CVPR 2021 Papers

Regression tracking has gained more and more attention thanks to its easy-to-implement characteristics, while existing regression trackers rarely consider the relationships between the object parts and the complete object. This would ultimately result in drift from the target object when missing some parts of the target object. Recently, Capsule Network (CapsNet) has shown promising results for image classification benefits from its part-object relationships mechanism, while CapsNet is known for its high computational demand even when carrying out simple tasks. Therefore, a primitive adaptation of CapsNet to regression tracking does not make sense, since this will seriously affect speed of a tracker. To solve these problems, we first explore the spatial-temporal relationships endowed by the CapsNet for regression tracking. The entire regression framework, dubbed CapsuleRRT, consists of three parts. One is S-Caps, which captures the spatial relationships between the parts and the object. Meanwhile, a T-Caps module is designed to exploit the temporal relationships within the target. The response of the target is obtained by STCaps Learning. Further, a prior-guided capsule routing algorithm is proposed to generate more accurate capsule assignments for subsequent frames. Apart from this, the heavy computation burden in CapsNet is addressed with a knowledge distillation pose matrix compression strategy that exploits more tight and discriminative representation with few samples. Extensive experimental results show that CapsuleRRT performs favorably against state-of-the-art methods in terms of accuracy and speed.

The Capsule Network is widely believed to be more robust than Convolutional Networks. However, there lack comprehensive comparisons between these two networks, and it is also unknown which components in the CapsNet affect its robustness. In this paper, we first carefully examine the special designs in CapsNet differing from that of a ConvNet, commonly used for image classification. The examination reveals 5 major new/different components in CapsNet: a transformation process, a dynamic routing layer, a squashing function, a marginal loss other than cross-entropy loss, and an additional class-conditional reconstruction loss for regularization. Along with these major differences, we comprehensively ablate their behavior on 3 kinds of robustness, including affine transformation, overlapping digits, and semantic representation. The study reveals that some designs which are thought critical to CapsNet actually can harm its robustness, i.e., the dynamic routing layer and the transformation process, while others are beneficial for the robustness. Based on these findings, we propose enhanced ConvNets simply by introducing the essential components behind the CapsNet's success. The proposed simple ConvNets can achieve better robustness than the CapsNet.

Human pose estimation from single images is a challenging problem in computer vision that requires large amounts of labeled training data to be solved accurately. Unfortunately, for many human activities (e.g. outdoor sports) such training data does not exist and is hard or even impossible to acquire with traditional motion capture systems. We propose a self-supervised approach that learns a single image 3D pose estimator from unlabeled multi-view data. To this end, we exploit multi-view consistency constraints to disentangle the observed 2D pose into the underlying 3D pose and camera rotation. In contrast to most existing methods, we do not require calibrated cameras and can therefore learn from moving cameras. Nevertheless, in the case of a static camera setup, we present an optional extension to include constant relative camera rotations over multiple views into our framework. Key to the success are new, unbiased reconstruction objectives that mix information across views and training samples. The proposed approach is evaluated on two benchmark datasets (Human3.6M and MPII-INF-3DHP) and on the in-the-wild SkiPose dataset.

The problem of expert model selection deals with choosing the appropriate pretrained network ("expert") to transfer to a target task. Methods, however, generally depend on two separate assumptions: the presence of labeled images and access to powerful "probe" networks that yield useful features. In this work, we demonstrate the current reliance on both of these aspects and develop algorithms to operate when either of these assumptions fail. In the unlabeled case, we show that pseudolabels from the probe network provide discriminative enough gradients to perform nearly-equal task selection even when the probe network is trained on imagery unrelated to the tasks. To compute the embedding with no probe network at all, we introduce the Task Tangent Kernel (TTK) which uses a kernelized distance across multiple random networks to achieve performance over double that of other methods with randomly initialized models. Code is available at https://github.com/BramSW/task_characterization_cvpr_2021/.

In this paper, we propose to make a systematic study on machines' multisensory perception under attacks. We use the audio-visual event recognition task against multimodal adversarial attacks as a proxy to investigate the robustness of audio-visual learning. We attack audio, visual, and both modalities to explore whether audio-visual integration still strengthens perception and how different fusion mechanisms affect the robustness of audio-visual models. For interpreting the multimodal interactions under attacks, we learn a weakly-supervised sound source visual localization model to localize sounding regions in videos. To mitigate multimodal attacks, we propose an audio-visual defense approach based on an audio-visual dissimilarity constraint and external feature memory banks. Extensive experiments demonstrate that audio-visual models are susceptible to multimodal adversarial attacks; audio-visual integration could decrease the model robustness rather than strengthen under multimodal attacks; even a weakly-supervised sound source visual localization model can be successfully fooled; our defense method can improve the invulnerability of audio-visual networks without significantly sacrificing clean model performance.

