CVPR 2020 论文列表

We consider the problem of unsupervised domain adaptation for semantic segmentation by easing the domain shift between the source domain (synthetic data) and the target domain (real data) in this work. State-of-the-art approaches prove that performing semantic-level alignment is helpful in tackling the domain shift issue. Based on the observation that stuff categories usually share similar appearances across images of different domains while things (i.e. object instances) have much larger differences, we propose to improve the semantic-level alignment with different strategies for stuff regions and for things: 1) for the stuff categories, we generate feature representation for each class and conduct the alignment operation from the target domain to the source domain; 2) for the thing categories, we generate feature representation for each individual instance and encourage the instance in the target domain to align with the most similar one in the source domain. In this way, the individual differences within thing categories will also be considered to alleviate over-alignment. In addition to our proposed method, we further reveal the reason why the current adversarial loss is often unstable in minimizing the distribution discrepancy and show that our method can help ease this issue by minimizing the most similar stuff and instance features between the source and the target domains. We conduct extensive experiments in two unsupervised domain adaptation tasks, i.e. GTA5 - Cityscapes and SYNTHIA - Cityscapes, and achieve the new state-of-the-art segmentation accuracy.

Learning-based 3D reconstruction methods have shown impressive results. However, most methods require 3D supervision which is often hard to obtain for real-world datasets. Recently, several works have proposed differentiable rendering techniques to train reconstruction models from RGB images. Unfortunately, these approaches are currently restricted to voxel- and mesh-based representations, suffering from discretization or low resolution. In this work, we propose a differentiable rendering formulation for implicit shape and texture representations. Implicit representations have recently gained popularity as they represent shape and texture continuously. Our key insight is that depth gradients can be derived analytically using the concept of implicit differentiation. This allows us to learn implicit shape and texture representations directly from RGB images. We experimentally show that our single-view reconstructions rival those learned with full 3D supervision. Moreover, we find that our method can be used for multi-view 3D reconstruction, directly resulting in watertight meshes.

This paper presents a novel attention-based algorithm for achieving adaptive computation called DACT, which, unlike existing ones, is end-to-end differentiable. Our method can be used in conjunction with many networks; in particular, we study its application to the widely know MAC architecture, obtaining a significant reduction in the number of recurrent steps needed to achieve similar accuracies, therefore improving its performance to computation ratio. Furthermore, we show that by increasing the maximum number of steps used, we surpass the accuracy of even our best non-adaptive MAC in the CLEVR dataset, demonstrating that our approach is able to control the number of steps without significant loss of performance. Additional advantages provided by our approach include considerably improving interpretability by discarding useless steps and providing more insights into the underlying reasoning process. Finally, we present adaptive computation as an equivalent to an ensemble of models, similar to a mixture of expert formulation. Both the code and the configuration files for our experiments are made available to support further research in this area.

Graph matching refers to finding vertex correspondence for a pair of graphs, which plays a fundamental role in many vision and learning related tasks. Directly applying gradient-based continuous optimization on graph matching can be attractive for its simplicity but calls for effective ways of converting the continuous solution to the discrete one under the matching constraint. In this paper, we show a novel regularization technique with the tool of determinant analysis on the matching matrix which is relaxed into continuous domain with gradient based optimization. Meanwhile we present a theoretical study on the property of our relaxation technique. Our paper strikes an attempt to understand the geometric properties of different regularization techniques and the gradient behavior during the optimization. We show that the proposed regularization is more gradient-efficient than traditional ones during early update stages. The analysis will also bring about insights for other problems under bijection constraints. The algorithm procedure is simple and empirical results on public benchmark show its effectiveness on both synthetic and real-world data.

We propose a simple yet effective proposal-based object detector, aiming at detecting highly-overlapped instances in crowded scenes. The key of our approach is to let each proposal predict a set of correlated instances rather than a single one in previous proposal-based frameworks. Equipped with new techniques such as EMD Loss and Set NMS, our detector can effectively handle the difficulty of detecting highly overlapped objects. On a FPN-Res50 baseline, our detector can obtain 4.9% AP gains on challenging CrowdHuman dataset and 1.0% \text MR ^ -2 improvements on CityPersons dataset, without bells and whistles. Moreover, on less crowed datasets like COCO, our approach can still achieve moderate improvement, suggesting the proposed method is robust to crowdedness.

