.8% on both benchmarks and on PACS with +12.3% in the single-source setting and +6.6% in multi-target adaptation. Additional analyses demonstrate that the proposed approach is robust to the noise, which re- sults in significantly more accurate pseudo-labels compared to state-of-the-art approaches.	1. Introduction Deep learning methods achieve remarkable performance in visual tasks when the training and test sets share a similar distribution, while their generalisation ability on unseen data decreases in presence of the so called domain shift [18, 48]. Moreover, DNNs require a huge amount of labelled data to be trained on a new domain entailing a considerable cost for collecting and labelling the data. Unsupervised DomainAdaptation (UDA) approaches aim to transfer the knowledge learned on a labelled source domain to an unseen target domain without requiring any target label [2, 13, 22, 57]. Common UDA techniques have the drawback of requiring access to source data while they are adapting the model to the target domain. This may not always be possible in many applications, i.e. when data privacy or transmission bandwidth become critical issues. In this work, we focus on the setting of Source-free adaptation (SF-UDA) [37, 55, 61, 66], where source data is no longer accessible during adaptation, but only unlabelled target data is available. As a result, standard UDA methods cannot be applied in the SF-UDA setting, since they require data from both domains. Recently, several SF-UDA methods have been proposed, focusing on generative models [36,43], class prototypes [37], entropy-minimisation [37, 61], self-training [37] and auxil- iary self-supervised tasks [55]. Yet, generative models re- quire a large computational effort to generate images/features in the target style [36], entropy-minimisation methods of- ten lead to posterior collapse [37] and the performance of self-training solutions [37] suffer from noisy pseudo-labels. Self supervision and pseudo-labelling have also been intro- duced as joint SF-UDA strategies [55, 60], raising the issue of choosing a suitable pretext task and refining pseudo-labels. For instance, Chen et al. [5] propose to refine predictions within a self-supervised strategy. Yet, their work does not take into account the noise inherent in pseudo-labels, which leads to equally weighting all samples without explicitly ac- counting for their uncertainty. This may cause pseudo-labels with high uncertainty to still contributing in the classification loss, resulting in detrimental noise overfitting and thus poor adaptation. In this work, we propose a novel Source-free adaptation approach that builds upon an initial pseudo-labels assign- ment (for all target samples) performed by using the pre- trained source model, always assumed to be available. To obtain robustness against the noise that inevitably affects such pseudo-labels, we propose to reweight the classification loss based on the reliability of the pseudo-labels. We mea- sure it by estimating pseudo-labels uncertainty, after they 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. 7640 Figure 1. (a) We obtain the refined pseudo-label ˆyt(green circle with black outline) for the current sample by looking at pseudo-labels of neighbour samples. (b) Predictions from neighbours are used to estimate the uncertainty of ˆytby computing the weight wthrough Entropy ˆH. (c) A temporal queue Qestoring predictions at past T epochs, i.e. {e−T, ..., e −1}, is used within the contrastive framework to exclude pairs of samples sharing the same class from the list of negative pairs (query, key) . have been refined by knowledge aggregation from neigh- bours sample, as shown in Figure 1 (a). The introduced loss reweighting strategy penalises pseudo-labels with high uncertainty to guide the learning through reliable pseudo- labels. Differently from previous reweighting strategies, we reweight the loss by estimating the uncertainty of the refined pseudo-labels by simply analysing neighbours’ predictions, as shown in Figure 1 (b). The process of refining pseudo-labels necessarily requires a regularised target feature space in which neighbourhoods are composed of semantically similar samples, possibly shar- ing the same class. With this objective, we exploit an aux- iliary self-supervised contrastive framework. Unlike prior works, we introduce a novel negative pairs exclusion strat- egy that is robust to noisy pseudo-labels, by leveraging past predictions stored in a temporal queue. This allows us to identify and exclude negative pairs made of samples belong- ing to the same class, even if their pseudo-labels are noisy, as shown in Figure 1 (c). We benchmark our method on three major domain adap- tation datasets outperforming the state-of-the-art by a large margin. Specifically, on VisDA-C and DomainNet, we set the new state-of-the-art with 90.0% and 69.6% accuracy, with a gain of +1.8% in both cases, while on PACS we im- prove the only existing SF-UDA baseline [1] by +12.3% and +6.6% in single-source and multi-target settings. Ablation studies demonstrate the effectiveness of individual compo- nents of our pipeline in adapting the model from the source to the target domain. We also show how our method is able to progressively reduce the noise in the pseudo-labels, betterthan the state-of-the-art. To summarise, the main contributions of this work are: •We introduce a novel loss re-weighting strategy that evaluates the reliability of refined pseudo-labels by es- timating their uncertainty. This enables our method to mitigate the impact of the noise in the pseudo-labels. To the best of our knowledge, this is the first work that estimates the reliability of pseudo-labels after their refinement. •We propose a novel negative pairs exclusion strategy which is robust to noisy pseudo-labels, being able to identify and exclude those negative pairs composed of same-class samples. •We validate our method on three benchmark datasets, outperforming SOTA by a large margin, while also proving the potential of the approach in progressively refining the pseudo-labels. The remainder of the paper is organized as follows. In Section 2, we discuss related works in the literature. Sec- tion 3 describes the proposed method. Section 4 illustrates the experimental setup and reports the obtained results. Fi- nally, conclusions are drawn in Section 5 together with a discussion of limitations.
Karim_C-SFDA_A_Curriculum_Learning_Aided_Self-Training_Framework_for_Efficient_Source_CVPR_2023	Abstract Unsupervised domain adaptation (UDA) approaches fo- cus on adapting models trained on a labeled source do- main to an unlabeled target domain. In contrast to UDA, source-free domain adaptation (SFDA) is a more practi- cal setup as access to source data is no longer required during adaptation. Recent state-of-the-art (SOTA) methods on SFDA mostly focus on pseudo-label refinement based self-training which generally suffers from two issues: i) in- evitable occurrence of noisy pseudo-labels that could lead to early training time memorization, ii) refinement process requires maintaining a memory bank which creates a signif- icant burden in resource constraint scenarios. To address these concerns, we propose C-SFDA, a curriculum learn- ing aided self-training framework for SFDA that adapts effi- ciently and reliably to changes across domains based on se- lective pseudo-labeling. Specifically, we employ a curricu- lum learning scheme to promote learning from a restricted amount of pseudo labels selected based on their reliabili- ties. This simple yet effective step successfully prevents la- bel noise propagation during different stages of adaptation and eliminates the need for costly memory-bank based label refinement. Our extensive experimental evaluations on both image recognition and semantic segmentation tasks confirm the effectiveness of our method. C-SFDA is also applica- ble to online test-time domain adaptation and outperforms previous SOTA methods in this task.	1. Introduction Deep neural network (DNN) models have achieved re- markable success in various visual recognition tasks [16, 22, 41, 44]. However, even very large DNN models of- ten suffer significant performance degradation when there *Most of this work was done during Nazmul Karim’s internship with SRI International. Project Page: https://sites.google.com/ view/csfdacvpr2023/homeis a distribution or domain shift [54, 78] between training (source) and test (target) domains. To address the prob- lem of domain shifts, various Unsupervised Domain Adap- tation (UDA) [19, 31] algorithms have been developed over recent years. Most UDA techniques require access to la- beled source domain data during adaptation, which limits their application in many real-world scenarios, e.g. source data is private, or adaptation in edge devices with lim- ited computational capacity. In this regard, source-free do- main adaptation setting has recently gained significant in- terest [34, 35, 84], which considers the availability of only source pre-trained model and unlabeled target domain data. Recent state-of-the-art SFDA methods (e.g., SHOT [42], NRC [83], G-SFDA [85], AdaContrast [4]) mostly rely on the self-training mechanism that is guided by the source pre- trained model generated pseudo-labels (PLs). Label refine- ment using the knowledge of per-class cluster structure in feature space is recurrently used in these methods. At early stages of adaptation, the label information formulated based on cluster structure can be severely misleading or noisy; shown in Fig. 1. As the adaptation progresses, this label noise can negatively impact the subsequent cluster struc- ture as the key to learning meaningful clusters hinges on the quality of pseudo-labels itself. Therefore, the inevitable presence of label noise at early training time is a critical is- sue in SFDA and requires proper attention. Furthermore, distributing cluster knowledge among neighbor samples re- quires a memory bank [4, 42] which creates a significant burden in resource-constraint scenarios. In addition, most memory bank dependent SFDA techniques are not suitable for online test-time domain adaptation [73,76]; an emerging area of UDA that has gained traction in recent times. De- signing a memory-bank-free SFDA approach that can guide the self-training with highly precise pseudo-labels is a very challenging task and a major focus of this work. In our work, we focus on increasing the reliability of generated pseudo-labels without using a memory-bank and clustering-based pseudo-label refinement. Our analy- 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. 24120 àCorrect Pseudo-labelsàIncorrect/Noisy Pseudo-labels Target DomainSource ModelExisting SFDA: SST with AllPseudo-Labels Ours : SST with SelectedPseudo-LabelsIteration = 0Iteration = 1Iteration = T-1àModel Corrected PredictionCluster A Cluster B Cluster “A” AnchorCluster “B” AnchorClass “A” InstanceClass “B” InstanceExisting SFDA Techniques Target Domain Feature SpaceMisleading Neighbors àSelected Labels (DR)Figure 1. Left: In SFDA, we only have a source model that needs to be adapted to the target data. Among the source-generated pseudo-labels, a large portion is noisy which is important to avoid during supervised self-training (SST) with regular cross-entropy loss. Instead of using all pseudo-labels, we choose the most reliable ones and effectively propagate high-quality label information to unreliable samples. As the training progresses, the proposed selection strategy tends to choose more samples for SST due to the improved average reliability of pseudo-labels. Such a restricted self-training strategy creates a model with better discrimination ability and eventually corrects the noisy predictions. Here, Tis the total number of iterations. Right: While existing SFDA techniques leverages cluster structure knowledge in the feature space, there may exist many misleading neighbors —pseudo-labels of neighbors’ that are different from the anchors’ true label. Therefore, clustering-based label propagation inevitably suffers from label noise in subsequent training. sis shows that avoiding early training-time memorization (ETM) of noisy labels encourages noise-free learning in subsequent stages of adaptation. We further analyze that even with an expensive label refinement technique in place, learning equally from all labels eventually leads to label- noise memorization. Therefore, we employ a curriculum learning-aided self-training framework, C-SFDA, that pri- oritizes learning from easy-to-learn samples first and hard samples later on. We show that one can effectively iden- tify the group of easy samples by utilizing the reliabil- ity of pseudo-labels, i.e. prediction confidence and uncer- tainty. We then follow a carefully designed curriculum learning pipeline to learn from highly reliable (easy) sam- ples first and gradually propagate more refined label infor- mation among less reliable (hard) samples later on. In ad- dition to the self-training, we facilitate unsupervised con- trastive representation learning that helps us prevent the early training-time memorization phenomenon. Our main contributions are summarized as follows: • We introduce a novel SFDA technique that focuses on noise-free self-training exploiting the reliability of generated pseudo-labels. With the help of curriculum learning, we aim to prevent early training time memo- rization of noisy pseudo-labels and improve the quality of subsequent self-training as shown in Fig. 1. • By prioritizing the learning from highly reliable pseudo-labels first, we aim to propagate refined and accurate label information among less reliable sam- ples. Such a selective self-training strategy elimi- nates the requirement of a computationally costly and memory-bank dependent label refinement framework. • C-SFDA achieves state-of-the-art performance on ma- jor benchmarks for image recognition and semantic segmentation. Being highly memory-efficient, the pro- posed method is also applicable to online test-time adaptation settings and obtains SOTA performance.
Li_Learning_Generative_Structure_Prior_for_Blind_Text_Image_Super-Resolution_CVPR_2023	Abstract Blind text image super-resolution (SR) is challenging as one needs to cope with diverse font styles and unknown degradation. To address the problem, existing methods perform character recognition in parallel to regularize the SR task, either through a loss constraint or intermediate feature condition. Nonetheless, the high-level prior could still fail when encountering severe degradation. The prob- lem is further compounded given characters of complex structures, e.g., Chinese characters that combine multiple pictographic or ideographic symbols into a single charac- ter. In this work, we present a novel prior that focuses more on the character structure. In particular, we learn to encapsulate rich and diverse structures in a StyleGAN and exploit such generative structure priors for restora- tion. To restrict the generative space of StyleGAN so that it obeys the structure of characters yet remains flexible in handling different font styles, we store the discrete fea- tures for each character in a codebook. The code subse- quently drives the StyleGAN to generate high-resolution structural details to aid text SR. Compared to priors based on character recognition, the proposed structure prior ex- erts stronger character-specific guidance to restore faithful and precise strokes of a designated character. Extensive experiments on synthetic and real datasets demonstrate the compelling performance of the proposed generative structure prior in facilitating robust text SR. Our code is available at https://github.com/csxmli2016/MARCONet .	1. Introduction Blind text super-resolution (SR) aims at restoring a high- resolution (HR) image from a low-resolution (LR) text image that is corrupted by unknown degradation. The problem is relatively less explored than SR of natural and face images. This seemingly easy task actually possesses many unique challenges. In particular, each character owns its unique structure. A suboptimal restoration that destroys the struc- ture with, e.g., distorted, missing or additional strokes, is easily perceptible and may alter the meaning of the original word. The task becomes more difficult when dealing with writing systems such as Chinese, Kanji and Hanja characters, in which the patterns of thousands of characters vary in com- plex ways and are mostly composed of different subunits (or compound ideographs). A slight deviation in strokes, say ‘太’ and ‘大’ carries entirely different meanings. The prob- lem becomes more intricate given the different typefaces. In addition, complex and unknown degradation make the data-driven convolutional neural networks (CNNs) incapable of generalizing well to real-world degraded observations. To ameliorate the difficulties in blind text SR, espe- cially in retaining the unique structure of each character, recent studies design new loss functions to constrain SR re- sults semantically close to the high-resolution (HR) ground- truth [7,8,45,50,79], or incorporate the recognition prior into the intermediate features [40,41,44]. Figure 1 shows several representative results of these two types of methods, i.e., TB- SRN [7] that applies content-aware constraint and TATT [41] 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. 10103 that embeds the recognition prior into the SR process. The results are not satisfactory despite the models were retrained with data synthesized with sophisticated degradation models, i.e., BSRGAN [75] and Real-ESRGAN [66]. TBSRN and TATT perform well when the LR character has fewer strokes. However, they both fail to generate the desired structures when the character is composed of complex strokes, e.g., the 4, 6, 8, 11- thcharacters in (a). The results suggest that high-level recognition prior cannot well benefit the SR task in restoring faithful and accurate structures. When dealing with complex characters, we can identify some commonly used strokes as constituents of characters or their subparts. These obvious structural regularities have not been studied thoroughly in the literature. In particular, most existing text SR studies [7, 40, 41, 45, 50, 79] focus on scene text images that are presented in different view perspectives or arranged in irregular patterns. In contrast, our goal is to explore structural regularities specific to complex char- acters and investigate an effective way to exploit the prior. The investigation has practical values in many applications, including the restoration of old printed media ( e.g., newspa- pers or books) for digitalization and preservation, and the enhancement of captions in digital media. The key idea of our solution is to mine structure prior from a generative network, and use the prior to facilitate blind SR of text images. To capture the structure prior of characters, we train a StyleGAN to generate characters of different styles. As our intent is not to generate infinite and non-repetitive results, we restrict the generative space of StyleGAN to obey the structure of characters by constraining the generation with a codebook. Specifically, the codebook stores the discrete code of each character, and each code serves as a constant to StyleGAN for generating a specific high-resolution character. The font style is governed by the Wlatent space of StyleGAN. Once the character StyleGAN is trained, we can extract its intermediate features serving as the generative structure prior, which encapsulates the intrinsic structure and style of the LR character. Using the prior for text SR can be conveniently achieved through an additional Transformer-based encoder and a text SR network that accepts prior. In particular, given a text image that constitutes multiple characters ( e.g., Figure 1), we use a Transformer-based encoder to predict the font style, character bounding boxes, and their respective indexes in the codebook jointly. The codebook indexes drive the structure prior generation for each character. In the text SR network, each LR character is super-resolved, guided by the respective priors that are aligned using their bounding boxes. Experiments demonstrate that our design can generate robust and consistent structure priors for diverse and severely degraded text input. With the embedded structure prior, our approach performs superior against state-of-the-art methods in generating photo-realistic text results and shows excellentgeneralization to real-world LR text images. While our study mainly employs Chinese characters as examples, the proposed method can be readily extended to characters of other languages. We call our approach as MARCONet . The main contributions are summarized as follows: •We show that blind SR task, especially for characters with complex structures, can be restored by using their structure prior encapsulated in a generative network. •We present a viable way of learning such generative structure prior through reformulating a StyleGAN by replacing its single constant with discrete codes that represent different characters. •To retrieve the prior accurately, we propose a Transformer-based encoder to jointly predict the font styles, character bounding boxes and their indexes in codebook from the LR input.