Camouflaged object segmentation (COS) aims to identify objects that are "perfectly" assimilate into their surroundings, which has a wide range of valuable applications. The key challenge of COS is that there exist high intrinsic similarities between the candidate objects and noise background. In this paper, we strive to embrace challenges towards effective and efficient COS. To this end, we develop a bio-inspired framework, termed Positioning and Focus Network (PFNet), which mimics the process of predation in nature. Specifically, our PFNet contains two key modules, i.e., the positioning module (PM) and the focus module (FM). The PM is designed to mimic the detection process in predation for positioning the potential target objects from a global perspective and the FM is then used to perform the identification process in predation for progressively refining the coarse prediction via focusing on the ambiguous regions. Notably, in the FM, we develop a novel distraction mining strategy for the distraction region discovery and removal, to benefit the performance of estimation. Extensive experiments demonstrate that our PFNet runs in real-time (72 FPS) and significantly outperforms 18 cutting-edge models on three challenging benchmark datasets under four standard metrics.

Recent years have witnessed significant progress in 3D hand mesh recovery. Nevertheless, because of the intrinsic 2D-to-3D ambiguity, recovering camera-space 3D information from a single RGB image remains challenging. To tackle this problem, we divide camera-space mesh recovery into two sub-tasks, i.e., root-relative mesh recovery and root recovery. First, joint landmarks and silhouette are extracted from a single input image to provide 2D cues for the 3D tasks. In the root-relative mesh recovery task, we exploit semantic relations among joints to generate a 3D mesh from the extracted 2D cues. Such generated 3D mesh coordinates are expressed relative to a root position, i.e., wrist of the hand. In the root recovery task, the root position is registered to the camera space by aligning the generated 3D mesh back to 2D cues, thereby completing camera-space 3D mesh recovery. Our pipeline is novel in that (1) it explicitly makes use of known semantic relations among joints and (2) it exploits 1D projections of the silhouette and mesh to achieve robust registration. Extensive experiments on popular datasets such as FreiHAND, RHD, and Human3.6M demonstrate that our approach achieves state-of-the-art performance on both root-relative mesh recovery and root recovery. Our code is publicly available at https://github.com/SeanChenxy/HandMesh.

Monocular depth predictors are typically trained on large-scale training sets which are naturally biased w.r.t the distribution of camera poses. As a result, trained predictors fail to make reliable depth predictions for testing examples captured under uncommon camera poses. To address this issue, we propose two novel techniques that exploit the camera pose during training and prediction. First, we introduce a simple perspective-aware data augmentation that synthesizes new training examples with more diverse views by perturbing the existing ones in a geometrically consistent manner. Second, we propose a conditional model that exploits the per-image camera pose as prior knowledge by encoding it as a part of the input. We show that jointly applying the two methods improves depth prediction on images captured under uncommon and even never-before-seen camera poses. We show that our methods improve performance when applied to a range of different predictor architectures. Lastly, we show that explicitly encoding the camera pose distribution improves the generalization performance of a synthetically trained depth predictor when evaluated on real images.

Complex backgrounds and similar appearances between objects and their surroundings are generally recognized as challenging scenarios in Salient Object Detection (SOD). This naturally leads to the incorporation of depth information in addition to the conventional RGB image as input, known as RGB-D SOD or depth-aware SOD. Meanwhile, this emerging line of research has been considerably hindered by the noise and ambiguity that prevail in raw depth images. To address the aforementioned issues, we propose a Depth Calibration and Fusion (DCF) framework that contains two novel components: 1) a learning strategy to calibrate the latent bias in the original depth maps towards boosting the SOD performance; 2) a simple yet effective cross reference module to fuse features from both RGB and depth modalities. Extensive empirical experiments demonstrate that the proposed approach achieves superior performance against 27 state-of-the-art methods. Moreover, the proposed depth calibration strategy as a preprocessing step, can be further applied to existing cutting-edge RGB-D SOD models and noticeable improvements are achieved.