We address the problem of detecting attention targets in video. Our goal is to identify where each person in each frame of a video is looking, and correctly handle the case where the gaze target is out-of-frame. Our novel architecture models the dynamic interaction between the scene and head features and infers time-varying attention targets. We introduce a new annotated dataset, VideoAttentionTarget, containing complex and dynamic patterns of real-world gaze behavior. Our experiments show that our model can effectively infer dynamic attention in videos. In addition, we apply our predicted attention maps to two social gaze behavior recognition tasks, and show that the resulting classifiers significantly outperform existing methods. We achieve state-of-the-art performance on three datasets: GazeFollow (static images), VideoAttentionTarget (videos), and VideoCoAtt (videos), and obtain the first results for automatically classifying clinically-relevant gaze behavior without wearable cameras or eye trackers.

Deep neural networks (DNNs) are notorious for their vulnerability to adversarial attacks, which are small perturbations added to their input images to mislead their prediction. Detection of adversarial examples is, therefore, a fundamental requirement for robust classification frameworks. In this work, we present a method for detecting such adversarial attacks, which is suitable for any pre-trained neural network classifier. We use influence functions to measure the impact of every training sample on the validation set data. From the influence scores, we find the most supportive training samples for any given validation example. A k-nearest neighbor (k-NN) model fitted on the DNN's activation layers is employed to search for the ranking of these supporting training samples. We observe that these samples are highly correlated with the nearest neighbors of the normal inputs, while this correlation is much weaker for adversarial inputs. We train an adversarial detector using the k-NN ranks and distances and show that it successfully distinguishes adversarial examples, getting state-of-the-art results on six attack methods with three datasets. Code is available at https://github.com/giladcohen/NNIF_adv_defense.

Human-Object Interaction (HOI) detection lies at the core of action understanding. Besides 2D information such as human/object appearance and locations, 3D pose is also usually utilized in HOI learning since its view-independence. However, rough 3D body joints just carry sparse body information and are not sufficient to understand complex interactions. Thus, we need detailed 3D body shape to go further. Meanwhile, the interacted object in 3D is also not fully studied in HOI learning. In light of these, we propose a detailed 2D-3D joint representation learning method. First, we utilize the single-view human body capture method to obtain detailed 3D body, face and hand shapes. Next, we estimate the 3D object location and size with reference to the 2D human-object spatial configuration and object category priors. Finally, a joint learning framework and cross-modal consistency tasks are proposed to learn the joint HOI representation. To better evaluate the 2D ambiguity processing capacity of models, we propose a new benchmark named Ambiguous-HOI consisting of hard ambiguous images. Extensive experiments in large-scale HOI benchmark and Ambiguous-HOI show impressive effectiveness of our method. Code and data are available at https://github.com/DirtyHarryLYL/DJ-RN.

This paper looks at this intriguing question: are single images with their details lost during deraining, reversible to their artifact-free status? We propose an end-to-end detail-recovery image deraining network (termed a DRDNet) to solve the problem. Unlike existing image deraining approaches that attempt to meet the conflicting goal of simultaneously deraining and preserving details in a unified framework, we propose to view rain removal and detail recovery as two seperate tasks, so that each part could specialize rather than trade-off between two conflicting goals. Specifically, we introduce two parallel sub-networks with a comprehensive loss function which synergize to derain and recover the lost details caused by deraining. For complete rain removal, we present a rain residual network with the squeeze-and-excitation (SE) operation to remove rain streaks from the rainy images. For detail recovery, we construct a specialized detail repair network consisting of welldesigned blocks, named structure detail context aggregation block (SDCAB), to encourage the lost details to return for eliminating image degradations. Moreover, the detail recovery branch of our proposed detail repair framework is detachable and can be incorporated into existing deraining methods to boost their performances. DRD-Net has been validated on several well-known benchmark datasets in terms of deraining robustness and detail accuracy. Comparisons show clear visual and numerical improvements of our method over the state-of-the-arts.

In this work, we present a new network design paradigm. Our goal is to help advance the understanding of network design and discover design principles that generalize across settings. Instead of focusing on designing individual network instances, we design network design spaces that parametrize populations of networks. The overall process is analogous to classic manual design of networks, but elevated to the design space level. Using our methodology we explore the structure aspect of network design and arrive at a low-dimensional design space consisting of simple, regular networks that we call RegNet. The core insight of the RegNet parametrization is surprisingly simple: widths and depths of good networks can be explained by a quantized linear function. We analyze the RegNet design space and arrive at interesting findings that do not match the current practice of network design. The RegNet design space provides simple and fast networks that work well across a wide range of flop regimes. Under comparable training settings and flops, the RegNet models outperform the popular EfficientNet models while being up to 5x faster on GPUs.