Li_AMT_All-Pairs_Multi-Field_Transforms_for_Efficient_Frame_Interpolation_CVPR_2023	Abstract We present All-Pairs Multi-Field Transforms ( AMT ), a new network architecture for video frame interpolation. Itis based on two essential designs. First, we build bidirec- tional correlation volumes for all pairs of pixels, and use the predicted bilateral ﬂows to retrieve correlations for up-dating both ﬂows and the interpolated content feature. Sec- ond, we derive multiple groups of ﬁne-grained ﬂow ﬁeldsfrom one pair of updated coarse ﬂows for performing back- ward warping on the input frames separately. Combiningthese two designs enables us to generate promising task- oriented ﬂows and reduce the difﬁculties in modeling large motions and handling occluded areas during frame interpo-lation. These qualities promote our model to achieve state- of-the-art performance on various benchmarks with high efﬁciency. Moreover , our convolution-based model com-petes favorably compared to Transformer-based models interms of accuracy and efﬁciency. Our code is available at https://github.com/MCG-NKU/AMT .	1. Introduction Video frame interpolation (VFI) is a long-standing video processing technology, aiming to increase the temporal res- olution of the input video by synthesizing intermediate frames from the reference ones. It has been applied to various downstream tasks, including slow-motion genera-tion [ 22,61], novel view synthesis [ 11,29,71], video com- pression [ 60], text-to-video generation [ 52], etc. Recently, ﬂow-based VFI methods [ 17,22,26,34,53,69] have been predominant in referenced research due to theireffectiveness. A common ﬂow-based technique estimates bilateral/bidirectional ﬂows from the given frames and thenpropagates pixels/features to the target time step via back- ward [ 2,17,26] or forward [ 14,37,38] warping. Thus, the quality of a synthesized frame relies heavily on ﬂow esti- mation results. In fact, it is cumbersome to approximate intermediate ﬂows through pretrained optical ﬂow models, *Equal contribution †C.L. Guo is the corresponding author.[26] [26][26] [17] [14][43] Figure 1. Performance vs. number of parameters and FLOPs. The PSNR values are obtained from the Vimeo90K dataset [ 65]. We use a 720p frame pair to calculate FLOPs. Our AMT outperformsthe state-of-the-art methods and is with higher efﬁciency. and these ﬂows are unqualiﬁed for VFI usage [ 14,17]. A feasible way to alleviate this issue is to estimate task- oriented ﬂows in an end-to-end training manner [ 22,26, 32,65]. However, some major challenges, such as large motions and occlusions, are still pending to be resolved. These challenges mainly arise from the defective estimation of optical ﬂows. Thus, a straightforward question shouldbe: Why do previous methods have difﬁculties in predict- ing promising task-oriented ﬂows when facing these chal- lenges? Inspired by the recent studies [ 26,65] that demon- strate task-oriented ﬂow is generally consistent with ground truth optical ﬂow but diverse in local details , we attempt to answer the above question from two perspectives: (i)The ﬂow ﬁelds predicted by existing VFI methods are not consistent enough with the true displacements, es- pecially when encountering large motions (see Fig. 2). Ex- isting methods mostly adopt the UNet-like architecture [ 48] with plain convolutions to build VFI models. However, this type of architecture is vulnerable to accumulating errors at early stages when modeling large motions [ 45,56,63,70]. As a result, the predicted ﬂow ﬁelds are not accurate. (ii)Existing methods predict one pair of ﬂow ﬁelds, re- 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. 9801 Overlaid IFRNet [ 26] Flow IFRNet [ 26] Result Ground Truth Our Flow Our Result Figure 2. Qualitative comparisons of estimated ﬂows and the inter- polated frames. Our AMT guarantees the general consistency of intermediate ﬂows and synthesizes fast-moving objects with oc-cluded regions precisely, while the previous state-of-the-art IFR-Net [ 26] fails to achieve them. stricting the solution set in a tight space. This makes them struggle to handle occlusions and details around the mo-tion boundaries, which consequently deteriorates the ﬁnal results (see Fig. 2and Fig. 5). In this paper, we present a new network architecture, dubbed All-pairs Multi-ﬁeld Transforms ( AMT ), for video frame interpolation. AMT explores two new designs to im-prove the ﬁdelity and diversity of predicted ﬂows regarding the above two main shortcomings of previous works. Our ﬁrst design is based on all-pairs correlation in RAFT [ 56], which adequately models the dense correspon- dence between frames, especially for large motions. We propose to build bidirectional correlation volumes instead of unidirectional one and introduce a scaled lookup strat- egy to solve the coordinate mismatch issue caused by the invisible frame. Besides, the retrieved correlations assist our model in jointly updating bilateral ﬂows and the inter- polated content feature in a cross-scale manner. Thus, the network guarantees the ﬁdelity of ﬂows across scales, lay- ing the foundation for the following reﬁnement. Second, considering that predicting one pair of ﬂow ﬁelds is hard to cope with the occlusions, we propose toderive multiple groups of ﬁne-grained ﬂow ﬁelds from one pair of updated coarse bilateral ﬂows. The input frames canbe separately backward warped to the target time step bythese ﬂows. Such diverse ﬂow ﬁelds provide adequate po-tential solutions for each pixel to be interpolated, particu-larly alleviating the ambiguity issue in the occluded areas. We examine the proposed AMT on several public bench- marks with different model scales, showing strong perfor-mance and high efﬁciency in contrast to the state-of-the-art (SOTA) methods (see Fig. 1). Our small model outper- forms IFRNet-B, a SOTA lightweight model, by +0.17dBPSNR on Vimeo90K [ 65] with only 60% of its FLOPs and parameters. For large-scale setting, our AMT exceeds the previous SOTA ( i.e., IFRNet-L) by +0.15 dB PSNR on Vimeo90K [ 65] with 75% of its FLOPs and 65% of its pa-rameters. Besides, we provide a huge model for comparison with the SOTA transformer-based method VFIFormer [ 34]. Our convolution-based AMT shows a comparable perfor- mance but only needs nearly 23 ×less computational cost compared to VFIFormer [ 34]. Considering its effectiveness, we hope our AMT could bring a new perspective for the ar-chitecture design in efﬁcient frame interpolation.
Liu_ML2P-Encoder_On_Exploration_of_Channel-Class_Correlation_for_Multi-Label_Zero-Shot_Learning_CVPR_2023	Abstract Recent studies usually approach multi-label zero- shot learning (MLZSL) with visual-semantic mapping on spatial-class correlation, which can be computationally costly, and worse still, fails to capture fine-grained class- specific semantics. We observe that different channels may usually have different sensitivities on classes, which can correspond to specific semantics. Such an intrinsic channel- class correlation suggests a potential alternative for the more accurate and class-harmonious feature representa- tions. In this paper, our interest is to fully explore the power of channel-class correlation as the unique base for MLZSL. Specifically, we propose a light yet efficient Multi-Label Multi-Layer Perceptron-based Encoder , dubbed (ML)2P- Encoder, to extract and preserve channel-wise semantics. We reorganize the generated feature maps into several groups, of which each of them can be trained independently with (ML)2P-Encoder. On top of that, a global group- wise attention module is further designed to build the multi- label specific class relationships among different classes, which eventually fulfills a novel Channel- ClassCorrelation MLZSL framework (C3-MLZSL)1. Extensive experiments on large-scale MLZSL benchmarks including NUS-WIDE and Open-Images-V4 demonstrate the superiority of our model against other representative state-of-the-art models.	1. Introduction The proliferation of smart devices has greatly enriched human life when it comes to the era of big data. These smart devices are usually equipped with cameras such that users can easily produce and share their images. With the increas- ing abundance of public images, how to analyze them ac- curately has become a challenging problem. Recent years *Jingcai Guo is the corresponding author. 1Released code: github.com/simonzmliu/cvpr23_mlzsl INPUT WATERTREEGARDENFeature MapsChannel-Class CorrelationPredictSeen LabelsUnseen LabelsFigure 1. Example of Channel-Class Correlation. Our method achieves the prediction of unseen classes by exploiting the unique distribution of channel responses as semantic information for the class and building correlations with responses from the same chan- nel (zoom in for a better view). have witnessed great success in classifying an image into a specific class [20, 37, 39], namely, single-label classifica- tion. However, in reality, the images [17,46] usually contain abundant information and thereby consist of multiple labels. In recent years, the multi-label classification has been widely investigated by exploring the relationship among different labels from multiple aspects [9, 13, 14, 16, 42]. However, in some scenarios where extensive collections of images exist, e.g., Flickr2, users can freely set one or more individual tags/labels for each image, while the presented objects and labels in these images may not be fully shown in any previous collection, and thus result in a domain gap for the recognition. Therefore, in real-world applications, the model is required to gain the ability to predict unseen classes as well. As one of the thriving research topics, zero- 2https://www.flickr.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. 23859 shot learning (ZSL) [1, 12, 15, 34] is designed to transfer tasks from seen classes to unseen classes, and naturally recognizes novel objects of unseen classes. Specifically, ZSL has made continuous success in single-label classifica- tion [19, 26, 31, 45, 48]. However, these methods can hardly be extended to the multi-label scenario since exploring the cross-class relationships in an image is non-trivial. Recently, some works have focused on multi-label zero- shot learning (MLZSL) tasks and obtained some promising results [33, 36, 49]. Other works considered incorporating attention mechanisms into their models, such as LESA [22] andBiAM [35]. LESA [22] designed an attention-sharing mechanism for different patches in the image so that each patch can output the corresponding class. In another way, BiAM [35] designed a bi-level attention to extract relations from regional context and scene context, which can enrich the regional features of the model and separate the features of different classes. Although previous works have made considerable progress, their designed methods have been limited to the processing of spatial-domain information. First of all, the over-reliance on spatial-class correlation fails to capture fine-grained class-specific semantics. In addition, the ad- ditional processing of spatial information greatly increases the computational cost of the model and limits the infer- ence speed. Given the shortcomings of the above methods, we found through analysis that the channel response can be used as the semantic information of the class. Firstly, the response of each class in the channel is unique, which creates conditions for obtaining the unique semantics. Sec- ondly, for classes with certain semantic associations, there must be some channels that capture their common infor- mation. Therefore, channel information, as an easily over- looked part after feature extraction, can complete the task of capturing multi-label information. In MLZSL, we can complete the prediction of unseen classes by obtaining the responses of seen classes in the channel domain, and the relationship between seen and unseen classes. Finally, the subsequent analysis of the channel response greatly saves computational costs. Specifically, as shown in Figure 1, as seen classes, “wa- ter” and “tree” have unique response distributions on feature channels, and these responses can be used as semantic infor- mation for classification tasks. Besides, in order to explore the correlation of classes, we found that although the se- mantic information of “water” and “tree” is different, there are still some channels that respond simultaneously (i.e. the blue channel). We need to build this correlation during the training process through modeling so that the model can learn multi-label correlations. In the ZSL process, for the unseen class “garden”, we know that it is related to “water” (i.e. purple layer) and “tree” (i.e. green, orange, and gray layer) by obtaining its semantic information and matchingwith seen classes. This observation suggests that channels can help not only to classify objects but also to establish as- sociations between classes. Previous methods which only consider spatial information are unable to obtain this intrin- sic channel-class correlation and dissimilarity, thus achiev- ing sub-optimal performance on the MLZSL task. To address the above challenges and construct a more ac- curate and robust MLZSL system, we propose to group the generated feature maps and process them in a group-wise manner, thus enhancing the model by fully exploring the channel-class correlations. Besides, by properly designing a light yet efficient Multi-Label Multi-Layer Perceptron- based Encoder, i.e., (ML)2P-Encoder, we can easily analyze the local relationship between channels while significantly reducing the computation overhead. Finally, these groups are recombined and then perform the calculation of group attention, indicating that the model is analyzed locally and globally from the perspective of the channels, which can ensure the integrity of the representation. In summary, our contributions are four-fold: 1. To the best of our knowledge, our method first suggests the concept of channel-class correlation in MLZSL, and proposes a channel-sensitive attention module (ML)2P-Encoder to extract and preserve channel-wise semantics for channel groups. 2. Different from previous works that use spatial-class correlation to extract global and local features, we al- ternatively explore the channel-class correlation as the unique base for MLZSL. 3. In conjunction with (ML)2P-Encoder, a global group- wise attention is also designed to establish the multi- label specific class relationships among classes. 4. Extensive experiments on large-scale datasets NUS- WIDE and Open-Images -V4 demonstrate the effec- tiveness of our method against other state-of-the-art models.
Lin_Optimal_Transport_Minimization_Crowd_Localization_on_Density_Maps_for_Semi-Supervised_CVPR_2023	Abstract The accuracy of crowd counting in images has improved greatly in recent years due to the development of deep neu- ral networks for predicting crowd density maps. However, most methods do not further explore the ability to localize people in the density map, with those few works adopting simple methods, like finding the local peaks in the density map. In this paper, we propose the optimal transport mini- mization (OT-M) algorithm for crowd localization with den- sity maps. The objective of OT-M is to find a target point map that has the minimal Sinkhorn distance with the in- put density map, and we propose an iterative algorithm to compute the solution. We then apply OT-M to gener- ate hard pseudo-labels (point maps) for semi-supervised counting, rather than the soft pseudo-labels (density maps) used in previous methods. Our hard pseudo-labels pro- vide stronger supervision, and also enable the use of recent density-to-point loss functions for training. We also propose a confidence weighting strategy to give higher weight to the more reliable unlabeled data. Extensive experiments show that our methods achieve outstanding performance on both crowd localization and semi-supervised counting. Code is available at https://github.com/Elin24/OT-M .	1. Introduction Crowd understanding gains much attention due to its wide applications in surveillance [33, 61] and crowd dis- aster prevention. Most studies in this area concentrate on crowd counting, whose objective is to provide the to- tal number and distribution of crowds in a scene automat- ically. Due to the development of deep learning, recent methods [5,58,59,62,67] have achieved success on a variety of counting benchmarks [50, 61, 62, 66]. Counting methods can be extended to other applications, such as traffic man- agement [63], animal protection [2], and health care [32]. Although crowd counting has been greatly developed, most methods do not explore further applications of the estimated density maps after obtaining the count. Specifi- CNN OT-M (ours)Figure 1. The relationship between crowd counting CNNs and OT- M algorithm. CNNs are trained to generate density maps (soft la- bels). OT-M produces point maps (hard labels) from the predicted density maps, without needing training. cally, there is limited research on crowd localization, track- ing, or group analysis with predicted density maps. Taking crowd localization as an example, only a few methods, such as local maximum [13, 58, 65], integer programming [36], or Gaussian mixture models (GMM) [14], have been pro- posed to locate pedestrians or tiny objects from density maps. Moreover, recent localization methods ignore the counting density map, and instead are based on point de- tection [40, 54], blob segmentation [1, 10, 17], or inverse distance maps [24]. However, this will increase the inef- ficiency of a crowd understanding system, since separate networks are required for counting and localization. To broaden the application of density maps for localiza- tion, in this paper we propose a parameter-free algorithm, Optimal Transport Minimization (OT-M), to estimate the point map indicating locations of objects from a counting density map (see Fig. 1). OT-M minimizes the Sinkhorn distance [7] between the density map (source) and point lo- calization map (target), through an alternating scheme that estimates the optimal transport plan from the current point map (the OT-step), and updates the point map by minimiz- ing their transport cost (the M-step). OT-M is parameter- free and requires no training, and thus can be applied to any crowd-counting method using density maps. To demonstrate the applicability of density-map based localization, we apply OT-M to semi-supervised counting. In previous work, [38] builds a baseline for semi-supervised counting based on the mean-teacher framework [55], but finds it ineffective. Looking closely, we note that the base- line in [38] uses a softpseudo-label (density map) to super- vise the student model, whereas successful semi-supervised classification [52] or segmentation [6, 55] methods usually 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. 21663 are based on hard pseudo-labels (e.g., class labels or binary segmentation masks). In the context of semi-supervised crowd counting, the hard label is the point map in which each person is pseudo-annotated with a point. Thus, in this paper we generate hard pseudo-labels using OT-M for the unlabeled crowd images for semi-supervised crowd count- ing. As an additional benefit, the hard pseudo-labels al- low training CNNs under semi-supervised learning using recent density-to-point loss functions (e.g., Bayesian loss (BL) [34] or generalized loss (GL) [58]), which are more effective than traditional losses, e.g. L2. Similar to other semi-supervised tasks, some estimated pseudo labels may be inaccurate due to limitations of the current trained model. To reduce the effect of these noisy pseudo-labels, we propose a confidence-weighted GL (C- GL) for semi-supervised counting. Specifically, we com- pute the unbalanced optimal transport plan between the stu- dent’s predicted density map (source) and the hard pseudo- labels from the teacher (target), and then define the pixel- wise and point-wise confidences based on the consistencies between source, target and the plan. Experiments show that the trained model is more robust with our C-GL. In summary, the contributions of this paper are 3-fold: •We propose an OT-M algorithm to estimate the loca- tions of objects from density maps, which is based on minimizing the Sinkhorn distance between the den- sity map and the target point map. Since OT-M is parameter-free, it can be applied to any density map without training. •We use OT-M to produce hard pseudo-labels for semi- supervised counting, which conforms with schemes in other semi-supervised tasks. The hard label also al- lows applying density-to-point loss like GL to unla- beled data for more effective training. •To mitigate risks brought by inaccurate pseudo-labels, we propose a confidence-weighted Generalized Loss to reduce the influence of inconsistencies between the teacher’s and student’s predictions. Experiments show that our loss improves semi-supervised counting per- formance.
Kim_NeuralField-LDM_Scene_Generation_With_Hierarchical_Latent_Diffusion_Models_CVPR_2023	Abstract Automatically generating high-quality real world 3D scenes is of enormous interest for applications such as vir- tual reality and robotics simulation. Towards this goal, we introduce NeuralField-LDM, a generative model capable of synthesizing complex 3D environments. We leverage Latent Diffusion Models that have been successfully utilized for ef- ficient high-quality 2D content creation. We first train a scene auto-encoder to express a set of image and pose pairs as a neural field, represented as density and feature voxel grids that can be projected to produce novel views of the scene. To further compress this representation, we train a latent-autoencoder that maps the voxel grids to a set of la- tent representations. A hierarchical diffusion model is then fit to the latents to complete the scene generation pipeline. We achieve a substantial improvement over existing state- of-the-art scene generation models. Additionally, we show how NeuralField-LDM can be used for a variety of 3D con- tent creation applications, including conditional scene gen- eration, scene inpainting and scene style manipulation.	1. Introduction There has been increasing interest in modelling 3D real- world scenes for use in virtual reality, game design, digi- *Equal contribution. †Work done during an internship at NVIDIA.tal twin creation and more. However, designing 3D worlds by hand is a challenging and time-consuming process, re- quiring 3D modeling expertise and artistic talent. Recently, we have seen success in automating 3D content creation via 3D generative models that output individual object as- sets [15, 46, 73]. Although a great step forward, automating the generation of real-world scenes remains an important open problem and would unlock many applications ranging from scalably generating a diverse array of environments for training AI agents ( e.g. autonomous vehicles) to the de- sign of realistic open-world video games. In this work, we take a step towards this goal with NeuralField-LDM (NF- LDM), a generative model capable of synthesizing complex real-world 3D scenes. NF-LDM is trained on a collection of posed camera images and depth measurements which are easier to obtain than explicit ground-truth 3D data, offering a scalable way to synthesize 3D scenes. Recent approaches [2, 6, 8] tackle the same problem of generating 3D scenes, albeit on less complex data. In [6,8], a latent distribution is mapped to a set of scenes using ad- versarial training, and in GAUDI [2], a denoising diffusion model is fit to a set of scene latents learned using an auto- decoder. These models all have an inherent weakness of attempting to capture the entire scene into a single vector that conditions a neural radiance field. In practice, we find that this limits the ability to fit complex scene distributions. Recently, diffusion models have emerged as a very pow- erful class of generative models, capable of generating high- 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. 8496 quality images, point clouds and videos [18, 24, 39, 48, 53, 73,78]. Yet, due to the nature of our task, where image data must be mapped to a shared 3D scene without an explicit ground truth 3D representation, straightforward approaches fitting a diffusion model directly to data are infeasible. In NeuralField-LDM, we learn to model scenes using a three-stage pipeline. First, we learn an auto-encoder that en- codes scenes into a neural field, represented as density and feature voxel grids. Inspired by the success of latent diffu- sion models for images [53], we learn to model the distribu- tion of our scene voxels in latent space to focus the genera- tive capacity on core parts of the scene and not the extrane- ous details captured by our voxel auto-encoders. Specif- ically, a latent-autoencoder decomposes the scene voxels into a 3D coarse, 2D fine and 1D global latent. Hierar- chichal diffusion models are then trained on the tri-latent representation to generate novel 3D scenes. We show how NF-LDM enables applications such as scene editing, birds- eye view conditional generation and style adaptation. Fi- nally, we demonstrate how score distillation [46] can be used to optimize the quality of generated neural fields, al- lowing us to leverage the representations learned from state- of-the-art image diffusion models that have been exposed to orders of magnitude more data. Our contributions are: 1) We introduce NF-LDM, a hi- erarchical diffusion model capable of generating complex open-world 3D scenes and achieving state of the art scene generation results on four challenging datasets. 2) We ex- tend NF-LDM to semantic birds-eye view conditional scene generation, style modification and 3D scene editing.
