CVPR 2020 Papers

3D human motion prediction, i.e., forecasting future sequences from given historical poses, is a fundamental task for action analysis, human-computer interaction, machine intelligence. Recently, the state-of-the-art method assumes that the whole human motion sequence involves a fully-connected graph formed by links between each joint pair. Although encouraging performance has been made, due to the neglect of the inherent and meaningful characteristics of the natural connectivity of human joints, unexpected results may be produced. Moreover, such a complicated topology greatly increases the training difficulty. To tackle these issues, we propose a deep generative model based on graph networks and adversarial learning. Specifically, the skeleton pose is represented as a novel dynamic graph, in which natural connectivities of the joint pairs are exploited explicitly, and the links of geometrically separated joints can also be learned implicitly. Notably, in the proposed model, the natural connection strength is adaptively learned, whereas, in previous schemes, it was constant. Our approach is evaluated on two representations (i.e., angle-based, position-based) from various large-scale 3D skeleton benchmarks (e.g., H3.6M, CMU, 3DPW MoCap). Extensive experiments demonstrate that our approach achieves significant improvements against existing baselines in accuracy and visualization. Code will be available at https://github.com/cuiqiongjie/LDRGCN.

3D object detection from a single image without LiDAR is a challenging task due to the lack of accurate depth information. Conventional 2D convolutions are unsuitable for this task because they fail to capture local object and its scale information, which are vital for 3D object detection. To better represent 3D structure, prior arts typically transform depth maps estimated from 2D images into a pseudo-LiDAR representation, and then apply existing 3D point-cloud based object detectors. However, their results depend heavily on the accuracy of the estimated depth maps, resulting in suboptimal performance. In this work, instead of using pseudo-LiDAR representation, we improve the fundamental 2D fully convolutions by proposing a new local convolutional network (LCN), termed Depth-guided Dynamic-Depthwise-Dilated LCN (D4LCN), where the filters and their receptive fields can be automatically learned from image-based depth maps, making different pixels of different images have different filters. D4LCN overcomes the limitation of conventional 2D convolutions and narrows the gap between image representation and 3D point cloud representation. Extensive experiments show that D^4LCN outperforms existing works by large margins. For example, the relative improvement of D4LCN against the state-of-the-art on KITTI is 9.1% in the moderate setting. D4LCN ranks 1st on KITTI monocular 3D object detection benchmark at the time of submission (car, December 2019). The code is available at https://github.com/dingmyu/D4LCN

The state-of-art 6D object pose detection methods use convolutional neural networks to estimate objects' 6D poses from RGB images. However, they require huge numbers of images with explicit 3D annotations such as 6D poses, 3D bounding boxes and 3D keypoints, either obtained by manual labeling or inferred from synthetic images generated by 3D CAD models. Manual labeling for a large number of images is a laborious task, and we usually do not have the corresponding 3D CAD models of objects in real environment. In this paper, we develop a keypoint-based 6D object pose detection method (and its deep network) called Object Keypoint based POSe Estimation (OK-POSE). OK-POSE employs relative transformation between viewpoints for training. Specifically, we use pairs of images with object annotation and relative transformation information between their viewpoints to automatically discover objects' 3D keypoints which are geometrically and visually consistent. Then, the 6D object pose can be estimated using a keypoint-based geometric reasoning method with a reference viewpoint. The relative transformation information can be easily obtained from any cheap binocular cameras or most smartphone devices, thus greatly lowering the labeling cost. Experiments have demonstrated that OK-POSE achieves acceptable performance compared to methods relying on the object's 3D CAD model or a great deal of 3D labeling. These results show that our method can be used as a suitable alternative when there are no 3D CAD models or a large number of 3D annotations.

Learning-based approaches to graph matching have been developed and explored for more than a decade, have grown rapidly in scope and popularity in recent years. However, previous learning-based algorithms, with or without deep learning strategy, mainly focus on the learning of node and/or edge affinities generation, and pay less attention on the learning of the combinatorial solver. In this paper we propose a fully trainable framework for graph matching, in which learning of affinities and solving for combinatorial optimization are not explicitly separated as in many previous arts. We firstly convert the problem of building node correspondences between two input graphs to the problem of selecting reliable nodes from a constructed assignment graph. Subsequently, the graph network block module is adopted to perform computation on the graph to form structured representations for each node. It finally predicts a label for each node that is used for node classification, and the training is performed under the supervision of both permutation differences and the one-to-one matching constraints. The proposed method is evaluated on four public benchmarks in comparison with several state-of-the-art algorithms, and the experimental results illustrate its excellent performance.

We present a novel approach to category-level 6D object pose and size estimation. To tackle intra-class shape variations, we learn canonical shape space (CASS), a unified representation for a large variety of instances of a certain object category. In particular, CASS is modeled as the latent space of a deep generative model of canonical 3D shapes with normalized pose. We train a variational auto-encoder (VAE) for generating 3D point clouds in the canonical space from an RGBD image. The VAE is trained in a cross-category fashion, exploiting the publicly available large 3D shape repositories. Since the 3D point cloud is generated in normalized pose (with actual size), the encoder of the VAE learns view-factorized RGBD embedding. It maps an RGBD image in arbitrary view into a poseindependent 3D shape representation. Object pose is then estimated via contrasting it with a pose-dependent feature of the input RGBD extracted with a separate deep neural networks. We integrate the learning of CASS and pose and size estimation into an end-to-end trainable network, achieving the state-of-the-art performance.