Depth sensing is a critical component of autonomous driving technologies, but today's LiDAR- or stereo camera- based solutions have limited range. We seek to increase the maximum range of self-driving vehicles' depth perception modules for the sake of better safety. To that end, we propose a novel three-camera system that utilizes small field of view cameras. Our system, along with our novel algorithm for computing metric depth, does not require full pre-calibration and can output dense depth maps with practically acceptable accuracy for scenes and objects at long distances not well covered by most commercial LiDARs.

Current 3D detection networks either rely on 2D object proposals or try to directly predict bounding box parameters from each point in a scene. While former methods are dependent on performance of 2D detectors, latter approaches are challenging due to the sparsity and occlusion in point clouds, making it difficult to regress accurate parameters. In this work, we introduce a novel approach for 3D object detection that is significant in two main aspects: a) cascaded modular approach that focuses the receptive field of each module on specific points in the point cloud, for improved feature learning and b) a class agnostic instance segmentation module that is initiated using unsupervised clustering. The objective of a cascaded approach is to sequentially minimize the number of points running through the network. While three different modules perform the tasks of background-foreground segmentation, class agnostic instance segmentation and object detection, through individually trained point based networks. We also evaluate bayesian uncertainty in modules, demonstrating the over all level of confidence in our prediction results. Performance of the network is evaluated on the SUN RGB-D benchmark dataset, that demonstrates an improvement as compared to state-of-the-art methods.

Semi-supervised learning (SSL) has been extensively studied to improve the generalization ability of deep neural networks for visual recognition. To involve the unlabelled data, most existing SSL methods are based on common density-based cluster assumption: samples lying in the same high-density region are likely to belong to the same class, including the methods performing consistency regularization or generating pseudo-labels for the unlabelled images. Despite their impressive performance, we argue three limitations exist: 1) Though the density information is demonstrated to be an important clue, they all use it in an implicit way and have not exploited it in depth. 2) For feature learning, they often learn the feature embedding based on the single data sample and ignore the neighborhood information. 3) For label-propagation based pseudo-label generation, it is often done offline and difficult to be end-to-end trained with feature learning. Motivated by these limitations, this paper proposes to solve the SSL problem by building a novel density-aware graph, based on which the neighborhood information can be easily leveraged and the feature learning and label propagation can also be trained in an end-to-end way. Specifically, we first propose a new Density-aware Neighborhood Aggregation(DNA) module to learn more discriminative features by incorporating the neighborhood information in a density-aware manner. Then a novel Density-ascending Path based Label Propagation(DPLP) module is proposed to generate the pseudo-labels for unlabeled samples more efficiently according to the feature distribution characterized by density. Finally, the DNA module and DPLP module evolve and improve each other end-to-end. Extensive experiments demonstrate the effectiveness of the newly proposed density-aware graph based SSL framework and our approach can outperform current state-of-the-art methods by a large margin.

Clustering has many applications in research and industry. However, traditional clustering methods, such as K-means, DBSCAN and HAC, impose oversimplifying assumptions and thus are not well-suited to face clustering. To adapt to the distribution of realistic problems, a natural approach is to use Graph Convolutional Networks (GCNs) to enhance features for clustering. However, GCNs can only utilize local information, which ignores the overall characterisitcs of the clusters. In this paper, we propose a Density-Aware Feature Embedding Network (DA-Net) for the task of face clustering, which utilizes both local and non-local information, to learn a robust feature embedding. Specifically, DA-Net uses GCNs to aggregate features locally, and then incorporates non-local information using a density chain, which is a chain of faces from low density to high density. This density chain exploits the non-uniform distribution of face images in the dataset. Then, an LSTM takes the density chain as input to generate the final feature embedding. Once this embedding is generated, traditional clustering methods, such as density-based clustering, can be used to obtain the final clustering results. Extensive experiments verify the effectiveness of the proposed feature embedding method, which can achieve state-of-the-art performance on public benchmarks.