Li_BioNet_A_Biologically-Inspired_Network_for_Face_Recognition_CVPR_2023	Abstract Recently, whether and how cutting-edge Neuroscience findings can inspire Artificial Intelligence (AI) confuse both communities and draw much discussion. As one of the most critical fields in AI, Computer Vision (CV) also pays much attention to the discussion. To show our ideas and exper- imental evidence to the discussion, we focus on one of the most broadly researched topics both in Neuroscience and CV fields, i.e., Face Recognition (FR). Neuroscience studies show that face attributes are essential to the human face- recognizing system. How the attributes contribute also be explained by the Neuroscience community. Even though a few CV works improved the FR performance with attribute enhancement, none of them are inspired by the human face- recognizing mechanism nor boosted performance signifi- cantly. To show our idea experimentally, we model the biological characteristics of the human face-recognizing system with classical Convolutional Neural Network Op- erators (CNN Ops) purposely. We name the proposed Biologically-inspired Network as BioNet . Our BioNet con- sists of two cascade sub-networks, i.e., the Visual Cor- tex Network (VCN) and the Inferotemporal Cortex Net- work (ICN). The VCN is modeled with a classical CNN backbone. The proposed ICN comprises three biologically- inspired modules, i.e., the Cortex Functional Compartmen- talization, the Compartment Response Transform, and the Response Intensity Modulation. The experiments prove that: 1) The cutting-edge findings about the human face- recognizing system can further boost the CNN-based FR network. 2) With the biological mechanism, both identity- related attributes ( e.g., gender) and identity-unrelated at- tributes ( e.g., expression) can benefit the deep FR models. Surprisingly, the identity-unrelated ones contribute even more than the identity-related ones. 3) The proposed BioNet significantly boosts state-of-the-art on standard FR bench- mark datasets. For example, BioNet boosts IJB-B@1e-6 from 52.12% to 68.28% and MegaFace from 98.74% to 99.19%. The source code is released in1. 1https://github.com/pengyuLPY/BioNet.git Figure 1. (A) Face recognition mechanism of human brains. (B) Architecture of BioNet.	1. Introduction It sparked much discussion in both communities that Zador, Bengio et al . [54] claimed fundamental Neuro- science research must be invested to accelerate AI progress. Some researchers agree with Zador’s opinion. For example, LeCun et al. [52] and Goodfellow et al. [10] proposed the Convolutional Neural Network (CNN) inspired by past clas- sical Neuroscience discoveries about the human visual cor- tex. However, other researchers have some concerns, e.g., 1) Except for the high-level and abstract senses from Neuro- science, can their specific studies support the AI fields? 2) Since CNN [9,10,26,52] was proposed many years ago, few AI works have been inspired by recent Neuroscience find- ings. It results in a lack of evidence to support that the latest Neuroscience studies can continue to drive AI progress. To show an idea and some experimental evidence to this discussion, we focus on one of the most broadly re- searched topics in both fields, i.e., Face Recognition (FR). The latest Neuroscience studies [1, 2, 7, 38, 51] found that 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. 10344 besides the visual cortex, the inferotemporal cortex plays a vital role in the human face-recognizing system. Because of the following three biological characteristics, the infer- otemporal cortex characterizes the complicated relationship between attributes and makes attributes contribute to FR. 1) The inferotemporal cortex is functionally compartmen- talized by the face stimuli ( i.e.,identity ,attributes ) [1, 7]. 2) The responses of the functional compartments are trans- formed into the successor neurons for processing com- plex tasks [38], e.g., FR. 3) The intensities of the func- tional compartment are variant in the inferotemporal cor- tex [2], demonstrating that the attributes are not equally essential to the human face-recognizing system. In the AI field, some CNN-based FR works proposed effective loss functions [4, 18, 21, 33, 40, 45] and others designed task-oriented face recognition structures [22,23,27–29,42]. Because of these excellent works, FR in AI achieved big progress. Although a few AI FR works [14, 19, 43] boosted the performance with attribute enhancement, none of them are inspired by the human face-recognizing mechanism nor boosted performance significantly. We experimentally prove that the deep FR models do not capture the biological characteristics of the human face-recognizing system intro- duced above. We suspect this is the factor limiting their performance improvement. To address the problem and experimentally support the opinion of Zador, Bengio et al. [54], we purposely model the biological characteristics of the human face-recognizing system with classical CNN Ops. We name the proposed Biologically-inspired Network as BioNet . The proposed BioNet is constituted by two cascade sub-networks, i.e., Visual Cortex Network (VCN) andInferotemporal Cor- tex Network (ICN) , as Fig.1 shows. The VCN is modeled with a CNN backbone ( e.g., CASIA-Net [53], ResNet [12]), which follows the common suggestions in [10, 52]. The proposed ICN is composed of three biologically-inspired modules, i.e. the Cortex Functional Compartmentalization (CFC), the Compartment Response Transform (CRT), and the Response Intensity Modulation (RIM). CFC is based on the attention mechanism [8, 15, 44] to functionally com- partmentalize the feature maps with face stimuli ( i.e.iden- tityand attributes ). CRT is implemented by Multilayer Perception (MLP) [36] to transform intra-compartment re- sponses to successor neurons for processing complex FR task. RIM follows the ensemble mechanism [55] to fuse inter-compartment features via adaptive weights to achieve the final FR representation. The proposed modules follow the human face-recognizing mechanism and empower ICN to characterize the complicated relationship between stim- uli. All of them are indispensable in BioNet. With such a Biologically-inspired Network, we achieve better attribute-enhanced deep FR models than ever. Fur- thermore, we conduct careful analyses of the proposedmodules and the impacts of attributes. We also compare the BioNet to the attention, multi-task learning, and ensem- ble mechanisms, verifying the advantage of our proposals. We think the experiments in this paper can support the con- clusions: 1. To the best of our knowledge, after CNN was pro- posed, our BioNet is the first network inspired by the latest Neuroscience studies. It provides experimental evidence that the latest Neuroscience studies can fur- ther boost the CNN-based FR network and continue to drive AI progress. 2. With the Neuroscience mechanism, both identity- related attributes ( e.g.,gender ) and identity-unrelated attributes ( e.g.,expression ) can benefit the deep FR models. Surprisingly, we find the identity-unrelated ones contribute even more than the identity-related ones. Besides, we also propose an online latent at- tributes mining method and prove that the latent at- tributes contribute to the FR task, too. 3. Without bells and whistles, Our BioNet consistently and significantly boosts state-of-the-art of FR models on standard FR benchmark datasets, e.g., IJB-A [24], IJB-B [46], IJB-C [46], and MegaFace [20].
Kim_Diffusion_Video_Autoencoders_Toward_Temporally_Consistent_Face_Video_Editing_via_CVPR_2023	Abstract Inspired by the impressive performance of recent face image editing methods, several studies have been naturally proposed to extend these methods to the face video editing task. One of the main challenges here is temporal consis- tency among edited frames, which is still unresolved. To this end, we propose a novel face video editing framework based on diffusion autoencoders that can successfully extract the decomposed features - for the ﬁrst time as a face video edit- ing model - of identity and motion from a given video. This modeling allows us to edit the video by simply manipulat- ing the temporally invariant feature to the desired direction for the consistency. Another unique strength of our model is that, since our model is based on diffusion models, it can satisfy both reconstruction and edit capabilities at the same time, and is robust to corner cases in wild face videos ( e.g. occluded faces) unlike the existing GAN-based methods.1	1. Introduction As one of the standard tasks in computer vision to change various face attributes such as hair color, gender, or glasses 1Project page: https://diff-video-ae.github.ioof a given face image, face editing has been continuously gaining attention due to its various applications and en- tertainment. In particular, with the improvement of analy- sis and manipulation techniques for recent Generative Ad- versarial Network (GAN) models [ 8,11,22,29,30], we simply can do this task by manipulating a given image’s latent feature. In addition, very recently, many methods for face image editing also have been proposed based on Diffusion Probabilistic Model (DPM)-based methods that show high-quality and ﬂexible manipulation performance [3,12,17,19,23,25]. Naturally, further studies [ 2,35,41] have been proposed to extend image editing methods to incorporate the tempo- ral axis for videos. Now, given real videos with a human face, these studies try to manipulate some target facial at- tributes with the other remaining features and motion in- tact. They all basically edit each frame of a video inde- pendently via off-the-shelf StyleGAN-based image editing techniques [ 22,29,41]. Despite the advantages of StyleGAN in this task such as high-resolution image generation capability and highly dis- entangled semantic representation space, one harmful draw- back of GAN-based editing methods is that the encoded 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. 6091 images cannot perfectly be recovered by the pretrained gen- erator [ 1,26,34]. Especially, if a face in a given image is unusually decorated or occluded by some objects, the ﬁxed generator cannot synthesize it. For perfect reconstruction, several methods [ 6,27] are suggested to further tune the generator for GAN-inversion [ 26,34] on one or a few tar- get images, which is computationally expensive. Moreover, after the ﬁne-tuning, the original editability of GANs can- not be guaranteed. This risk could be worse in video domain since we have to ﬁnetune the model on the multiple frames. Aside from the reconstruction issue of existing GAN- based methods, it is critical in video editing tasks to con- sider the temporal consistency among edited frames to pro- duce realistic results. To address this, some prior works rely on the smoothness of the latent trajectory of the original frames [ 35] or smoothen the latent features directly [ 2] by simply taking the same editing step for all frames. However, smoothness does not ensure temporal consistency. Rather, the same editing step can make different results for differ- ent frames because it can be unintentionally entangled with irrelevant motion features. For example, in the middle row of Fig. 1, eyeglasses vary across time and sometimes dimin- ish when the man closes his eyes. In this paper, we propose a novel video editing frame- work for human face video, termed Diffusion Video Au- toencoder , that resolves the limitations of the prior works. First, instead of GAN-based editing methods suffering from imperfect reconstruction quality, we newly introduce diffu- sion based model for face video editing tasks. As the re- cently proposed diffusion autoencoder (DiffAE) [ 23] does, thanks to the expressive latent spaces of the same size as input space, our model learns a semantically meaningful la- tent space that can perfectly recover the original image back and are directly editable. Not only that, for the ﬁrst time as a video editing model, encode the decomposed features of the video: 1) identity feature shared by all frames, 2) fea- ture of motion or facial expression in each frame, and 3) background feature that could not have high-level represen- tation due to large variances. Then, for consistent editing, we simply manipulate a single invariant feature for the de- sired attribute (single editing operation per video), which is also computationally beneﬁcial compared to the prior works that require editing the latent features of all frames. We experimentally demonstrate that our model appropri- ately decomposes videos into time-invariant and per-frame variant features and can provide temporally consistent ma- nipulation. Speciﬁcally, we explore two ways of manipula- tion. The ﬁrst one is to edit features in the predeﬁned set of attributes by moving semantic features to the target di- rection found by learning linear classiﬁer in semantic repre- sentation space on annotated CelebA-HQ dataset [ 10]. Ad- ditionally, we explore the text-based editing method that op- timizes a time-invariant latent feature with CLIP loss [ 7]. Itis worth noting that since we cannot fully generate edited images for CLIP loss due to the computational cost, we pro- pose the novel strategy that rather uses latent state of inter- mediate time step for the efﬁciency. To summarize, our contribution is four-fold: •We devise diffusion video autoencoders based on dif- fusion autoencoders [ 23] that decompose the video into a single time-invariant and per-frame time-variant features for temporally consistent editing. •Based on the decomposed representation of diffusion video autoencoder, face video editing can be con- ducted by editing only the single time-invariant iden- tity feature and decoding it together with the remaining original features. •Owing to the nearly-perfect reconstruction ability of diffusion models, our framework can be utilized to edit exceptional cases such that a face is partially occluded by some objects as well as usual cases. •In addition to the existing predeﬁned attributes edit- ing method, we propose a text-based identity editing method based on the local directional CLIP loss [ 7,24] for the intermediately generated product of diffusion video autoencoders.
Kulkarni_Learning_To_Predict_Scene-Level_Implicit_3D_From_Posed_RGBD_Data_CVPR_2023	Abstract We introduce a method that can learn to predict scene- level implicit functions for 3D reconstruction from posed RGBD data. At test time, our system maps a previously unseen RGB image to a 3D reconstruction of a scene via implicit functions. While implicit functions for 3D recon- struction have often been tied to meshes, we show that we can train one using only a set of posed RGBD images. This setting may help 3D reconstruction unlock the sea of ac- celerometer+RGBD data that is coming with new phones. Our system, D2-DRDF , can match and sometimes outper- form current methods that use mesh supervision and shows better robustness to sparse data.	1. Introduction Consider the image in Figure 1. From this ordinary RGB image, we can understand the complete 3D geometry of the scene, including the floor and walls behind the chairs. Our goal is to enable computers to recover this geometry from a single RGB image. To this end, we present a method that does so while learning only on posed RGBD images. The task of reconstructing the full 3D of a scene from a single previously unseen RGB image has long been known to be challenging. Early work on full 3D relied on vox- els [6, 17] or meshes [19], but these representations fail on scenes due to topology and memory challenges. Implicit functions (or learning to map each point in R3to a value like the distance to the nearest surface) have shown substantial promise at overcoming these challenges. When conditioned on an image, these have led to several successful methods. Unfortunately, the implicit function status quo mainly ties implicit function reconstruction methods to mesh su- pervision. This symbiosis has emerged since meshes give excellent direct supervision. However, methods are limited to training with an image-aligned mesh that is usually wa- tertight (and often artist-created) [35, 38, 44] and occasion- ally non-watertight but professionally-scanned [27, 61]. We present a method, Depth to DRDF (D2-DRDF), that breaks the implicit-function/mesh connection and can train an effective single RGB image implicit 3D reconstruction (Ours) Posed Depth (Prior work) Explicit 3D MeshInference: Single RGB Image to 3D Training Data Input Figure 1. We propose a method, D2-DRDF, that reconstructs full 3D from a single RGB image. During inference (left), our method uses implicit functions to reconstruct the complete 3D in- cluding visible and occluded surfaces (visualized as surface nor- mals ) such as the occluded wall and empty floor. For training, our method uses only RGBD images and poses, unlike most prior works that need an explicit and often watertight 3D mesh. system using a set of RGBD images with known pose. We envision that being able to entirely skip meshing will enable the use of vast quantities of lower-quality data from con- sumers (e.g., from increasingly common consumer phones with LiDAR scanners and accelerometers) as well as robots. In addition to permitting training, the bypassing of meshing may enable adaptation in a new environment on raw data without needing an expert to ensure mesh acquisition. Our key insight is that we can use segments of observed free space in depth maps in other views to constrain dis- tance functions. We show this using the Directed Ray Dis- tance Function (DRDF) [27] that has recently shown good performance in 3D reconstruction using implicit functions and has the benefit of not needing watertight meshes for training. Given an input reference view, the DRDF breaks the problem of predicting the 3D surface into a set of in- dependent 1D distance functions, each along a ray through a pixel in the reference view and accounting for only sur- faces on the ray. Rather than use an ordinary unsigned dis- tance function, the DRDF signs the distance using the lo- cation of the camera’s pinhole and the intersections along the ray. While [27] showed their method could be trained onnon-watertight meshes, their method is still dependent 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. 17256 on meshes. In our paper, we show that the DRDF can be cleanly supervised using auxiliary views of RGBD obser- vations and their poses. We derive constraints on the DRDF that power loss functions for training our system. While the losses on any one image are insufficient to produce a recon- struction, they provide a powerful signal when accumulated across thousands of training views. We evaluate our method on realistic indoor scene datasets and compare it with methods that train with full mesh supervision. Our method is competitive and some- times even better compared to the best-performing mesh- supervised method [28] with full, professional captures. As we degrade scan completeness, our method largely main- tains its performance while mesh-based methods perform substantially worse. We conclude by showing that fine- tuning of our method on a handful of images enables a sim- ple, effective method for fusing and completing scenes from a handful of RGBD images.