Garment transfer shows great potential in realistic applications with the goal of transfering outfits across different people images. However, garment transfer between images with heavy misalignments or severe occlusions still remains as a challenge. In this work, we propose Complementary Transfering Network (CT-Net) to adaptively model different levels of geometric changes and transfer outfits between different people. In specific, CT-Net consists of three modules: i) A complementary warping module first estimates two complementary warpings to transfer the desired clothes in different granularities. ii) A layout prediction module is proposed to predict the target layout, which guides the preservation or generation of the body parts in the synthesized images. iii) A dynamic fusion module adaptively combines the advantages of the complementary warpings to render the garment transfer results. Extensive experiments conducted on DeepFashion dataset demonstrate that our network synthesizes high-quality garment transfer images and significantly outperforms the state-of-art methods both qualitatively and quantitatively. Our source code will be available online.

Semi-supervised learning on class-imbalanced data, although a realistic problem, has been under studied. While existing semi-supervised learning (SSL) methods are known to perform poorly on minority classes, we find that they still generate high precision pseudo-labels on minority classes. By exploiting this property, in this work, we propose Class-Rebalancing Self-Training (CReST), a simple yet effective framework to improve existing SSL methods on class-imbalanced data. CReST iteratively retrains a baseline SSL model with a labeled set expanded by adding pseudo-labeled samples from an unlabeled set, where pseudo-labeled samples from minority classes are selected more frequently according to an estimated class distribution. We also propose a progressive distribution alignment to adaptively adjust the rebalancing strength dubbed CReST+. We show that CReST and CReST+ improve state-of-the-art SSL algorithms on various class-imbalanced datasets and consistently outperform other popular rebalancing methods. Code has been made available at https://github.com/google-research/crest.

Face detection is a fundamental problem for many downstream face applications, and there is a rising demand for faster, more accurate yet support for higher resolution face detectors. Recent smartphones can record a video in 8K resolution, but many of the existing face detectors still fail due to the anchor size and training data. We analyze the failure cases and observe a large number of correct predicted boxes with incorrect confidences. To calibrate these confidences, we propose a confidence ranking network with a pairwise ranking loss to re-rank the predicted confidences locally within the same image. Our confidence ranker is model-agnostic, so we can augment the data by choosing the pairs from multiple face detectors during the training, and generalize to a wide range of face detectors during the testing. On WiderFace, we achieve the highest AP on the single-scale, and our AP is competitive with the previous multi-scale methods while being significantly faster. On 8K resolution, our method solves the GPU memory issue and allows us to indirectly train on 8K. We collect 8K resolution test set to show the improvement, and we will release our test set as a new benchmark for future research.

In this paper, we study two challenging problems in incomplete multi-view clustering analysis, namely, i) how to learn an informative and consistent representation among different views without the help of labels and ii) how to recover the missing views from data. To this end, we propose a novel objective that incorporates representation learning and data recovery into a unified framework from the view of information theory. To be specific, the informative and consistent representation is learned by maximizing the mutual information across different views through contrastive learning, and the missing views are recovered by minimizing the conditional entropy of different views through dual prediction. To the best of our knowledge, this could be the first work to provide a theoretical framework that unifies the consistent representation learning and cross-view data recovery. Extensive experimental results show the proposed method remarkably outperforms 10 competitive multi-view clustering methods on four challenging datasets. The code is available at https://pengxi.me.

In this paper, we present CLCC, a novel contrastive learning framework for color constancy. Contrastive learning has been applied for learning high-quality visual representations for image classification. One key aspect to yield useful representations for image classification is to design illuminant invariant augmentations. However, the illuminant invariant assumption conflicts with the nature of the color constancy task, which aims to estimate the illuminant given a raw image. Therefore, we construct effective contrastive pairs for learning better illuminant-dependent features via a novel raw-domain color augmentation. On the NUS-8 dataset, our method provides 17.5% relative improvements over a strong baseline, reaching state-of-the-art performance without increasing model complexity. Furthermore, our method achieves competitive performance on the Gehler dataset with 3x fewer parameters compared to top-ranking deep learning methods. More importantly, we show that our model is more robust to different scenes under close proximity of illuminants, significantly reducing 28.7% worst-case error in data-sparse regions. Our code is available at https://github.com/howardyclo/clcc-cvpr21.

Previous point cloud semantic segmentation networks use the same process to aggregate features from neighbors of the same category and different categories. However, the joint area between two objects usually only occupies a small percentage in the whole scene. Thus the networks are well-trained for aggregating features from the same category point while not fully trained on aggregating points of different categories. To address this issue, this paper proposes to utilize different aggregation strategies between the same category and different categories. Specifically, it presents a customized module, termed as Category Guided Aggregation (CGA), where it first identifies whether the neighbors belong to the same category with the center point or not, and then handles the two types of neighbors with two carefully-designed modules. Our CGA presents a general network module and could be leveraged in any existing semantic segmentation network. Experiments on three different backbones demonstrate the effectiveness of our method.