Neural architecture search (NAS) has dramatically advanced the development of neural network design. We revisit the search space design in most previous NAS methods and find the number and widths of blocks are set manually. However, block counts and block widths determine the network scale (depth and width) and make a great influence on both the accuracy and the model cost (FLOPs/latency). In this paper, we propose to search block counts and block widths by designing a densely connected search space, i.e., DenseNAS. The new search space is represented as a dense super network, which is built upon our designed routing blocks. In the super network, routing blocks are densely connected and we search for the best path between them to derive the final architecture. We further propose a chained cost estimation algorithm to approximate the model cost during the search. Both the accuracy and model cost are optimized in DenseNAS. For experiments on the MobileNetV2-based search space, DenseNAS achieves 75.3% top-1 accuracy on ImageNet with only 361MB FLOPs and 17.9ms latency on a single TITAN-XP. The larger model searched by DenseNAS achieves 76.1% accuracy with only 479M FLOPs. DenseNAS further promotes the ImageNet classification accuracies of ResNet-18, -34 and -50-B by 1.5%, 0.5% and 0.3% with 200M, 600M and 680M FLOPs reduction respectively. The related code is available at https://github.com/JaminFong/DenseNAS.

We address the problem of video grounding from natural language queries. The key challenge in this task is that one training video might only contain a few annotated starting/ending frames that can be used as positive examples for model training. Most conventional approaches directly train a binary classifier using such imbalance data, thus achieving inferior results. The key idea of this paper is to use the distances between the frame within the ground truth and the starting (ending) frame as dense supervisions to improve the video grounding accuracy. Specifically, we design a novel dense regression network (DRN) to regress the distances from each frame to the starting (ending) frame of the video segment described by the query. We also propose a simple but effective IoU regression head module to explicitly consider the localization quality of the grounding results (i.e., the IoU between the predicted location and the ground truth). Experimental results show that our approach significantly outperforms state-of-the-arts on three datasets (i.e., Charades-STA, ActivityNet-Captions, and TACoS).

Our goal is to capture the pose of real animals using synthetic training examples, without using any manual supervision. Our focus is on neuroscience model organisms, to be able to study how neural circuits orchestrate behaviour. Human pose estimation attains remarkable accuracy when trained on real or simulated datasets consisting of millions of frames. However, for many applications simulated models are unrealistic and real training datasets with comprehensive annotations do not exist. We address this problem with a new sim2real domain transfer method. Our key contribution is the explicit and independent modeling of appearance, shape and pose in an unpaired image translation framework. Our model lets us train a pose estimator on the target domain by transferring readily available body keypoint locations from the source domain to generated target images. We compare our approach with existing domain transfer methods and demonstrate improved pose estimation accuracy on Drosophila melanogaster (fruit fly), Caenorhabditis elegans (worm) and Danio rerio (zebrafish), without requiring any manual annotation on the target domain and despite using simplistic off-the-shelf animal characters for simulation, or simple geometric shapes as models. Our new datasets, code and trained models will be published to support future computer vision and neuroscientific studies.

Siamese-based trackers have achieved excellent performance on visual object tracking. However, the target template is not updated online, and the features of target template and search image are computed independently in a Siamese architecture. In this paper, we propose Deformable Siamese Attention Networks, referred to as SiamAttn, by introducing a new Siamese attention mechanism that computes deformable self-attention and cross-attention. The self-attention learns strong context information via spatial attention, and selectively emphasizes interdependent channel-wise features with channel attention. The crossattention is capable of aggregating rich contextual interdependencies between the target template and the search image, providing an implicit manner to adaptively update the target template. In addition, we design a region refinement module that computes depth-wise cross correlations between the attentional features for more accurate tracking. We conduct experiments on six benchmarks, where our method achieves new state-of-the-art results, outperforming recent strong baseline, SiamRPN++, by 0.464 to 0.537 and 0.415 to 0.470 EAO on VOT 2016 and 2018.

Recently, a new paradigm of the adversarial attack on the quantized neural network weights has attracted great attention, namely, the Bit-Flip based adversarial weight attack, aka. Bit-Flip Attack (BFA). BFA has shown extraordinary attacking ability, where the adversary can malfunction a quantized Deep Neural Network (DNN) as a random guess, through malicious bit-flips on a small set of vulnerable weight bits (e.g., 13 out of 93 millions bits of 8-bit quantized ResNet-18). However, there are no effective defensive methods to enhance the fault-tolerance capability of DNN against such BFA. In this work, we conduct comprehensive investigations on BFA and propose to leverage binarization-aware training and its relaxation -- piece-wise clustering as simple and effective countermeasures to BFA. The experiments show that, for BFA to achieve the identical prediction accuracy degradation (e.g., below 11% on CIFAR-10), it requires 19.3x and 480.1x more effective malicious bit-flips on ResNet-20 and VGG-11 respectively, compared to defend-free counterparts.