Lang_SCOOP_Self-Supervised_Correspondence_and_Optimization-Based_Scene_Flow_CVPR_2023	Abstract Scene flow estimation is a long-standing problem in com- puter vision, where the goal is to find the 3D motion of a scene from its consecutive observations. Recently, there have been efforts to compute the scene flow from 3D point clouds. A common approach is to train a regression model that consumes source and target point clouds and outputs the per-point translation vector. An alternative is to learn point matches between the point clouds concurrently with regressing a refinement of the initial correspondence flow. In both cases, the learning task is very challenging since the flow regression is done in the free 3D space, and a typical solution is to resort to a large annotated synthetic dataset. We introduce SCOOP , a new method for scene flow esti- mation that can be learned on a small amount of data with- out employing ground-truth flow supervision. In contrast to previous work, we train a pure correspondence model fo- cused on learning point feature representation and initial- ize the flow as the difference between a source point and its softly corresponding target point. Then, in the run-time phase, we directly optimize a flow refinement component with a self-supervised objective, which leads to a coherent and accurate flow field between the point clouds. Experi- ments on widespread datasets demonstrate the performance gains achieved by our method compared to existing leading techniques while using a fraction of the training data. Our code is publicly available1.	1. Introduction Scene flow estimation [27] is a fundamental problem in computer vision with various use-cases, such as au- tonomous driving, scene parsing, pose estimation, and ob- ject tracking, to name a few. Given two consecutive obser- vations of a 3D scene, the aim is to compute the dynamics of the scene between the observations. Scene flow predic- tion based on 2D images has been thoroughly investigated in the literature [17, 19, 28, 32, 33]. However, in light of the 1https://github.com/itailang/SCOOP *The work was done during an internship at Google Research. 100 1,000 10,000 Train set size (#, in a log10 scale)0102030405060708090100Strict accuracy (%) FlowNet3DFLOT FESTA3DFlow BiPFNSCOOP (ours)SCOOP+(ours) RigidFlowJust Go with the Flow Self-Point-FlowSCOOP (ours)SCOOP+(ours) Full supervision, synthetic train data (FlyingThings3D) Self supervision, synthetic train data (FlyingThings3D) Self supervision, real train data (KITTI)Figure 1. Flow accuracy on the KITTI benchmark vs. the train set size. Our method is trained on one or two orders of magnitude less data while surpassing the performance of the competing tech- niques [4, 13, 14, 16, 21, 24, 30, 31] by a large margin. Please see Table 1 for the complete details of the evaluation settings. recent proliferation of 3D sensors, such as LiDAR, there is a surge of interest in scene flow methods that operate directly on the 3D data [10, 13, 16, 21, 34]. Liu et al . [16] were among the first to pursue this research avenue. They proposed FlowNet3D, a fully- supervised neural network that learned to regress the flow between 3D point clouds and showed remarkable perfor- mance improvement over image-based techniques [1, 19, 29]. Since their method required ground-truth flow anno- tations, which are scarce for real-world data, they turned to training on a large synthetic dataset that compromised the generalization capability to real-world LiDAR data. Follow-up works devised self-supervised learning schemes [13, 21] and narrowed the domain gap by training on unannotated LiDAR point cloud pairs. However, similar to Liu et al. [16], they used a regression approach in which the model should learn to compute the flow in the free 3D space . This task is extremely challenging, given the irreg- ular nature of point clouds, and requires a large amount of training data for the network to converge. In another line of work [8, 11, 24], researchers leveraged point cloud correspondence for scene flow prediction. In this approach, the flow is computed as the translation of a point in the first point cloud (source) to its softly corre- 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. 5281 Correspondence Learning Regr ession Learning Refinement Learning Point clouds Flow Point clouds Flow Point clouds Flow Point clouds Flow Flow Point clouds Train Test Correspondence Model Refinement Model Regression Model Correspondence Learning Refinement Optimization Initial Flow Flow Point clouds Point clouds Optimize Freeze Correspondence Model Flow Optimization No Training aaaaaaaFlowNet3D [16] aaaaaaaaaaFLOT [24] aaaaaaaaaaaNeural Prior [15] aSCOOP (ours) Figure 2. Comparison of scene flow approaches. Given a pair of point clouds, FlowNet3D [16] learns to regress the flow in the free 3D space, and the trained model is frozen for testing. FLOT [24] concurrently trains two network components: one that computes point correspondence and another that regresses a correction to the resulting correspondence flow. Neural Prior [15] optimizes the flow between the point clouds from scratch without learning. In contrast to previous work, we take a hybrid approach. We train a pure correspondence model without flow regression , which serves for flow initialization. Then, we directly optimize only the flow refinement at the test-time. sponding point in the second one (target). The softly cor- responding point is a weighted sum of target points based on point similarity in a learned latent space. Thus, rather than the challenging regression problem in the 3D ambi- ent space, the flow estimation task boils down to point feature learning and is reduced to the convex combination space [26] of existing target points. However, to relax this constraint, another network component is trained to regress flow corrections. The joint training of point representation and flow refinement burdens the learning process and re- tains the reliance on large datasets with flow supervision. Another emerging approach is an optimization-only flow computation [15, 23]. In this case, no training data is in- volved, and the flow is optimized at run-time for each scene separately. Despite the high accuracy such a dedicated op- timization achieves, it requires a long processing time. We present SCOOP, a hybrid flow estimation method that can be learned from a small amount of training data. SCOOP consists of two parts: a self-supervised neural net- work for point cloud correspondence and a direct flow re- finement optimization module. During the training phase, the network learns to extract point features for soft point matches, which initialize the flow between the point clouds. In contrast to previous work, our network is focused on learning just the point embeddings, allowing its training on a very small dataset, as shown in Figure 1. Additionally, we consider the confidence of the network in the computed correspondences to guide the learning process better. Then, instead of training another network for regress- ing flow updates, we define an optimization problem and directly optimize residual flow refinement vectors at run- time. The optimization objective encourages a coherent flow field while retaining the translated source points close to the target point cloud. Our design choices improve the accuracy compared to learning-based methods and reduce the processing time with respect to the optimization-only approach [15, 23]. For both correspondence learning and refinement optimization, we use a self-supervised distance objective and a smoothness prior instead of ground-truthflow labels. Figure 2 presents the difference between our approach and leading previous ones. In summary, we propose a hybrid flow prediction ap- proach for point clouds based on self-supervised correspon- dence learning and direct run-time residual flow optimiza- tion. Using well-established datasets in the scene flow liter- ature, we show that our approach yields clear performance improvement over existing state-of-the-art methods while using a fraction of the training data and without employing any ground-truth flow supervision.
Lin_Towards_Fast_Adaptation_of_Pretrained_Contrastive_Models_for_Multi-Channel_Video-Language_CVPR_2023	Abstract Multi-channel video-language retrieval require models to understand information from different channels (e.g. video +question, video +speech) to correctly link a video with a textual response or query. Fortunately, contrastive multimodal models are shown to be highly effective at align- ing entities in images/videos and text, e.g., CLIP [20]; text contrastive models are extensively studied recently for their strong ability of producing discriminative sentence embed- dings, e.g., SimCSE [5]. However, there is not a clear way to quickly adapt these two lines to multi-channel video- language retrieval with limited data and resources. In this paper, we identify a principled model design space with two axes: how to represent videos and how to fuse video and text information. Based on categorization of recent methods, we investigate the options of representing videos using contin- uous feature vectors or discrete text tokens; for the fusion method, we explore the use of a multimodal transformer or a pretrained contrastive text model. We extensively evalu- ate the four combinations on five video-language datasets. We surprisingly find that discrete text tokens coupled with a pretrained contrastive text model yields the best perfor- mance, which can even outperform state-of-the-art on the iVQA and How2QA datasets without additional training on millions of video-text data. Further analysis shows that this is because representing videos as text tokens captures the key visual information and text tokens are naturally aligned with text models that are strong retrievers after the con- trastive pretraining process. All the empirical analysis es- tablishes a solid foundation for future research on afford- able and upgradable multimodal intelligence.	1. Introduction From retrieving a trending video on TikTok with natural language descriptions to asking a bot to solve your tech- nical problem with the question and a descriptive video,AI agents handling multi-channel video-language retrieval- style tasks have been increasingly demanded in this post- social-media era. These tasks require the agent to fuse in- formation from multiple channels, i.e., video and text to retrieve a text response or return a multi-channel sample for a text query. To power such agents, a popular ap- proach [9, 10, 15, 34, 40] consists of two rounds of pre- training: (1) The 1st round is to obtain unimodal pre- trained models, such as visual-only encoders [3, 6, 17, 20] (e.g., S3D,CLIP) and text-only encoders [4,13,22,24] (e.g., BERT) (2) The 2nd round aims at pretraining on visual- text dataset - specifically, researchers leverage techniques like masked token modeling [9, 40] or contrastive learn- ing [10,15,33,34] to align and fuse unimodal features from model pretrained in the 1st round. Such methods achieve good performance on multi- channel retrieval-style tasks but they suffer from two ma- jor limitations: 1) huge amounts of data and computational resources are required for the second-round “pretraining”, which significantly limits the research exploration without such resources; 2) the domain of video data used in the second round “pretraining” has to be strongly correlated with downstream tasks [9], which may restrict such meth- ods from being generally applicable. To alleviate such limitations, we study a novel problem: fast adaptation of pretrained contrastive models on multi- channel video-language retrieval under limited resources. Specifically, we propose to adapt both contrastive multi- modal models [16,20] and contrastive text models [5,22] to enjoy their strong encoding ability and discriminative em- bedding space. There has been tremendous progress re- cently on large-scale contrastive multimodal models [16, 17, 20]. Through pretraining on millions of images/videos, these models are highly effective at encoding visual inputs and linking entities across modalities . Meanwhile, con- trastive text models [5, 22] have been also densely stud- ied to obtain discriminative sentence embeddings . These models are shown to perform well on challenging text re- trieval tasks such as semantic search [19], which requires 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. 14846 the model to understand both language and real-world knowledge [5, 22]. Such an ability allows the model to retrieve a text response or encode a text query in multi- channel video-language retrieval-style tasks. Thereby, once the video information is effectively incorporated, it elimi- nates the necessity of second-round “pretraining” on large scale multimodal datasets, enabling fast adaptation to any video-language retrieval-style tasks. We first conduct a systematic analysis on potential model designs, as shown in Figure 1. We identify the model de- sign space with two design principles: how to represent the video information and how to fuse this video information with questions or other text such as speech. To represent video information, we could either adopt Continuous Fea- tures which is commonly used in existing work [18, 34], or project videos into unified Text Tokens [11] from vari- ous modalities. To fuse information from multiple channels, i.e., video and question/speech, there are two potential op- tions, namely, a Multimodal Transformer [34] or a Text Transformer [18]. Hence, there are four combinations de- rived from this model design space, namely, Continuous Features + Multimodal Transformer [34], Continuous Features + Text Transformer ,Text Tokens + Multimodal Transformer ,Text Tokens + Text Transformer . Our exploration of this model design space results into a simple yet effective approach that allows fast adaptation of pretrained contrastive models, which first leverages con- trastive multimodal models to retrieve a sequence of Text Tokens for the visual input and then feeds these tokens to- gether with other text to contrastive Text Transformer for answer retrieval. Its fast adaptation ability not only comes from the ability of linking entities across modalities of the contrastive multimodal model, but it also enjoys the natural alignment with contrastive text models to produce discrimi- native embeddings. To the best of our knowledge, this is the first proposal to adapt pretrained contrastive models in this manner, although each individual design choice may have been adopted previously. We further conduct in-depth anal- ysis to understand 1) the trade-off between data efficiency and accuracy, 2) the impact of pretrained contrastive text model, and 3) the possible limitation of this framework. The contribution could be summarized three-fold: • We identified the principled model design space for fast adaption of pretrained contrastive multimodal models and pretrained contrastive text models. • We conducted extensive experiments on five video- language datasets, observed a consistent trend across these four variants, and even obtained state-of-the-art performance (e.g., 6.5% improvement on How2QA) with the proposed Text Tokens + Text Transformer variant without using millions of extra multimodal data samples , which is essential to democratize thecommunity of video+language. • The proposed Text Tokens + Text Transformer vari- ant scales significantly better than the other variants, w.r.t. the quality of pretrained text models. The code will be released at https://github.com/ XudongLinthu/upgradable- multimodal- intelligence to facilitate future research.
Liu_MixMAE_Mixed_and_Masked_Autoencoder_for_Efficient_Pretraining_of_Hierarchical_CVPR_2023	Abstract In this paper, we propose Mixed and Masked AutoEn- coder (MixMAE), a simple but efficient pretraining method that is applicable to various hierarchical Vision Transform- ers. Existing masked image modeling (MIM) methods for hierarchical Vision Transformers replace a random subset of input tokens with a special [MASK] symbol and aim at reconstructing original image tokens from the corrupted im- age. However, we find that using the [MASK] symbol greatly slows down the training and causes pretraining-finetuning inconsistency, due to the large masking ratio (e.g., 60% in SimMIM). On the other hand, MAE does not introduce [MASK] tokens at its encoder at all but is not applicable for hierarchical Vision Transformers. To solve the issue and accelerate the pretraining of hierarchical models, we replace the masked tokens of one image with visible tokens of an- other image, i.e., creating a mixed image. We then conduct dual reconstruction to reconstruct the two original images from the mixed input, which significantly improves efficiency. While MixMAE can be applied to various hierarchical Trans- formers, this paper explores using Swin Transformer with a large window size and scales up to huge model size (to reach 600M parameters). Empirical results demonstrate that Mix- MAE can learn high-quality visual representations efficiently. Notably, MixMAE with Swin-B/W14 achieves 85.1% top-1 accuracy on ImageNet-1K by pretraining for 600 epochs. Besides, its transfer performances on the other 6 datasets show that MixMAE has better FLOPs / performance tradeoff than previous popular MIM methods.	1. Introduction Utilizing unlabeled visual data in self-supervised manners to learn representations is intriguing but challenging. Follow- ing BERT [12] in natural language processing, pretraining with masked image modeling (MIM) shows great success Corresponding author.in learning visual representations for various downstream vision tasks [4, 16, 33, 39, 40], including image classifica- tion [11], object detection [25], semantic segmentation [42], video classification [15], and motor control [39]. While those state-of-the-art methods [4, 16] achieved superior per- formance on vanilla Vision Transformer (ViT) [14, 34], it is still an open question that how to effectively pretrain hierar- chical ViT to purchase further efficiencies [8, 10, 26, 28, 35] on broad vision tasks. In general, existing MIM approaches replace a portion of input tokens with a special [MASK] symbol and aim at re- covering the original image patches [4, 40]. However, using [MASK] symbol leads to two problems. On the one hand, the[MASK] symbol used in pretraining never appears in the finetuning stage, resulting in pretraining-finetuning incon- sistency [12]. On the other hand, the pretrained networks waste much computation on processing the less informative [MASK] symbols, making the pretraining process inefficient. Those problems become severer when a large masking ratio is used [4,16,33,40]. For example, in SimMIM [40], a mask- ing ratio of 60% is used during the pretraining, i.e., 60% of the input tokens are replaced with the [MASK] symbols. As a result, SimMIM needs relatively more epochs (i.e., 800) for pretraining. In addition, as the high masking ratio causes much pretraining-finetuning inconsistency, the performances of SimMIM on downstream tasks are limited. In contrast, MAE [16] does not suffer from the above problems by discarding the masked tokens in the encoder and uses the [MASK] symbols only in the lightweight decoder. MAE utilizes the vanilla ViT [14] as the encoder, which can process the partial input efficiently with the self-attention operation. However, the design also limits the application of MAE on hierarchical ViTs as the hierarchical ViTs cannot process 1D token sequences with arbitrary lengths [28, 35]. In this work, we propose MixMAE, a generalized pre- training method that takes advantage of both SimMIM [40] and MAE [40] while avoiding their limitations. In particular, given two random images from the training set, MixMAE creates a mixed image with random mixing masks as input 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. 6252 ApproachCompatible with hierarchical ViTPretraining efficientPretrain-finetune consistent BEiT [4] ✓ ✗ ✗ SimMIM [40] ✓ ✗ ✗ MAE [16] ✗ ✓ ✓ MixMAE ✓ ✓ ✓ Table 1. Key differences between MixMAE and related works. and trains a hierarchical ViT to reconstruct the two original images to learn visual representations. From one image’s perspective, instead of replacing the masked tokens of the image with the special [MASK] symbols, the masked tokens are replaced by visible tokens of the other image. MixMAE adopts an encoder-decoder design. The encoder is a hierar- chical ViT and processes the mixed image to obtain hidden representations of the two partially masked images. Before the decoding, the hidden representations are unmixed and filled with the [MASK] tokens. Following MAE [16], the decoder is a small ViT to reconstruct the two original images. We illustrate the proposed MixMAE in Figure 1. MixMAE can be widely applied to pretrain different hi- erarchical ViTs, such as Swin Transformer [28], Twins [8], PVT [35], etc. Thanks to the utilization of the hierarchical architecture, we can naturally apply the pretrained encoder to object detection and semantic segmentation tasks. Em- pirically, with similar model sizes and FLOPs, MixMAE consistently outperforms BEiT [4] and MAE [16] on a wide spectrum of downstream tasks, including image classifica- tion on iNaturalist [18] and Places [41], object detection and instance segmentation on COCO [25], and semantic seg- mentation on ADE20K [42]. By abandoning using [MASK] tokens in the encoder, MixMAE shows much better pretrain- ing efficiency than SimMIM [40] on various hierarchical ViTs [8, 28, 35].
Lei_Blind_Video_Deflickering_by_Neural_Filtering_With_a_Flawed_Atlas_CVPR_2023	Abstract Many videos contain flickering artifacts; common causes of flicker include video processing algorithms, video gener- ation algorithms, and capturing videos under specific situ- ations. Prior work usually requires specific guidance such as the flickering frequency, manual annotations, or extra consistent videos to remove the flicker. In this work, we propose a general flicker removal framework that only re- ceives a single flickering video as input without additional guidance. Since it is blind to a specific flickering type or guidance, we name this “blind deflickering. ” The core of our approach is utilizing the neural atlas in cooperation with a neural filtering strategy. The neural atlas is a uni- fied representation for all frames in a video that provides temporal consistency guidance but is flawed in many cases. To this end, a neural network is trained to mimic a filter to learn the consistent features (e.g., color, brightness) and avoid introducing the artifacts in the atlas. To validate our method, we construct a dataset that contains diverse real- world flickering videos. Extensive experiments show that our method achieves satisfying deflickering performance and even outperforms baselines that use extra guidance on a public benchmark. The source code is publicly available at *Equal contribution. †Corresponding authors.https://chenyanglei.github.io/deflicker .	1. Introduction A high-quality video is usually temporally consistent, but many videos suffer from flickering artifacts for var- ious reasons, as shown in Figure 2. For example, the brightness of old movies can be very unstable since some old cameras with low-quality hardware cannot set the ex- posure time of each frame to be the same [16]. Be- sides, high-speed cameras with very short exposure time can capture the high-frequency (e.g., 60 Hz) changes of in- door lighting [25]. Effective processing algorithms such as enhancement [38, 45], colorization [29, 60], and style transfer [33] might bring flicker when applied to tempo- rally consistent videos. Videos from video generations ap- proaches [46, 50, 61] also might contain flickering artifacts. Since temporally consistent videos are generally more vi- sually pleasing, removing the flicker from videos is highly desirable in video processing [9, 13, 14, 54, 59] and compu- tational photography. In this work, we are interested in a general approach for deflickering: (1) it is agnostic to the patterns or levels of flickering (e.g., old movies, high-speed cameras, process- ing artifacts), (2) it only takes a single flickering video and does not require other guidance (e.g., flickering types, ex- tra consistent videos). That is to say, this model is blind to flickering types and guidance, and we name this task as 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. 10439 Frame 1 Frame 2 Old movies Old cartoons Time-lapse Slow-motion Colorization White balance Intrinsic Text-to-video Figure 2. Videos with flickering artifacts. Flickering artifacts exist in unprocessed videos, including old movies, old cartoons, time-lapse videos, and slow-motion videos. Besides, some processing algorithms [6, 18, 60] can introduce the flicker to temporally consistent unpro- cessed videos. Synthesized videos from video generations approaches [46, 50, 61] also might be temporal inconsistent. Blind deflickering aims to remove various types of flickering with only an unprocessed video as input. blind deflickering . Thanks to the blind property, blind de- flickering has very wide applications. Blind deflickering is very challenging since it is hard to enforce temporal consistency across the whole video with- out any extra guidance. Existing techniques usually de- sign specific strategies for each flickering type with specific knowledge. For example, for slow-motion videos captured by high-speed cameras, prior work [25] can analyze the lighting frequency. For videos processed by image process- ing algorithms, blind video temporal consistency [31, 32] obtains long-term consistency by training on a temporally consistent unprocessed video. However, the flickering types or unprocessed videos are not always available, and existing flickering-specific algorithms cannot be applied in this case. One intuitive solution is to use the optical flow to track the correspondences. However, the optical flow from the flick- ering videos is not accurate, and the accumulated errors of optical flow are also increasing with the number of frames due to inaccurate estimation [7]. With two key observations and designs, we successfully propose the first approach for blind deflickering that can re- move various flickering artifacts without extra guidance or knowledge of flicker. First, we utilize a unified video rep- resentation named neural atlas [26] to solve the major chal- lenge of solving long-term inconsistency. This neural atlas tracks all pixels in the video, and correspondences in differ- ent frames share the same pixel in this atlas. Hence, a se- quence of temporally consistent frames can be obtained by sampling from the shared atlas. Secondly, while the frames from the shared atlas are consistent, the structures of images are flawed: the neural atlas cannot easily model dynamic objects with large motion; the optical flow used to construct the atlas is imperfect. Hence, we propose a neural filtering strategy to take the treasure and throw the trash from the flawed atlas. A neural network is trained to learn the invari- ant under two types of distortion, which mimics the artifacts in the atlas and the flicker in the video, respectively. At test time, this network works as a filter to preserve the consis- tency property and block the artifacts from the flawed atlas. We construct the first dataset containing various types of flickering videos to evaluate the performance of blinddeflickering methods faithfully. Extensive experimental re- sults show the effectiveness of our approach in handling dif- ferent flicker. Our approach also outperforms blind video temporal consistency methods that use an extra input video as guidance on a public benchmark. Our contributions can be summarized as follows: • We formulate the problem of blind deflickering and construct a deflickering dataset containing diverse flickering videos for further study. • We propose the first blind deflickering approach to re- move diverse flicker. We introduce the neural atlas to the deflickering problem and design a dedicated strat- egy to filter the flawed atlas for satisfying deflickering performance. • Our method outperforms baselines on our dataset and even outperforms methods that use extra input videos on a public benchmark.