Recently, the ever-increasing capacity of large-scale annotated datasets has led to profound progress in stereo matching. However, most of these successes are limited to a specific dataset and cannot generalize well to other datasets. The main difficulties lie in the large domain differences and unbalanced disparity distribution across a variety of datasets, which greatly limit the real-world applicability of current deep stereo matching models. In this paper, we propose CFNet, a Cascade and Fused cost volume based network to improve the robustness of the stereo matching network. First, we propose a fused cost volume representation to deal with the large domain difference. By fusing multiple low-resolution dense cost volumes to enlarge the receptive field, we can extract robust structural representations for initial disparity estimation. Second, we propose a cascade cost volume representation to alleviate the unbalanced disparity distribution. Specifically, we employ a variance-based uncertainty estimation to adaptively adjust the next stage disparity search space, in this way driving the network progressively prune out the space of unlikely correspondences. By iteratively narrowing down the disparity search space and improving the cost volume resolution, the disparity estimation is gradually refined in a coarse-to-fine manner. When trained on the same training images and evaluated on KITTI, ETH3D, and Middlebury datasets with the fixed model parameters and hyperparameters, our proposed method achieves the state-of-the-art overall performance and obtains the 1st place on the stereo task of Robust Vision Challenge 2020. The code will be available at https://github.com/gallenszl/CFNet.

DNN-based frame interpolation--that generates the intermediate frames given two consecutive frames--typically relies on heavy model architectures with a huge number of features, preventing them from being deployed on systems with limited resources, e.g., mobile devices. We propose a compression-driven network design for frame interpolation (CDFI), that leverages model pruning through sparsity-inducing optimization to significantly reduce the model size while achieving superior performance. Concretely, we first compress the recently proposed AdaCoF model and show that a 10X compressed AdaCoF performs similarly as its original counterpart; then we further improve this compressed model by introducing a multi-resolution warping module, which boosts visual consistencies with multi-level details. As a consequence, we achieve a significant performance gain with only a quarter in size compared with the original AdaCoF. Moreover, our model performs favorably against other state-of-the-arts in a broad range of datasets. Finally, the proposed compression-driven framework is generic and can be easily transferred to other DNN-based frame interpolation algorithm. Our source code is available at https://github.com/tding1/CDFI.

Recent advances in self-supervised learning (SSL) have largely closed the gap with supervised ImageNet pretraining. Despite their success these methods have been primarily applied to unlabeled ImageNet images, and show marginal gains when trained on larger sets of uncurated images. We hypothesize that current SSL methods perform best on iconic images, and struggle on complex scene images with many objects. Analyzing contrastive SSL methods shows that they have poor visual grounding and receive poor supervisory signal when trained on scene images. We propose Contrast Attention-Supervised Tuning (CAST) to overcome these limitations. CAST uses unsupervised saliency maps to intelligently sample crops, and to provide grounding supervision via a Grad-CAM attention loss. Experiments on COCO show that CAST significantly improves the features learned by SSL methods on scene images, and further experiments show that CAST-trained models are more robust to changes in backgrounds. We hope that CAST can improve the ability of SSL methods to learn from complex non-iconic images. Our code is available at https://github.com/salesforce/CAST.

Backpropagation image saliency aims at explaining model predictions by estimating model-centric importance of individual pixels in the input. However, class-insensitivity of the earlier layers in a network only allows saliency computation with low resolution activation maps of the deeper layers, resulting in compromised image saliency. Remedifying this can lead to sanity failures. We propose CAMERAS, a technique to compute high-fidelity backpropagation saliency maps without requiring any external priors and preserving the map sanity. Our method systematically performs multi-scale accumulation and fusion of the activation maps and backpropagated gradients to compute precise saliency maps. From accurate image saliency to articulation of relative importance of input features for different models, and precise discrimination between model perception of visually similar objects, our high-resolution mapping offers multiple novel insights into the black-box deep visual models, which are presented in the paper. We also demonstrate the utility of our saliency maps in adversarial setup by drastically reducing the norm of attack signals by focusing them on the precise regions identified by our maps. Our method also inspires new evaluation metrics and a sanity check for this developing research direction.