Li_DISC_Learning_From_Noisy_Labels_via_Dynamic_Instance-Specific_Selection_and_CVPR_2023	Abstract Existing studies indicate that deep neural networks (DNNs) can eventually memorize the label noise. We ob- serve that the memorization strength of DNNs towards each instance is different and can be represented by the con- fidence value, which becomes larger and larger during the training process. Based on this, we propose a Dy- namic Instance-specific Selection and Correction method (DISC) for learning from noisy labels (LNL). We first use a two-view-based backbone for image classification, ob- taining confidence for each image from two views. Then we propose a dynamic threshold strategy for each instance, based on the momentum of each instance’s memorization strength in previous epochs to select and correct noisy la- beled data. Benefiting from the dynamic threshold strategy and two-view learning, we can effectively group each in- stance into one of the three subsets (i.e., clean, hard, and purified) based on the prediction consistency and discrep- ancy by two views at each epoch. Finally, we employ differ- ent regularization strategies to conquer subsets with differ- ent degrees of label noise, improving the whole network’s robustness. Comprehensive evaluations on three control- lable and four real-world LNL benchmarks show that our method outperforms the state-of-the-art (SOTA) methods to leverage useful information in noisy data while allevi- ating the pollution of label noise. Code is available at https://github.com/JackYFL/DISC .	1. Introduction Label noise is inevitable in image classification model learning, especially for large-scale database annotations through web-crawling [31,40], crowd-sourcing [52], or pre- This research was supported in part by the National Key R&D Pro- gram of China (grant 2021ZD0111901), and the National Natural Science Foundation of China (grant 62176249). (a) (b) (c) Clean class - 1 feature Class - 2 prototype Noisy class - 1 feature Clean class - 2 feature Noisy class - 2 feature Class - 1 prototype 80 81 75 49 43 73 72 83 65 80 85 95 97 95 96 97 94 94 95 88 82 82 63 78 80 80 (a) (b) (c) Clean class - 1 feature Class - 2 prototype Noisy class - 1 feature Clean class - 2 feature Noisy class - 2 feature Class - 1 prototype 80 81 75 49 43 73 72 83 65 80 85 95 97 95 96 97 94 94 95 88 82 63 78 80Figure 1. An illustration of DNN’s increasing memorization strength during network training. The class prototypes are weights of the DNN classifier, and for simplicity, we only take a two-class case as an example. (a) In the beginning, DNN first fits clean data whose features are closer to class prototypes than noisy data, which is more sparsely distributed in feature space. (b) As training progresses, DNN begins to fit slightly noisy data, some of which can also be classified to its labeled class with relatively high con- fidence. (c) By the end of the training, the DNN has greatly in- creased its memorization strength, and even extremely noisy data can also be grouped into its labeled class with high confidence. trained models [12], etc. Recent studies show that DNNs are susceptible to label noise and could fit to the entire data set [2, 55] including the noisy set. Meanwhile, researchers found that DNNs have a memorization effect [2], i.e., the learning process of DNNs follows a curriculum, in which simple patterns are memorized first, followed by more diffi- cult ones like data with noisy labels. Recent studies have explored the use of memorization effect for LNL tasks, with many of these approaches being "early-learning"- based methods [1, 7, 11, 17, 23, 24, 29, 30, 32, 44, 53, 57, 58]. These methods leverage an early-stage DNN to improve the model robustness and generalization ability. Early-learning-based LNL methods can be divided into three main directions: sample selection [7,17,23,24,29,53], label correction [1,30,44,57] and regularization [11,32,37, 56, 58, 60]. Sample selection-based methods usually utilize the early-stage DNN’s losses or confidence to select reliable 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. 24070 1 2 N 2 1 Confidence Instances K(a) 1 2 N 2 1 K Confidence Instances (b) 1 2 N K 2 1 Condence Instances (c) Figure 2. An illustration of different threshold strategies, including (a) the global threshold, (b) the class-wise threshold, and (c) the proposed dynamic instance-specific threshold. instances, which are utilized to update the network. Some of these methods require a predefined threshold [24, 30] or prior knowledge about the noise label rate [17, 53] to select instances. Such a global predefined threshold, as shown in Fig. 2 (a), is usually difficult to determine, and may require prior knowledge about the noise label rate to avoid exces- sive or insufficient noisy data selection, which will further lead to over-fitting and confirmation bias issues [17]. La- bel correction-based methods try to learn [44] or generate pseudo-labels [30, 57] to replace the original noisy ones. Many of these methods employ the semi-supervised learn- ing (SSL) techniques to pseudo-label the noisy data, and most of them use global (say MixMatch [3], FixMatch [39]) or class-wise threshold (say FlexMatch [54]) to recalibrate labels. As shown in Fig. 2 (b), while a class-wise thresh- old considers the fitting difficulty of different classes, it still applies a uniform threshold for individual instances in each class. This remains sub-optimum if we consider an example of face images, in which profile face images are more dif- ficult to fit (relatively lower classification confidence) than frontal ones (relatively higher classification confidence) of the same subject. Regularization-based methods aim to de- sign robust loss functions [11, 32, 37, 58, 60, 61] or regular- ization techniques such as augmentation [56] that can uti- lize all instances to improve the model robustness against label noise. While these methods work well on moderately noisy data, they may have poor generalization ability under extremely noisy data (see Table 1), since all instances are utilized during the training process. Based on the observation that the memorization strength for individual instances increases during network training, we argue that neither a global threshold nor a class-wise threshold is optimum for LNL. Therefore, we propose a Dy- namic Instance-specific Selection and Correction (DISC) approach (see Fig. 3 (a)) for LNL. DISC leverages a dy- namic instance-specific threshold strategy (Fig. 2 (c)) fol- lowing a memorization curriculum to select reliable in- stances and correct noisy labels. Each threshold is ob- tained through the momentum of each instance’s memoriza- tion strength in previous epochs. Such a dynamic threshold strategy can determine a reasonable threshold for each in- stance according to its memorization strength by the net- work. Inspired by previous methods of RRL [30], AugDisc [35] and FixMatch [39], DISC also adopts weak and strongaugmentations to produce two different views for image classification via a shared-weight model. Unlike previous methods, which use predictions from one view to select reliable instances or generate pseudo-labels for unlabeled data, DISC considers the consistency and discrepancy of two views and divides the noisy data into reliable instances (clean set), hard instances (hard set), and recalibrate noisy labeled instances (purified set), reflecting different degrees of label noise. By dividing the noisy data into three differ- ent subsets, DISC can alleviate the contamination of noisy labels to LNL model learning by conquering them via dif- ferent regularization strategies. As a result, the method can better make full use of the whole noisy dataset. The contri- butions of this paper include: • We observe the memorization strength of DNNs towards individual instances can be represented by confidence value, which increases along with training. We provide evidence and experimental analyses to validate this claim. • Based on the insight of memorization strength, we pro- pose a simple yet effective dynamic instance-specific threshold strategy of LNL that selects reliable instances and recalibrates noisy labels following an easy to hard curriculum. • Additionally, we leverage the dynamic threshold strategy to group noisy data into three subsets based on predictions from two views generated from weak and strong augmen- tations. We then adopt different regularization strategies to handle individual subsets.
Kong_Understanding_Masked_Autoencoders_via_Hierarchical_Latent_Variable_Models_CVPR_2023	Abstract Masked autoencoder (MAE), a simple and effective self- supervised learning framework based on the reconstruction of masked image regions, has recently achieved prominent success in a variety of vision tasks. Despite the emergence of intriguing empirical observations on MAE, a theoreti- cally principled understanding is still lacking. In this work, we formally characterize and justify existing empirical in- sights and provide theoretical guarantees of MAE. We for- mulate the underlying data-generating process as a hierar- chical latent variable model, and show that under reason- able assumptions, MAE provably identifies a set of latent variables in the hierarchical model, explaining why MAE can extract high-level information from pixels. Further, we show how key hyperparameters in MAE (the masking ratio and the patch size) determine which true latent variables to be recovered, therefore influencing the level of semantic information in the representation. Specifically, extremely large or small masking ratios inevitably lead to low-level representations. Our theory offers coherent explanations of existing empirical observations and provides insights for potential empirical improvements and fundamental limita- tions of the masked-reconstruction paradigm. We conduct extensive experiments to validate our theoretical insights.	1. Introduction Self-supervised learning (SSL) has achieved tremendous success in learning transferable representations without la- bels, showing strong results in a variety of downstream tasks [12,14, 16,23,49]. As a major SSL paradigm, masked image modeling (MIM) [1–3, 11,13,22,41,63,69] performs the reconstruction of purposely masked image pixels as the pretraining task. Among MIM methods, masked autoen- coding (MAE) [22] has gained significant traction due to its computational efficiency and state-of-the-art performance in a wide range of downstream tasks. Empirical observations from previous work reveal vari- ous intriguing properties of MAE. In particular, aggressive *Joint first author. †Joint senior author. Figure 1. Masking-reconstruction under a hierarchical generating process. In a hierarchical data-generating process, high-level latent vari- ables (e.g., z1) represent high-level information such as semantics, and low-level latent variables (e.g., [z2,z3,z4]) represent low-level informa- tion such as texture. We show that through proper masking, MAE learns to recover high-level latent variables with identifiability guarantees. masking has been shown critical to downstream task per- formances [22, 28,61,63]. It is conjectured that such mask- ing forces the model to learn meaningful high-level seman- tic understanding of the objects and scenes rather than the low-level information such as texture. However, it remains largely unclear whether such intuitions are sound in princi- ple. Theoretically verifying and characterizing these empir- ical insights would not only grant a certificate to the current approaches but would also offer theoretical insights for al- gorithmic advancements. In this work, we establish a principled yet intuitive framework for understanding MAE and providing identifia- bility guarantees. Concretely, we first formulate the under- lying data-generating process as a hierarchical latent vari- able model (Figure 1), with high-level variables correspond- ing to abstract and semantic information like classes, and low-level variables corresponding to elaborate and granular information like texture. Such latent variable models have been studied in causal discovery [29, 62]. In [27, 50], it is hypothesized that complex data, such as images, follow a hierarchical latent structure. Stemming from this formulation, we show that under reasonable assumptions, MAE can recover a subset of 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. 7918 true latent variables within the hierarchy, where the levels of the learned latent variables are explicitly determined by how masking is performed. Our theoretical framework not only unifies existing empirical observations in a coherent fashion but also gives rise to insights for potential empir- ical improvemen
Kim_Coreset_Sampling_From_Open-Set_for_Fine-Grained_Self-Supervised_Learning_CVPR_2023	Abstract Deep learning in general domains has constantly been extended to domain-speciﬁc tasks requiring the recognition of ﬁne-grained characteristics. However, real-world appli- cations for ﬁne-grained tasks suffer from two challenges: a high reliance on expert knowledge for annotation and ne- cessity of a versatile model for various downstream tasks in a speciﬁc domain (e.g., prediction of categories, bounding boxes, or pixel-wise annotations). Fortunately, the recent self-supervised learning (SSL) is a promising approach to pretrain a model without annotations, serving as an effective initialization for any downstream tasks. Since SSL does not rely on the presence of annotation, in general, it utilizes the large-scale unlabeled dataset, referred to as an open-set. In this sense, we introduce a novel Open-Set Self-Supervised Learning problem under the assumption that a large-scale unlabeled open-set is available, as well as the ﬁne-grained target dataset, during a pretraining phase. In our problem setup, it is crucial to consider the distribution mismatch be- tween the open-set and target dataset. Hence, we propose SimCore algorithm to sample a coreset, the subset of an open-set that has a minimum distance to the target dataset in the latent space. We demonstrate that SimCore signiﬁcantly improves representation learning performance through ex- tensive experimental settings, including eleven ﬁne-grained datasets and seven open-sets in various downstream tasks.	1. Introduction The success of deep learning in general computer vision tasks has encouraged its widespread applications to speciﬁc domains of industry and research [ 21,46,73], such as facial recognition or vehicle identiﬁcation. We particularly focus on the visual recognition of ﬁne-grained datasets, where the goal is to differentiate between hard-to-distinguish images. However, real-world application for ﬁne-grained tasks poses two challenges for practitioners and researchers developing algorithms. First, it requires a number of experts for anno- *equal contributiontation, which incurs a large cost [ 3,14,60]. For example, ordinary people do not have professional knowledge about aircraft types or ﬁne-grained categories of birds. Therefore, a realistic presumption for a domain-speciﬁc ﬁne-grained dataset is that there may be no or very few labeled samples. Second, ﬁne-grained datasets are often re-purposed or used for various tasks according to the user’s demand, which mo- tivates development of a versatile pretrained model. One might ask, as a target task, that bird images be classiﬁed by species or even segmented into foreground and back- ground. A good initialization model can handle a variety of annotations for ﬁne-grained datasets, such as multiple attributes [ 43,47,74], pixel-level annotations [ 53,74], or bounding boxes [ 35,47,53]. Recently, self-supervised learning (SSL) [ 11,14,24,28] has enabled learning how to represent the data even without annotations, such that the representations serve as an effec- tive initialization for any future downstream tasks. Since labeling is not necessary, SSL generally utilizes an open- set, or large-scale unlabeled dataset, which can be easily obtained by web crawling [ 23,64], for the pretraining. In this paper, we introduce a novel Open-Set Self-Supervised Learning (OpenSSL) problem, where we can leverage the open-set as well as the training set of ﬁne-grained target dataset. Refer to Figure 1for the overview of OpenSSL. In the OpenSSL setup, since the open-set may contain instances irrelevant to the ﬁne-grained dataset, we should consider the distribution mismatch. A large distribution mismatch might inhibit representation learning for the target task. For instance, in Figure 2, SSL on the open-set ( OS) does not always outperform SSL on the ﬁne-grained dataset ( X) because it depends on the semantic similarity between Xand OS. This is in line with the previous observations [ 21,22,64] that the performance of self-supervised representation on downstream tasks is correlated with similarity of pretraining and ﬁne-tuning datasets. To alleviate this mismatch issue, we could exploit a core- set, a subset of an open-set, which shares similar semantics with the target dataset. As a motivating experiment, we man- ually selected the relevant classes from ImageNet ( OSoracle) that are supposed to be helpful according to each target 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. 7537 ? ObjectDetectionSemanticSegmentationImageClassificationSelf-SupervisedPretrainingDownstream Tasks airplane: 93%manufacturervariant Open-SetFine-Grained Dataset?+Coreset Figure 1. Overview of an OpenSSL problem. For any downstream tasks, we pretrain an effective model with the ﬁne-grained dataset via self-supervised learning (SSL). Here, the assumption for a large-scale unlabeled open-set in the pretraining phase is well-suited for a real-world scenario. The main goal is to ﬁnd a coreset, highlighted by the blue box, among the open-set to enhance ﬁne-grained SSL.XOSX+OSX+OSrandX+OSoracle Aircraft30%35%40%45%50%Cars30%40%50%60%Pet40%50%60%70%80%Birds20%25%30%35%40%Target (X)Classes forOSoracle(#)Aircraftairliner, warplane, ... (8)Carsconvertible, jeep, ... (10)PetPersian cat, beagle, ... (24)Birdsgoldﬁnch, junco, ... (20)Figure 2. Linear evaluation performance on the ﬁne-grained target dataset. Each color corresponds to a pretraining dataset, while “+” means merging two datasets. The table on the right side shows the manually selected categories from an open-set ( OS), ImageNet-1k [ 20] in this case, according to each target dataset ( X). Selected categories and exact numbers are detailed in Appendix A. We followed the typical linear evaluation protocol [ 14,24] and used the SimCLR method [ 14] on ResNet50 encoder [ 29]. dataset. Interestingly, in Figure 2, merging OSoracle toX shows a signiﬁcant performance gain, and its superiority over merging the entire open-set ( X+OS) or the randomly sampled subset ( X+OSrand) implies the necessity of a sam- pling algorithm for the coreset in the OpenSSL problem. Therefore, we propose SimCore , a simple yet effective coreset sampling algorithm from an unlabeled open-set. Our main goal is to ﬁnd a subset semantically similar to the target dataset. We formulate the data subset selection problem to obtain a coreset that has a minimum distance to the target dataset in the latent space. SimCore signiﬁcantly improves performance in extensive experimental settings (eleven ﬁne- grained datasets and seven open-sets), and shows consistent gains with different architectures, SSL losses, and down- stream tasks. Our contributions are outlined as follows: •We ﬁrst propose a realistic OpenSSL task, assuming an unlabeled open-set available during the pretraining phase on the ﬁne-grained dataset. •We propose a coreset selection algorithm, SimCore, to leverage a subset semantically similar to the target dataset. •Our extensive experiments with eleven ﬁne-grained datasets and seven open-sets substantiate the signiﬁcance of data selection in our OpenSSL problem.
Lai_Spherical_Transformer_for_LiDAR-Based_3D_Recognition_CVPR_2023	Abstract LiDAR-based 3D point cloud recognition has benefited various applications. Without specially considering the Li- DAR point distribution, most current methods suffer from information disconnection and limited receptive field, es- pecially for the sparse distant points. In this work, we study the varying-sparsity distribution of LiDAR points and present SphereFormer to directly aggregate information from dense close points to the sparse distant ones. We de- sign radial window self-attention that partitions the space into multiple non-overlapping narrow and long windows. It overcomes the disconnection issue and enlarges the re- ceptive field smoothly and dramatically, which significantly boosts the performance of sparse distant points. Moreover, to fit the narrow and long windows, we propose exponen- tial splitting to yield fine-grained position encoding and dynamic feature selection to increase model representation ability. Notably, our method ranks 1ston both nuScenes and SemanticKITTI semantic segmentation benchmarks with 81.9%and74.8%mIoU, respectively. Also, we achieve the 3rdplace on nuScenes object detection benchmark with 72.8%NDS and 68.5%mAP . Code is available at https: //github.com/dvlab-research/SphereFormer.git.	1. Introduction Nowadays, point clouds can be easily collected by Li- DAR sensors. They are extensively used in various indus- trial applications, such as autonomous driving and robotics. In contrast to 2D images where pixels are arranged densely and regularly, LiDAR point clouds possess the varying- sparsity property — points near the LiDAR are quite dense, while points far away from the sensor are much sparser, as shown in Fig. 2 (a). However, most existing work [12, 13, 24, 25, 55, 70–72] does not specially consider the the varying-sparsity point distribution of outdoor LiDAR point clouds. They inherit from 2D CNNs or 3D indoor scenarios, and conduct local operators ( e.g., SparseConv [24, 25]) uniformly for all lo- cations. This causes inferior results for the sparse distant points. As shown in Fig. 1, although decent performance 78.7951.54 13.2879.2154.31 19.3180.860.7830.38 0102030405060708090 close (<20m)medium (>=20m, <50m)far (>=50m)SparseConvCubic Self-attentionOurs(%) mIoUFigure 1. Semantic segmentation performance on nuScenes valset for points at different distances. is yielded for the dense close points, it is difficult for these methods to deal with the sparse distant points optimally. We note that the root cause lies in limited receptive field. For sparse distant points, there are few surrounding neigh- bors. This not only results in inconclusive features, but also hinders enlarging receptive field due to information discon- nection. To verify this finding, we visualize the Effective Receptive Field (ERF) [40] of the given feature (shown with the yellow star) in Fig. 2 (d). The ERF cannot be expanded due to disconnection, which is caused by the extreme spar- sity of the distant car. Although window self-attention [22, 30], dilated self- attention [42], and large-kernel CNN [10] have been pro- posed to conquer the limited receptive field, these methods do not specially deal with LiDAR point distribution, and re- main to enlarge receptive field by stacking local operators as before, leaving the information disconnection issue still unsolved. As shown in Fig. 1, the method of cubic self- attention brings a limited improvement. In this paper, we take a new direction to aggregate long- range information directly in a single operator to suit the varying-sparsity point distribution. We propose the module ofSphereFormer to perceive useful information from points 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. 17545 (a) LiDAR Point Cloud(b) Radial Window Partition (c) Sparse Distant Points(d) ERF of SparseConv(e) ERF of our method carroad vegetation building pole fence trunkFigure 2. Effective Receptive Field (ERF) of SparseConv and ours. (a) LiDAR point cloud. (b) Radial window partition. Only a single radial window is shown. Points inside the window are marked in red. (c) Zoom-in sparse distant points. A sparse caris circled in yellow. (d) ERF of SparseConv, given the point of interest (with yellow star). White and red denote high contribution. (e) ERF of ours. 50+ meters away and yield large receptive field for feature extraction. Specifically, we represent the 3D space using spherical coordinates (r, θ, ϕ )with the sensor being the ori- gin, and partition the scene into multiple non-overlapping windows. Unlike the cubic window shape, we design radial windows that are long and narrow. They are obtained by partitioning only along the θandϕaxis, as shown in Fig. 2 (b). It is noteworthy that we make it a plugin module to conveniently insert into existing mainstream backbones. The proposed module does not rely on stacking local op- erators to expand receptive field, thus avoiding the discon- nection issue, as shown in Fig. 2 (e). Also, it facilitates the sparse distant points to aggregate information from the dense-point region, which is often semantically rich. So, the performance of the distant points can be improved sig- nificantly ( i.e., +17.1% mIoU) as illustrated in Fig. 1. Moreover, to fit the long and narrow radial windows, we propose exponential splitting to obtain fine-grained relative position encoding. The radius rof a radial window can be over 50 meters, which causes large splitting intervals. It thus results in coarse position encoding when converting relative positions into integer indices. Besides, to let points at varying locations treat local and global information dif- ferently, we propose dynamic feature selection to make fur- ther improvements. In total, our contribution is three-fold. • We propose SphereFormer to directly aggregate long- range information from dense-point region. It in- creases the receptive field smoothly and helps improvethe performance of sparse distant points . • To accommodate the radial windows, we develop ex- ponential splitting for relative position encoding. Our dynamic feature selection further boosts performance. • Our method achieves new state-of-the-art results on multiple benchmarks of both semantic segmentation and object detection tasks.
Liu_Revisiting_Temporal_Modeling_for_CLIP-Based_Image-to-Video_Knowledge_Transferring_CVPR_2023	Abstract Image-text pretrained models, e.g., CLIP , have shown impressive general multi-modal knowledge learned from large-scale image-text data pairs, thus attracting increas- ing attention for their potential to improve visual represen- tation learning in the video domain. In this paper, based on the CLIP model, we revisit temporal modeling in the context of image-to-video knowledge transferring, which is the key point for extending image-text pretrained models to the video domain. We find that current temporal model- ing mechanisms are tailored to either high-level semantic- dominant tasks (e.g., retrieval) or low-level visual pattern- dominant tasks (e.g., recognition), and fail to work on the two cases simultaneously. The key difficulty lies in modeling temporal dependency while taking advantage of both high- level and low-level knowledge in CLIP model. To tackle this problem, we present Spatial-Temporal Auxiliary Net- work (STAN) – a simple and effective temporal modeling mechanism extending CLIP model to diverse video tasks. Specifically, to realize both low-level and high-level knowl- edge transferring, STAN adopts a branch structure with decomposed spatial-temporal modules that enable multi- level CLIP features to be spatial-temporally contextual- ized. We evaluate our method on two representative video tasks: Video-Text Retrieval and Video Recognition. Exten- sive experiments demonstrate the superiority of our model over the state-of-the-art methods on various datasets, in- cluding MSR-VTT, DiDeMo, LSMDC, MSVD, Kinetics-400, and Something-Something-V2. Codes will be available at https://github.com/farewellthree/STAN	1. Introduction Recent years have witnessed the great success of image- text pretrained models such as CLIP [31]. Pretrained on over 400M image-text data pairs, these models learned transferable rich knowledge for various image understand- ing tasks. Similarly, video domains also call for a CLIP -likemodel to solve downstream video tasks. However, it is hard to get a pretrained model as powerful as CLIP in the video domain due to the unaffordable demands on computation re- sources and the difficulty of collecting video-text data pairs as large and diverse as image-text data. Instead of directly pursuing video-text pretrained models [16, 26], a potential alternative solution that benefits video downstream tasks is to transfer the knowledge in image-text pretrained models to the video domain, which has attracted increasing atten- tion in recent years [11, 12, 25, 28, 29, 40]. Extending pretrained 2D image models to the video do- main is a widely-studied topic in deep learning [4, 7], and the key point lies in empowering 2D models with the ca- pability of modeling temporal dependency between video frames while taking advantages of knowledge in the pre- trained models. In this paper, based on CLIP [31], we revisit temporal modeling in the context of image-to-video knowl- edge transferring, and present Spatial-Temporal Auxiliary Network (STAN) – a new temporal modeling method that is easy and effective for extending image-text pretrained model to diverse downstream video tasks. We find that current efforts on empowering CLIP with temporal modeling capability can be roughly divided into posterior structure based methods and intermediate struc- ture based methods as shown in Fig. 1(a). Posterior struc- ture based methods [11,12,25] employ a late modeling strat- egy, which take CLIP as a feature extractor and conduct temporal modeling upon the embeddings of video frames extracted independently from CLIP . Upon the highly se- mantic embeddings, though the structure is beneficial for transferring the well-aligned visual-language representation (i.e.,high-level knowledge) to downstream tasks, it hardly captures the spatial-temporal visual patterns ( i.e.,low-level knowledge) among different frames, which is important for video understanding. As shown in Fig. 1(b), compared to theCLIP baseline that employs a naive mean pooling to aggregate the features of all frames to obtain a video rep- resentation, the performance improvement brought by the typical posterior structure, i.e.CLIP4clip-seqTrans [25] is trivial, especially on the video action recognition task where 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. 6555 VideoOutput⊕ VideoOutput VideoOutput CLIP layerTemporal modeling layerData flow(a)1.40.23.80.53.74.2 012345Improvement over CLIP Retrieval (R@1)Recognition (Top1 Acc) (b)Posteriorstructure(CLIP4clip-seqTrans)Intermediatestructure(XCLIP)Branchstructure(Ours)Figure 1. (a) Illustration of temporal modeling with posterior structure (left), intermediate structure (middle) and our branch structure(right). (b) Performance comparison among the posterior structure based CLIP4clip-seqTrans [25] , intermediate structure based XCLIP [28] and our branch structure based STAN. We take the CLIP model with a naive mean pooling to aggregate the features of all frames into video representations as the baseline. We present the improvement brought by different methods over this baseline w.r.t. Recall@1 on MSRVTT for video-text retrieval and Top-1 accuracy on Kinetics-400 for video recognition. spatial-temporal visual patterns are crucial. In contrast to posterior structure based methods, inter- mediate structure based methods [4, 28, 29] strengthen the spatial-temporal modeling capability of CLIP via plugging temporal modeling modules directly between CLIP layers, and achieve 3.7% improvement over the baseline on the video action recognition task. However, we find that in- serting additional modules into CLIP would impact the pre- trained high-level semantic knowledge in the model, which only outperforms the baseline by 0.2% on the video-text retrieval tasks. Therefore, modeling temporal dependency while taking advantage of knowledge in different levels of representation is important for extending the CLIP model to the video domain. Unlike the above methods, inspired by FPN [22] that introduces a branch network to strengthen multi-level rep- resentation learning for CNNs, our proposed STAN em- ploys a new branch structure outside of the visual back- bone, as shown in Fig. 1(a). Thanks to the branch structure, STAN augments the features of video frames with spatial- temporal contexts at different CLIP output levels without affecting the forward-propagating of CLIP itself. Thus, it is able to take advantage of both high-level and low-level knowledge in the pretrained model simultaneously, and ef- fectively extends CLIP to diverse downstream video tasks. STAN consists of multiple layers with a spatial-temporal separated design. Specifically, the layer operates spatial- temporal modeling via alternatively stacking two separate modules – an intra-frame module and a cross-frame module, which enables the layer to boost the performance of model via reusing the pretrained parameter of CLIP layers to ini- tialize the intra-frame spatial modules. We further investi- gate two instantiations of cross-frame modules, i.e.,the self- attention-based module and 3D convolution based module, to facilitate the comprehensive understanding of STAN indifferent implementations. We evaluate our STAN on both the high-level semantic- dominant task ( i.e.,video-text retrieval) and low-level vi- sual pattern-dominant task ( i.e.,, video recognition), trial- ing our methods from the two different perspectives. Ex- tensive experiments demonstrate our expanded models are generally effective on the two different tasks. For video- text retrieval, we surpass the CLIP4clip by +3.7%, +3.1%, and +2.1% R@1 on MSRVTT, DiDemo, and LSMDC. For video recognition, we achieve competitive performance on Kinetics-400, with 88×fewer FLOPs than Swin3D-L [24] and improve CLIP baseline by 20%+ on Something- Something-V2. Our main contributions are summarized as: (1) we re- visit temporal modeling in the context of image-to-video knowledge transferring and figure out that the key challenge lies in modeling temporal dependency while taking advan- tage of both high-level and low-level knowledge; (2) we propose Spatial-Temporal Auxiliary Network (STAN) – a new branch structure for temporal modeling, which facil- itates representation learning of video frames with includ- ing spatial-temporal contexts at different levels and better transfer the pretrained knowledge in CLIP to diverse video tasks; (3) our method achieves competitive results on both video-text retrieval and video recognition tasks compared to SOTA methods.
Lee_Fix_the_Noise_Disentangling_Source_Feature_for_Controllable_Domain_Translation_CVPR_2023	Abstract Recent studies show strong generative performance in domain translation especially by using transfer learning techniques on the unconditional generator. However, the control between different domain features using a single model is still challenging. Existing methods often require additional models, which is computationally demanding and leads to unsatisfactory visual quality. In addition, they have restricted control steps, which prevents a smooth tran- sition. In this paper, we propose a new approach for high- quality domain translation with better controllability. The key idea is to preserve source features within a disentangled subspace of a target feature space. This allows our method to smoothly control the degree to which it preserves source features while generating images from an entirely new do- main using only a single model. Our extensive experiments show that the proposed method can produce more consistent and realistic images than previous works and maintain pre- cise controllability over different levels of transformation. The code is available at LeeDongYeun/FixNoise.	1. Introduction Image translation between different domains is a long- standing problem in computer vision [8, 9,13,20,22,24, 35,52,62]. Controllability in domain translation is impor- tant since it allows the users to set the desired properties. Recently, several studies have shown promising results in domain translation using a pre-trained unconditional gen- erator, such as StyleGAN2 [27], and its fine-tuned version [29,30,42,48]. These studies implemented domain trans- lation by embedding an image from the source domain to the latent space of the source model and by providing the obtained latent code into the target model to generate a tar- get domain image. To preserve semantic correspondence between different domains, previous works commonly fo- cused on the hierarchical design of the unconditional gener- ator. They used several techniques like freezing [30] and swapping [42] layers or both [29]. In these approaches, users can control the degree of preserved source features by setting the number of freezing or swapping layers of the target model differently. However, one of the notable limitations of the previous 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. 14224 methods is that they cannot control features across domains in a single model. Imagine morphing between two images x0andx1. Previous methods approximated midpoints be- tween x0andx1by either building a new hybrid model by converting weights or training a new model. In these ap- proaches, each intermediate point is drawn from the output distribution of different models, which would produce in- consistent results. Moreover, getting an additional model for each intermediate point (image) also increases the com- putational cost. Another common limitation of these layer- based methods is that their control levels are discrete and restricted to the number of layers, which prevents fine-grain control. In this paper, we introduce a new training strategy, FixNoise, for cross-domain controllable domain translation. To control features across domains in a single model, we argue that the source features should be preserved but dis- entangled with the target in the model’s inherited space. To this end, we focus on the fact that the noise input of Style- GAN2, which is added after each convolution, expands the functional space composed of the latent code expression. In other words, the feature space could be seen as a set of sub- spaces corresponding to each random noise. To preserve the source features only to a particular subset of the fea- ture space of the target model, we fix the noise input when applying a simple feature matching loss. The disentangled feature space allows our method to fine-grain control the preserved source features only in a single model without limited control steps through linear interpolation between the fixed and random noise. The extensive experiments demonstrate that our approach can generate more consistent and realistic results than existing methods on cross-domain feature control and also show better performance on domain translation qualitatively and quantitatively.
Li_Efficient_and_Explicit_Modelling_of_Image_Hierarchies_for_Image_Restoration_CVPR_2023	Abstract The aim of this paper is to propose a mechanism to efficiently and explicitly model image hierarchies in the global, regional, and local range for image restoration. To achieve that, we start by analyzing two important prop- erties of natural images including cross-scale similarity and anisotropic image features. Inspired by that, we pro- pose the anchored stripe self-attention which achieves a good balance between the space and time complexity of self-attention and the modelling capacity beyond the re- gional range. Then we propose a new network architec- ture dubbed GRL to explicitly model image hierarchies in the G lobal, R egional, and L ocal range via anchored stripe self-attention, window self-attention, and channel attention enhanced convolution. Finally, the proposed network is ap- plied to 7 image restoration types, covering both real and synthetic settings. The proposed method sets the new state- of-the-art for several of those. Code will be available at https://github.com/ofsoundof/GRL-Image- Restoration.git .	1. Introduction Image restoration aims at recovering high-quality images from low-quality ones, resulting from an image degrada- tion processes such as blurring, sub-sampling, noise cor- ruption, and JPEG compression. Image restoration is an ill- posed inverse problem since important content information about the image is missing during the image degradation processes. Thus, in order to recover a high-quality image, the rich information exhibited in the degraded image should be fully exploited. Natural images contain a hierarchy of features at global, regional, and local ranges which could be used by deep neu- ral networks for image restoration. First , the local range covers a span of several pixels and typical features are edges and local colors. To model such local features, convo- lutional neural networks (CNNs) with small kernels ( e.g. 3×3) are utilized. Second , the regional range is charac- terized by a window with tens of pixels. This range of pix- (a)bridge from ICB, 2749×4049(b)0848x4 from DIV2K, 1020×768 (c)073from Urban100, 1024×765 Figure 1. Natural images show a hierarchy of features in a global, regional, and local range. The local (edges, colors) and regional features (the pink squares) could be well modelled by CNNs and window self-attention. By contrast, it is difficult to efficiently and explicitly model the rich global features (cyan rectangles). els can cover small objects and components of large objects (pink squares in Fig. 1). Due to the larger range, modelling the regional features (consistency, similarity) explicitly with large-kernel CNNs would be inefficient in both parameters and computation. Instead, transformers with a window at- tention mechanism are well suited for this task. Third , be- yond local and regional, some features have a global span (cyan rectangles in Fig. 1), incl. but not limited to symme- try, multi-scale pattern repetition (Fig. 1a), same scale tex- ture similarity (Fig. 1b), and structural similarity and con- sistency in large objects and content (Fig. 1c). To model features at this range, global image understanding is needed. Different from the local and regional range features, there are two major challenges to model the global range features. Firstly, existing image restoration networks based on convolutions and window attention could not capture long-range dependencies explicitly by using a single com- putational module. Although non-local operations are used in some works, they are either used sparsely in the network or applied to small image crops. Thus, global image under- 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. 18278 Figure 2. The proposed GRL achieves state-of-the-art performances on various image restoration tasks. Details provided in Sec. 5. standing still mainly happens via progressive propagation of features through repeated computational modules. Sec- ondly, the increasing resolution of today’s images poses a challenge for long-range dependency modelling. High im- age resolution leads to a computational burden associated with pairwise pixel comparisons and similarity searches. The aforementioned discussion leads to a series of re- search questions: 1) how to efficiently model global range features in high-dimensional images for image restora- tion; 2) how to model image hierarchies (local, regional, global) explicitly by a single computational module for high-dimensional image restoration; 3) and how can this joint modelling lead to a uniform performance improvement for different image restoration tasks. The paper tries to an- swer these questions in Sec. 3, Sec. 4, and Sec. 5, resp. First , we propose anchored stripe self-attention for effi- cient dependency modelling beyond the regional range. The proposed self-attention is inspired by two properties of nat- ural images including cross-scale similarity and anisotropic image features. Cross-scale similarity means that struc- tures in a natural image are replicated at different scales. Inspired by that, we propose to use anchors as an inter- mediate to approximate the exact attention map between queries and keys in self-attention. Since the anchors sum- marize image information into a lower-dimensional space, the space and time complexity of self-attention can be sig- nificantly reduced. In addition, based on the observation of anisotropic image features, we propose to conduct an- chored self-attention within vertical and horizontal stripes. Due to the anisotropic shrinkage of the attention range, a further reduction of complexity is achieved. And the com- bination of axial stripes also ensures a global view of the image content. When equipped with the stripe shift oper- ation, the four stripe self-attention modes (horizontal, ver- tical, shifted horizontal, shifted vertical) achieves a good balance between computational complexity and the capac- ity of global range dependency modelling. Furthermore, the proposed anchored stripe self-attention is analyzed from the perspective of low-rankness and similarity propagation. Secondly , a new transformer network is proposed to ex- plicitly model global, regional, and local range dependen- cies in a single computational module. The hierarchical modelling of images is achieved by the parallel computa- tion of the proposed anchored stripe self-attention, window self-attention, and channel-attention enhanced convolution. And the transformer architecture is dubbed GRL .Thirdly , the proposed GRL transformer is applied to var- ious image restoration tasks. Those tasks could be classi- fied into three settings based on the availability of data in- cluding real image restoration, synthetic image restoration, and data synthesis based real image restoration. In total, seven tasks are explored for the proposed network includ- ing image super-resolution, image denoising, JPEG com- pression artifacts removal, demosaicking, real image super- resolution, single image motion deblurring, and defocus de- blurring. As shown in Fig. 2, the proposed network shows promising results on the investigated tasks.
Kim_DCFace_Synthetic_Face_Generation_With_Dual_Condition_Diffusion_Model_CVPR_2023	Abstract Generating synthetic datasets for training face recogni- tion models is challenging because dataset generation en- tails more than creating high fidelity images. It involves generating multiple images of same subjects under different factors (e.g., variations in pose, illumination, expression, aging and occlusion) which follows the real image condi- tional distribution. Previous works have studied the gener- ation of synthetic datasets using GAN or 3D models. In this work, we approach the problem from the aspect of combin- ing subject appearance (ID) and external factor (style) con- ditions. These two conditions provide a direct way to con- trol the inter-class and intra-class variations. To this end, we propose a Dual Condition Face Generator (DCFace) based on a diffusion model. Our novel Patch-wise style ex- tractor and Time-step dependent ID loss enables DCFace to consistently produce face images of the same subject under different styles with precise control. Face recognition mod- els trained on synthetic images from the proposed DCFace provide higher verification accuracies compared to previ- ous works by 6.11% on average in 4out of 5test datasets, LFW, CFP-FP , CPLFW, AgeDB and CALFW. Code Link	1. Introduction What does it take to create a good training dataset for visual recognition? An ideal training dataset for recogni- tion tasks would have 1) large inter-class variation, 2) large intra-class variation and 3) small label noise. In the context of face recognition (FR), it means, the dataset has a large number of unique subjects, large intra-subject variations, and reliable subject labels. For instance, large-scale face datasets such as WebFace4M [73] contain over 1M subjects and large number of images/subject. Both the number of subjects and the number of images per subject are impor- tant for training FR models [11,31]. Also, datasets amassed by crawling the web are not free from label noise [7, 73]. In various domains, synthetic datasets are traditionally 1. Subject Variation 2. Style Variation 3. Subject ConsistencyAC B LabeledDatasetFigure 1. Illustration of three factors that characterize a labeled face dataset. It contains large subject variation, style variation and label consistency. Synthetic face datasets should be created with all three factors in mind. Face images in this figure are samples generated by our proposed method which combines arbitrary ID condition with style condition while preserving subject identity. used to help generalize deep models when only limited real datasets could be collected [13, 21, 63, 74] or when bias ex- ists in the real dataset [34, 64]. Lately, more attention has been drawn to training with only synthetic datasets in the face domain, as synthetic data can avoid leaking the privacy of real individuals. This is important as real face datasets have been under scrutiny for their lack of informed consent, as web-crawling is the primary means of large-scale data collection [17, 22, 73]. Also, synthetic training datasets can remedy some long-standing issues in real datasets, e.g.the long tail distribution, demographic bias, etc. When it comes to generating synthetic training datasets, the following questions should be raised. (i) How many novel subjects can be synthesized (ii) How well can we mimic the distribution of real images in the target domain and (iii) How well can we consistently generate multiple images of the same subjects? We start with the hypothesis that face dataset generation can be formulated as a problem that maximizes these criteria together. Previous efforts in generating synthetic face datasets touch on one of the three aspects but do not consider all of them together [3, 49]. SynFace [49] generates high-fidelity face images based on DiscoFaceGAN [12], coming close to real images in terms of FID metric [18]. However, we were 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. 12715 Style Image 𝑿𝒔𝒕𝒚𝑩ID Image 𝑿𝒊𝒅𝑨Predicted"𝑿𝒔𝒕𝒚𝑨𝑬𝒔𝒕𝒚Generate ID image𝑁(0,1)𝑮𝒊𝒅 Dual Condition Generator 𝑮𝒎𝒊𝒙𝑬𝒊𝒅Style BankChoose Style image Subject : ASubject : BSubject : A𝝐𝜽1. Sampling Stage2. Mixing StageIdentity GeneratorABCLabeled DatasetFigure 2. Two stage dataset generation paradigm. In the sampling stage, 1) Gidgenerates a high-quality face image Xidthat defines how a person looks and 2) the style bank selects a style image Xstythat defines the overall style of the final image. The mixing stage generates image with identity from Xidand style from Xsty. Repeating this process multiple times, one can generate a labeled synthetic face dataset. surprised to find that the actual number of unique subjects that can be generated by DiscoFaceGAN is less than 500, a finding that will be discussed in Sec. 3.1. The recent state of the art (SoTA), DigiFace [3], can generate 1M large-scale synthetic face images with many unique subjects based on 3D parametric model rendering. However, it falls short in matching the quality and style of real face images. We propose a new data generation scheme that addresses all three criteria, i.e.the large number of novel subjects (uniqueness ), real dataset style matching ( diversity ) and la- bel consistency ( consistency ). In Fig. 1, we illustrate the high-level idea by showcasing some of our generated face samples. The key motivation of our paper is that the syn- thetic dataset generator needs to control the number of unique subjects, match the training dataset’s style distribu- tion and be consistent in the subject label. In light of this, we formulate the face image generation as a dual condition inverse problem, retrieving the unknown imageYfrom the observable Identity condition Xidand Style condition Xsty. Specifically, Xidspecifies how a per- son looks and Xstyspecifies how Xidshould be portrayed in an image. Xstycontains identity-independent informa- tion such as pose, expression, and image quality. Our choice of dual conditions (identity and style) is im- portant in how we generate a synthetic dataset as ID and style conditions are controllable factors that govern the dataset’s characteristics. To achieve this, we propose a two-stage generation paradigm. First, we generate a high- quality face image Xidusing a face image generator and sample a style image Xstyfrom a style bank. Secondly, we mix these two conditions using a dual condition generator which predicts an image that has the ID of Xidand a style ofXsty. An illustration is given in Fig. 2. Training the mixing generator in stage 2 is not trivial as it would require a triplet of ( XA id,XB sty,XA sty) where XA sty is a hypothetical combination of the ID of subject Aand the style of subject B. To solve this problem, we propose a new dual condition generator that can learn from ( XA id,XA sty), a tuple of same subject images that can always be obtained in a labeled dataset. The novelty lies in our style condi-tion extractor and ID loss which prevents the training from falling into a degenerate solution. We modify the diffusion model [19, 55] to take in dual conditions and apply an aux- iliary time-dependent ID loss that can control the balance between sample diversity and label consistency. We show that our Dual Condition Face Dataset Gener- ator (DCFace) is capable of surpassing the previous meth- ods in terms of FR performance, establishing a new bench- mark in face recognition with synthetic face datasets. We also show the roles dataset subject uniqueness, diversity and consistency play in face recognition performance. The followings are the contributions of the paper. • We propose a two-stage face dataset generator that controls subject uniqueness, diversity and consistency. • For this, we propose a dual condition generator that mixes the two independent conditions XidandXsty. • We propose uniqueness, consistency and diversity met- rics that quantify the respective properties of a given dataset, useful measures that allow one to compare datasets apart from the recognition performance. • We achieve SoTA in FR with 0.5M image synthetic training dataset by surpassing the previous methods by 6.11% on average in 5popular test datasets.
Lao_Simultaneously_Short-_and_Long-Term_Temporal_Modeling_for_Semi-Supervised_Video_Semantic_CVPR_2023	Abstract In order to tackle video semantic segmentation task at a lower cost, e.g., only one frame annotated per video, lots of efforts have been devoted to investigate the utiliza- tion of those unlabeled frames by either assigning pseudo labels or performing feature enhancement. In this work, we propose a novel feature enhancement network to si- multaneously model short- and long-term temporal corre- lation. Compared with existing work that only leverage short-term correspondence, the long-term temporal corre- lation obtained from distant frames can effectively expand the temporal perception field and provide richer contex- tual prior. More importantly, modeling adjacent and dis- tant frames together can alleviate the risk of over-fitting, hence produce high-quality feature representation for the distant unlabeled frames in training set and unseen videos in testing set. To this end, we term our method SSLTM , short for Simultaneously Short- and Long-Term Temporal Modeling. In the setting of only one frame annotated per video, SSLTM significantly outperforms the state-of-the-art methods by 2%∼3%mIoU on the challenging VSPW dataset. Furthermore, when working with a pseudo label based method such as MeanTeacher, our final model only exhibits 0.13% mIoU less than the ceiling performance ( i.e., all frames are manually annotated).	1. Introduction Deep neural networks have been the de-facto solution for many vision tasks such as image recognition [12], ob- ject detection [15] and semantic segmentation [21]. These state-of-the-art results are generally achieved by training very deep networks on large-scale labeled datasets, e.g., Im- ageNet [27], COCO [16] and Cityscapes [4], etc. However, building such labeled datasets is labor-intensive and com- plicated. Hence, it is very appealing to explore less label- dependent alternatives that only requires a (small) portion of the dataset to be annotated [14, 24–26, 37]. FirstLabeledFrameDistantUnlabeledFrameGTw/DistantFramew/oDistantFrame(a) The model trained with distant frames demonstrates better segmentation output for distant unlabeled frames in the training set. DistantFrameCurrentFrameGT w/DistantFramew/oDistantFrame (b) The adjacent frames do not contain sufficient visual clues to segment the car in current frame, while the distant frame can provide richer context to guide the segmentation model. Figure 1. The importance of involving distant frames in train- ing. The exploitation of distant frames not only reduces the risk of over-fitting to the labeled frame and its adjacent ones, but also provides temporally long-term context to enhance the feature rep- resentation. In this work, we aim to train the video semantic seg- mentation model under an extreme setting of annotation availability, i.e., each video in the training set only has its first frame annotated. The significance of this problem is twofold: 1). Video semantic segmentation is a fundamen- tal task in computer vision, with wide applications in many scenarios like autonomous driving [9], robot controlling [7]; 2). Compared with dense annotations, it takes much less (if not the least) cost to label one frame per video. More im- portantly, given the information redundancy [18, 41] within a video, it seems intuitively unnecessary to annotate ev- ery frame at all costs. Thus, it is of great practical and theoretical interests to explore the feasibility of conduct- ing video semantic segmentation with one-frame-per-video- 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. 14763 annotation. Existing methods for this problem can be grouped into Pseudo Label based approaches and Feature Enhancement based ones, depending on whether explicit pseudo labels are generated for the unlabeled frames. For the former ones [1, 6, 40], a pseudo-label generator is often trained with the annotated frames, then the model is updated using both la- beled and unlabeled data. As a comparison, the latter group of methods [18, 41] concentrates on obtaining high-quality representations based on the features from both labeled and unlabeled frames. Thus, these methods rely on feature en- hancement modules that are specially designed for temporal feature fusion. Note, Pseudo Label based approaches and Feature Enhancement based ones are orthogonal, i.e., they pay attention to different aspects of the semi-supervised video semantic segmentation task, and can usually work to- gether to combine the best of two worlds as shown in Sec- tion 4.5. In this work, we will focus on the latter ones - designing innovative feature enhancement modules. Prior arts on feature enhancement mostly focus on mod- eling short-term temporal correlation, under the assumption of temporal consistency [18, 41] among adjacent frames. Nevertheless, the distant frame is less exploited in existing work, due to its severe content changes and weak temporal consistency. However, in the setting of partial annotation, the absence of distant frames in training results in signifi- cant drawbacks: 1). As illustrated in Figure 1a, if the distant frame is not involved in the training phase, the model will be over adapted (or even over-fitted) to the labeled frame and its adjacent ones. Consequently, the generalization to dis- tant frames and unseen videos in the testing set is severely hurt, leading to poor segmentation performance in the test- ing stage. 2). Since the distant frame can provide long-term temporal context, the representation quality of the current frame can be improved by leveraging the information from its distant frame. A qualitative sample is given in Figure 1b, where a severely occluded car is correctly segmented with the help of long-term context from the distant frame. To address the aforementioned drawbacks, we propose a novel Simultaneously Short- and Long-Term Temporal Modeling (SSLTM) method to capture the temporal rela- tionship from both adjacent and distant frames. To achieve this goal, we design three novel components in our model for representation learning: 1). We refer to the labeled frame as query frame, for its adjacent frames in the same video, we model the short-term inter-frame correlations by a Spatial-Temporal Transformer (STT). 2). For the pur- pose of long-term temporal modeling, we obtain a refer- ence frame by randomly sampling a distant frame from the same video of the query frame, then feeding the reference frame’s feature to our proposed Reference Frame Context Enhancement (RFCE) module, so as to enhance the rep- resentation of query frame. Meanwhile, as the referenceframe is selected randomly from the entire video, our model is potentially trained with all data, rather than just the la- beled frames and their adjacent ones. As such, we expect the model to be prevented from over-fitting to some extent. 3). To compensate for the semantic category representation from RFCE, we further propose a Global Category Context (GCC) module to model the global information across the whole dataset. In summary, our method is a pioneer work to exploit both short- and long-term inter-frame correlations in the video semantic segmentation task. Thanks to the effective model- ing of distant frames, our RFCE demonstrates outstanding performance, especially under the setting of partial annota- tion. Specifically, on the challenging VSPW dataset [22], the mIoU of our final model only decreases by 0.13% when switching from per-frame-annotation to the one-frame-per- video-annotation setting. To our knowledge, this is the first work that nearly closes the gap between dense anno- tations and one-frame-per-video ones, which is of great sig- nificance in practical applications. Compared with exist- ing Feature Enhancement based video semantic segmenta- tion methods [5,18,22,23,30,41], our SSLTM demonstrates advantageous results (mIoU as 39.79%) by a large margin (2%∼3%mIoU) on the VSPW dataset.
Kim_Feature_Separation_and_Recalibration_for_Adversarial_Robustness_CVPR_2023	Abstract Deep neural networks are susceptible to adversarial at- tacks due to the accumulation of perturbations in the fea- ture level, and numerous works have boosted model robust- ness by deactivating the non-robust feature activations that cause model mispredictions. However, we claim that these malicious activations still contain discriminative cues and that with recalibration, they can capture additional use- ful information for correct model predictions. To this end, we propose a novel, easy-to-plugin approach named Fea- ture Separation and Recalibration (FSR) that recalibrates the malicious, non-robust activations for more robust fea- ture maps through Separation and Recalibration. The Sep- aration part disentangles the input feature map into the robust feature with activations that help the model make correct predictions and the non-robust feature with activa- tions that are responsible for model mispredictions upon adversarial attack. The Recalibration part then adjusts the non-robust activations to restore the potentially useful cues for model predictions. Extensive experiments verify the superiority of FSR compared to traditional deactivation techniques and demonstrate that it improves the robustness of existing adversarial training methods by up to 8.57% with small computational overhead. Codes are available athttps://github.com/wkim97/FSR .	1. Introduction Despite the advancements of deep neural networks (DNNs) in computer vision tasks [2,10,20,39], they are vul- nerable to adversarial examples [19,43] that are maliciously crafted to subvert the decisions of these models by adding imperceptible noise to natural images. Adversarial exam- ples are also known to be successful in real-world cases, including autonomous driving [17] and biometrics [27, 41], and to be effective even when target models are unknown to the attacker [26, 31, 43]. Thus, it has become crucial to devise effective defense strategies against this insecurity. To this end, numerous defense techniques have been pro- posed, including defensive distillation [38], input denois- ing [30], and attack detection [36, 54]. Among these meth-ods, adversarial training [19, 33], which robustifies a model by training it on a set of worst-case adversarial examples, has been considered to be the most successful and popular. Even with adversarial training, however, small adversar- ial perturbations on the pixel-level accumulate to a much larger degree in the intermediate feature space and ruin the final output of the model [50]. To solve this problem, re- cent advanced methods disentangled and deactivated the non-robust feature activations that cause model mispredic- tions. Xie et al. [50] applied classical denoising techniques to deactivate disrupted activations, and Bai et al. [4] and Yan et al. [55] deactivated channels that are irrelevant to correct model decisions. These approaches, however, in- evitably neglect discriminative cues that potentially lie in these non-robust activations. Ilyas et al. [22] have shown that a model can learn discriminative information from non- robost features in the input space. Based on this finding, we argue that there exist potential discriminative cues in the non-robust activations, and deactivating them could lead to loss of these useful information that can provide the model with better guidance for making correct predictions. For the first time, we argue that with appropriate adjust- ment, the non-robust activations that lead to model mis- predictions could recapture discriminative cues for correct model decisions. To this end, we propose a novel Feature Separation and Recalibration (FSR) module that aims to improve the feature robustness . We first separate the in- termediate feature map of a model into the malicious non- robust activations that are responsible for model mispre- dictions and the robust activations that still provide useful cues for correct model predictions even under adversarial attacks. Exploiting only the robust feature just like the ex- isting methods [4, 50, 55], however, could lead to loss of potentially useful cues in the non-robust feature. Thus, we recalibrate the non-robust activations to capture cues that provide additional useful guidance for correct model deci- sions. These additional cues can better guide the model to make correct predictions and thus boost its robustness. Fig. 1 visualizes the attention maps [40] on the features of natural images by a naturally trained model ( fnat) and the robust ( f+), non-robust ( f−), and recalibrated features (˜f−) on adversarial examples ( x′) obtained from an adver- 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. 8183 𝑥′ 𝑓+𝑓− ሚ𝑓−𝑓𝑛𝑎𝑡 Figure 1. Visualization of attention maps [40] on the features of natural images on a naturally trained model ( fnat) and the robust (f+), non-robust ( f−), and recalibrated features ( ˜f−) of adver- sarial examples ( x′) obtained from an adversarial training [33] model equipped with our FSR module on CIFAR-10 dataset. The robust feature captures discriminative cues regarding the ground truth class, while the non-robust feature captures irrelevant cues. To further boost feature robustness, we recalibrate the non-robust feature ( f−→˜f−) and capture additional useful cues for model predictions. Adversarial images x′are upscaled for clearer visual representations. sarial training [33] model equipped with our FSR module. Given an adversarial example, while the non-robust fea- ture (f−) captures cues irrelevant to the ground truth class, the robust feature ( f+) captures discriminative cues ( e.g., horse’s leg). Our FSR module recalibrates the non-robust activations ( f−→˜f−), which are otherwise neglected by the existing methods, and restores additional useful cues not captured by the robust activations ( e.g., horse’s body). With these additional cues, FSR further boosts the model’s ability to make correct decisions on adversarial examples. Thanks to its simplicity, our FSR module can be easily plugged into any layer of a CNN model and is trained with the entire model in an end-to-end manner. We extensively evaluate the robustness of our FSR module on benchmark datasets against various white-box and black-box attacks and demonstrate that our approach improves the robustness of different variants of adversarial training (Sec. 4.2) with small computational overhead (Sec. 4.4). We also show that our approach of recalibrating non-robust activations is superior to existing techniques [4, 50, 55] that simply deac- tivate them (Sec. 4.2). Finally, through ablation studies, we demonstrate that our Separation stage can effectively dis- entangle feature activations based on their effects on model decision and that our Recalibration stage successfully re- captures useful cues for model predictions (Sec. 4.3). In summary, our contributions are as follow: • In contrast to recent methods that deactivate distorted feature activations, we present a novel point of viewthat these activations can be recalibrated to capture useful cues for correct model decisions. • We introduce an easy-to-plugin Feature Separation and Recalibration (FSR) module, which separates non-robust activations from feature maps and recali- brates these feature units for additional useful cues. • Experimental results demonstrate the effectiveness of our FSR module on various white- and black-box at- tacks with small computational overhead and verify our motivation that recalibration restores discrimina- tive cues in non-robust activations.
Liu_Delving_StyleGAN_Inversion_for_Image_Editing_A_Foundation_Latent_Space_CVPR_2023	Abstract GAN inversion and editing via StyleGAN maps an in- put image into the embedding spaces ( W,W+, andF) to simultaneously maintain image fidelity and meaningful manipulation. From latent space Wto extended latent spaceW+to feature space Fin StyleGAN, the editability of GAN inversion decreases while its reconstruction quality increases. Recent GAN inversion methods typically explore W+andFrather than Wto improve reconstruction fidelity while maintaining editability. As W+andFare derived fromWthat is essentially the foundation latent space of StyleGAN, these GAN inversion methods focusing on W+ andFspaces could be improved by stepping back to W. In this work, we propose to first obtain the proper latent code in foundation latent space W. We introduce contrastive learning to align Wand the image space for proper la- tent code discovery. Then, we leverage a cross-attention en- coder to transform the obtained latent code in WintoW+ andF, accordingly. Our experiments show that our explo- *Y . Song and Q. Chen are the joint corresponding authors.ration of the foundation latent space Wimproves the repre- sentation ability of latent codes in W+and features in F, which yields state-of-the-art reconstruction fidelity and ed- itability results on the standard benchmarks. Project page: https://kumapowerliu.github.io/CLCAE .	1. Introduction StyleGAN [29–31] achieves numerous successes in im- age generation. Its semantically disentangled latent space enables attribute-based image editing where image content is modified based on the semantic attributes. GAN in- version [62] projects an input image into the latent space, which benefits a series of real image editing methods [4,36, 49, 65, 72]. The crucial part of GAN inversion is to find the inversion space to avoid distortion while enabling ed- itability. Prevalent inversion spaces include the latent space W+[1] and the feature space F[28]. W+is shown to balance distortion and editability [56, 71]. It attracts many editing methods [1, 2, 5, 20, 25, 53] to map real images into this latent space. On the other hand, Fcontains spatial 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. 10072 age representation and receives extensive studies from the image embedding [28, 48, 59, 63] or StyleGAN’s parame- ters [6, 14] perspectives. The latent space W+and feature space Freceive wide investigations. In contrast, Karras et al. [31] put into explor- ingWand the results are unsatisfying. This may be because that manipulation in Wwill easily bring content distortions during reconstruction [56], even though Wis effective for editability. Nevertheless, we observe that W+andFare indeed developed from W, which is the foundation latent space in StyleGAN. In order to improve image editability while maintaining reconstruction fidelity (i.e., W+andF), exploring Wis necessary. Our motivation is similar to the following quotation: “You can’t build a great building on a weak foundation. You must have a solid foundation if you’re going to have a strong superstructure. ” —Gordon B. Hinckley In this paper, we propose a two-step design to improve the representation ability of the latent code in W+andF. First, we obtain the proper latent code in W. Then, we use the latent code in Wto guide the latent code in W+andF. In the first step, we propose a contrastive learning paradigm to align the Wand image space. This paradigm is derived from CLIP [51] where we switch the text branch with W. Specifically, we construct the paired data that consists of one image Iand its latent code w∈ W with pre-trained StyleGAN. During contrastive learning, we train two en- coders to obtain two feature representations of Iandw, re- spectively. These two features are aligned after the train- ing process. During GAN inversion, we fix this contrastive learning module and regard it as a loss function. This loss function is set to make the one real image and its latent code wsufficiently close. This design improves existing stud- ies [31] on Wthat their loss functions are set on the image space (i.e., similarity measurement between an input image and its reconstructed image) rather than the unified image and latent space. The supervision on the image space only does not enforce well alignment between the input image and its latent code in W. After discovering the proper latent code in W, we lever- age a cross-attention encoder to transform wintow+∈ W+andf∈ F . When computing w+, we set was the query and w+as the value and key. Then, we calculate the cross-attention map to reconstruct w+. This cross-attention map enforces the value w+close to the query w, which en- ables the editability of w+to be similar to that of w. Be- sides, w+is effective in preserving the reconstruction abil- ity. When computing f, we set the was the value and key, while setting fas the query. So wwill guide ffor fea- ture refinement. Finally, we use w+andfin StyleGAN to generate the reconstruction result. We named our method CLCAE (i.e., StyleGAN in-version with Contrastive Learning and Cross-Attention Encoder). We show that our CLCAE can achieve state- of-the-art performance in both reconstruction quality and editing capacity on benchmark datasets containing human portraits and cars. Fig. 1 shows some results. This indi- cates the robustness of our CLCAE. Our contributions are summarized as follows: • We propose a novel contrastive learning approach to align the image space and foundation latent space W of StyleGAN. This alignment ensures that we can ob- tain proper latent code wduring GAN inversion. • We propose a cross-attention encoder to transform la- tent codes in WintoW+andF. The representation of latent code in W+and feature in Fare improved to benefit reconstruction fidelity and editability. • Experiments indicate that our CLCAE achieves state- of-the-art fidelity and editability results both qualita- tively and quantitatively.
Lin_Zero-Shot_Everything_Sketch-Based_Image_Retrieval_and_in_Explainable_Style_CVPR_2023	Abstract This paper studies the problem of zero-short sketch- based image retrieval (ZS-SBIR), however with two sig- nificant differentiators to prior art (i) we tackle all vari- ants (inter-category, intra-category, and cross datasets) of ZS-SBIR with just one network (“everything”), and (ii) we would really like to understand how this sketch-photo matching operates (“explainable”). Our key innovation lies with the realization that such a cross-modal matching prob- lem could be reduced to comparisons of groups of key local patches – akin to the seasoned “bag-of-words” paradigm. Just with this change, we are able to achieve both of the aforementioned goals, with the added benefit of no longer requiring external semantic knowledge. Technically, ours is a transformer-based cross-modal network, with three novel components (i) a self-attention module with a learnable tok- enizer to produce visual tokens that correspond to the most informative local regions, (ii) a cross-attention module to compute local correspondences between the visual tokens across two modalities, and finally (iii) a kernel-based rela- tion network to assemble local putative matches and pro- duce an overall similarity metric for a sketch-photo pair. Experiments show ours indeed delivers superior perfor- mances across all ZS-SBIR settings. The all important ex- plainable goal is elegantly achieved by visualizing cross- modal token correspondences, and for the first time, via sketch to photo synthesis by universal replacement of all matched photo patches. Code and model are available at https://github.com/buptLinfy/ZSE-SBIR .	1. Introduction Zero-shot sketch-based image retrieval (ZS-SBIR) is a central problem to sketch understanding [8, 16, 19, 20, 28, 34, 50, 54, 55, 57, 60, 67]. The zero-shot setting is largely driven by the prevailing data scarcity problem of human sketches [19,28,58] – they are much harder to acquire com- pared with photos. As research matures on the non zero- ∗Equal contribution. †Corresponding author. (a) (b) Unseen categoryTraining data Transfer Shareable Local Match (c) Figure 1. Attentive regions of self-/cross-attention and the learned visual correspondence for tackling unseen cases. (a) The proposed retrieval token [ Ret] can attend to informative regions. Different colors are attention maps from different heads. (b) Cross-attention offers explainability by explicitly constructing local visual cor- respondence. The local matches learned from training data are shareable knowledge, which enables ZS-SBIR to work under di- verse settings (inter- / intra-category and cross datasets) with just one model. (c) An input sketch can be transformed into its image by the learned correspondence, i.e., sketch patches are replaced by the closest image patches from the retrieved image. shot [4,25,33,41–43,59,62], and in a push to make sketch- based retrieval commercially viable, recent research efforts had mainly focused on ZS-SBIR (or the simpler few-shot setting) [16, 19, 28, 34, 50, 57, 60]. Great strides have been made but attempts have largely aligned with the larger photo-based zero-shot literature, where the key lies in leveraging external knowledge for cross-category adaptation [19, 34]. That of conducting cross-modal matching is, however, less studied, and most prior art relies on a gold standard triplet loss with some auxiliary modules [16] to learn a joint embedding. Further- more, as problems such as domain shift and fine-grained matching come to play, research efforts are mostly done in silo for different settings: category-level (standard) [28, 50, 60], fine-grained [2], and cross-dataset [40]. Last but defi- nitely not least, one can not help but wonder why many of the proposed algorithm work – what is matched, and how is the transfer conducted? This paper aims to tackle all said problems associated with the current status quo for ZS-SBIR. In particular, we advocate for (i) a single model to tackle all three settings 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. 23349 ZS-SBIR, (ii) ditching the requirement on external knowl- edge to conduct category transfer, and more importantly, (iii) a way to explain why our model works (or not). At the very core of our contribution lies with a well- explored insight that predates “deep vision”, that image matching can be achieved by establishing local patch cor- respondences and computing a distance score based on that – yes, loosely similar to that of “bag of visual words” [11, 26, 51] (without building dictionaries). In our context of ZS-SBIR, we would like to conduct this matching (i) cross two diverse modalities in sketch and photo, and (ii) cross- category, granularity (fine-grained or not) and dataset (do- main difference) boundaries. The biggest upside of this re- alization is that just as how “bag of visual words” is explain- able, we can directly visualize the patch correspondences to achieve a similar level of explainability (see Figure 1). Our solution first is a transformer-based cross-modal net- work, that (i) sources local patches independently in each modality, (ii) establishes patch-to-patch correspondences across two modalities, and (iii) computes matching scores based on putative correspondences. We put forward a novel design for each of the three components. We approach (i) by proposing a novel CNN-based learnable tokenizer, that is specifically tailored to sketch data. This is because the vanilla non-overlapping patch-wise tokenization proposed in ViT [18] is not friendly to the sparse nature of sketches (as most patches would belong to the uninformative blank). Our tokenizer on the other hand attends to a larger recep- tive field [37] hence more keen to sketch data. With this tokenizer, visual cues from nearby regions are aggregated when constructing visual tokens, so that structural informa- tion is preserved. In the same spirit of class token developed in ViT for image recognition, we introduce a learnable re- trieval token to prioritize tokens for cross-modal matching. To establish patch-to-patch correspondences, a novel cross-attention module is proposed that operates across sketch-photo modalities. Specifically, we propose cross- modal multi-head attention, in which the query embeddings are exchanged between sketch and photo branches to rea- son patch-level correspondences with only category-level supervision. With the putative matches in place, inspired by relation networks [53], we propose a kernel-based relation network to aggregate the correspondences and calculate a similarity score between each sketch-photo pair. We achieve state-of-the-art performance across all said ZS-SBIR settings. Explainability is offered (i) as per tradi- tion in terms of visualizing patch correspondences, where interesting local matches can be observed, such as the antlers ofdeer in Figure 1(b), regardless of a sketch being very abstract, and (ii) by replacing all patches in a sketch with their photo correspondences, to perform sketch to photo synthesis as shown in Figure 1(c).
Liu_Hierarchical_Supervision_and_Shuffle_Data_Augmentation_for_3D_Semi-Supervised_Object_CVPR_2023	Abstract State-of-the-art 3D object detectors are usually trained on large-scale datasets with high-quality 3D annotations. However , such 3D annotations are often expensive and time-consuming, which may not be practical for real ap- plications. A natural remedy is to adopt semi-supervised learning (SSL) by leveraging a limited amount of labeled samples and abundant unlabeled samples. Current pseudo-labeling-based SSL object detection methods mainly adopt a teacher-student framework, with a single ﬁxed thresholdstrategy to generate supervision signals, which inevitably brings confused supervision when guiding the student net-work training. Besides, the data augmentation of the point cloud in the typical teacher-student framework is too weak, and only contains basic down sampling and ﬂip-and-shift(i.e., rotate and scaling), which hinders the effective learn- ing of feature information. Hence, we address these is- sues by introducing a novel approach of Hierarchical Su-pervision and Shufﬂe Data Augmentation (HSSDA), whichis a simple yet effective teacher-student framework. The teacher network generates more reasonable supervision forthe student network by designing a dynamic dual-threshold strategy. Besides, the shufﬂe data augmentation strategy is designed to strengthen the feature representation ability of the student network. Extensive experiments show thatHSSDA consistently outperforms the recent state-of-the-art methods on different datasets. The code will be released athttps://github.com/azhuantou/HSSDA .	1. Introduction Kinds of important applications, especially autonomous driving, have been motivating the rapid development of 3D *Corresponding author.Vanilla TeacherTeacher NetworkUnlabeled scene Pseudo-labeling scene Ours Teacher Vanilla StudentStudent NetworkUnlabeled scene Augmented scene Student NetworkUnlabeled scene Augmented scene RGB Image Full labeled scene (a) (b)Ours Student High threshold Low threshold Teacher NetworkUnlabeled scene Pseudo-labeling sceneDual-threshold strategyFlip&Shift Split&Shuffle OR Figure 1. Illustration of (a) the previous teacher compared to our teacher framework and (b) the previous student compared to ourstudent framework. The black dashed box includes the RGB im- age and the corresponding fully annotated 3D point cloud (greenbox). The left side of the yellow dotted line in (a) representsthe pseudo-labeling scene generated by the single threshold ofthe vanilla teacher network, causing the student network may beseverely misled due to missing mined objects (high threshold) orfalse positive objects (low threshold), while our proposed teachernetwork generates three groups of pseudo labels(shown as green,red, blue) to provide hierarchical supervision for the student net- work. (b) shows our student network adopts stronger shufﬂed data augmentation than the vanilla student network to learn the strongerability of feature representation. object detection by the range sensor data ( e.g., LiDAR point cloud). Up to now, many point-based and point-voxel-based methods [ 10,30,31,49] have been proposed. Despite the im- pressive progress, a large amount of accurate instance-level 3D annotations have to be provided for training, which is more time-consuming and expensive than 2D object annota-tion. This hinders the application and deployment of exist- 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. 23819 ing advanced detection models. To this end, how to reduce the dependence on huge annotated datasets has achieved growing interest in the object detection community. As one of the important machine learning schemes for reducing data annotation, semi-supervised learning (SSL)aims to improve the generalization ability of model train-ing with a small number of labeled data together withlarge-scale unlabeled samples. In the semi-supervised ob- ject detection community, most works focus on 2D object detection [ 13,17,34,40,43], among which the teacher- student model is the mainstream framework. Speciﬁcally,the teacher network with weakly augmented labeled datagenerates pseudo labels to train the student network withstrong data augmentation. This pipeline has been widelyveriﬁed to be effective, in which the quality of pseudo labels and the data augmentation strategy are the key and manyworks have been proposed to tackle them [ 2,6,33,34]. Ben- eﬁting from 2D semi-supervised object detection, several3D semi-supervised object detection methods have been proposed [ 23,39,47,52], which still mainly adopted the teacher-student model. These methods, as well as 2D semi- supervised object detection methods [ 21,34,55], mainly use a hard way, e.g., a score threshold, to get pseudo labels to train the student network. This kind of strategy is difﬁcult to guarantee the quality of pseudo labels. Taking the scorethreshold strategy as an example, if the threshold is too low, the pseudo labels will contain many false objects, while ifit is too high, the pseudo labels will miss many real ob- jects which will be improperly used as background (see in Fig. 1(a)). As exhibited in [ 39], only about 30% of the ob- jects can be mined from the unlabeled scenes even at theend of the network training. Thus, both of those two cases will bring the student network confused supervision, whichharms the performance of the teacher-student model. This would inevitably happen for the single threshold strategy,even adopting some optimal threshold search method [ 39]. Thus, how to produce reasonable pseudo labels from the teacher network output is an important issue to address for better training the student networks. Besides the quality of pseudo labels, data augmentation is also the key to the teacher-student model as mentionedpreviously. Extensive works in 2D semi-supervised object detection have shown that strong data augmentation is very important to learn the strong feature representation ability of the student network. Thus, kinds of strong data aug- mentation strategies, e.g., Mixup [ 48], Cutout [ 6], and Mo- saic [ 4] have been widely adopted. However, current 3D semi-supervised object detection methods adopt some weakdata augmentation strategies, e.g., ﬂip-and-shift. These kinds of data augmentations are not able to well drive the student network to learn strong feature representation abil- ity. Thus, the good effect of data augmentation in 2D semi-supervised object detection does not appear obviously in 3Dsemi-supervised object detection. To tackle the above issues of the quality of pseudo labels and data augmentation, we propose a Hierarchical Super-vision and Shufﬂe Data Augmentation (HSSDA) methodfor 3D semi-supervised object detection. We still adoptthe teacher-student model as our mainframe. For obtain-ing more reasonable pseudo labels for the student network,we design a dynamic dual-threshold strategy to divide the output of the teacher network into three groups: (1) high- conﬁdence level pseudo labels, (2) ambiguous level pseudolabels, and (3) low-conﬁdence level pseudo labels, as shown in Fig. 1(a). This division provides hierarchical supervision signals for the student network. Speciﬁcally, the ﬁrst group is used as the strong labels to learn the student network,while the second join learning through a soft-weight way. The higher the score is, the more it affects learning. The third group is much more likely to tend to be false objects.We directly delete them from the point cloud to avoid con- fusing parts of the object point cloud into the background. For strengthening the feature representation ability of the student network, we design a shufﬂe data augmentation strategy in this paper. As shown in Fig. 1(b), we ﬁrst gener- ate shufﬂed scenes by splitting and shufﬂing the point cloud patches in BEV (bird-eye view) and use them as inputs tothe student model. Next, the feature maps extracted from the detector backbone are unshufﬂed back to the original point cloud geometry location. To summarize, our contributions are as follows: • We propose a novel hierarchical supervision gener- ation and learning strategy for the teacher-student model. This strategy can provide the student networkhierarchical supervision signal, which can fully utilizethe output of the teacher network. • We propose a shufﬂe data augmentation strategy that can strengthen the feature representation ability of the student network. • Our proposed hierarchical supervision strategy and shufﬂe data augmentation strategy can be directly ap-plied to the off-the-shelf 3D semi-supervised point cloud object detector and extensive experimentsdemonstrate that our method has achieved state-of-the- art results.

Subsets and Splits

No community queries yet

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