We leverage the powerful lossy image compression algorithm BPG to build a lossless image compression system. Specifically, the original image is first decomposed into the lossy reconstruction obtained after compressing it with BPG and the corresponding residual. We then model the distribution of the residual with a convolutional neural network-based probabilistic model that is conditioned on the BPG reconstruction, and combine it with entropy coding to losslessly encode the residual. Finally, the image is stored using the concatenation of the bitstreams produced by BPG and the learned residual coder. The resulting compression system achieves state-of-the-art performance in learned lossless full-resolution image compression, outperforming previous learned approaches as well as PNG, WebP, and JPEG2000.

Data augmentation can impact the generalization performance of an image classification model in a significant way. However, it is currently conducted on the basis of trial and error, and its impact on the generalization performance cannot be predicted during training. This paper considers an influence function that predicts how generalization performance, in terms of validation loss, is affected by a particular augmented training sample. The influence function provides an approximation of the change in validation loss without actually comparing the performances that include and exclude the sample in the training process. Based on this function, a differentiable augmentation network is learned to augment an input training sample to reduce validation loss. The augmented sample is fed into the classification network, and its influence is approximated as a function of the parameters of the last fully-connected layer of the classification network. By backpropagating the influence to the augmentation network, the augmentation network parameters are learned. Experimental results on CIFAR-10, CIFAR-100, and ImageNet show that the proposed method provides better generalization performance than conventional data augmentation methods do.

Scene understanding has been of high interest in computer vision. It encompasses not only identifying objects in a scene, but also their relationships within the given context. With this goal, a recent line of works tackles 3D semantic segmentation and scene layout prediction. In our work we focus on scene graphs, a data structure that organizes the entities of a scene in a graph, where objects are nodes and their relationships modeled as edges. We leverage inference on scene graphs as a way to carry out 3D scene understanding, mapping objects and their relationships. In particular, we propose a learned method that regresses a scene graph from the point cloud of a scene. Our novel architecture is based on PointNet and Graph Convolutional Networks (GCN). In addition, we introduce 3DSSG, a semiautomatically generated dataset, that contains semantically rich scene graphs of 3D scenes. We show the application of our method in a domain-agnostic retrieval task, where graphs serve as an intermediate representation for 3D-3D and 2D-3D matching.

Image compression is a fundamental research field and many well-known compression standards have been developed for many decades. Recently, learned compression methods exhibit a fast development trend with promising results. However, there is still a performance gap between learned compression algorithms and reigning compression standards, especially in terms of widely used PSNR metric. In this paper, we explore the remaining redundancy of recent learned compression algorithms. We have found accurate entropy models for rate estimation largely affect the optimization of network parameters and thus affect the rate-distortion performance. Therefore, in this paper, we propose to use discretized Gaussian Mixture Likelihoods to parameterize the distributions of latent codes, which can achieve a more accurate and flexible entropy model. Besides, we take advantage of recent attention modules and incorporate them into network architecture to enhance the performance. Experimental results demonstrate our proposed method achieves a state-of-the-art performance compared to existing learned compression methods on both Kodak and high-resolution datasets. To our knowledge our approach is the first work to achieve comparable performance with latest compression standard Versatile Video Coding (VVC) regarding PSNR. More importantly, our approach generates more visually pleasant results when optimized by MS-SSIM.

While deep neural networks have been achieving state-of-the-art performance across a wide variety of applications, their vulnerability to adversarial attacks limits their widespread deployment for safety-critical applications. Alongside other adversarial defense approaches being investigated, there has been a very recent interest in improving adversarial robustness in deep neural networks through the introduction of perturbations during the training process. However, such methods leverage fixed, pre-defined perturbations and require significant hyper-parameter tuning that makes them very difficult to leverage in a general fashion. In this study, we introduce Learn2Perturb, an end-to-end feature perturbation learning approach for improving the adversarial robustness of deep neural networks. More specifically, we introduce novel perturbation-injection modules that are incorporated at each layer to perturb the feature space and increase uncertainty in the network. This feature perturbation is performed at both the training and the inference stages. Furthermore, inspired by the Expectation-Maximization, an alternating back-propagation training algorithm is introduced to train the network and noise parameters consecutively. Experimental results on CIFAR-10 and CIFAR-100 datasets show that the proposed Learn2Perturb method can result in deep neural networks which are 4-7% more robust on l_inf FGSM and PDG adversarial attacks and significantly outperforms the state-of-the-art against l_2 C&W attack and a wide range of well-known black-box attacks.

Handwritten text and scene text suffer from various shapes and distorted patterns. Thus training a robust recognition model requires a large amount of data to cover diversity as much as possible. In contrast to data collection and annotation, data augmentation is a low cost way. In this paper, we propose a new method for text image augmentation. Different from traditional augmentation methods such as rotation, scaling and perspective transformation, our proposed augmentation method is designed to learn proper and efficient data augmentation which is more effective and specific for training a robust recognizer. By using a set of custom fiducial points, the proposed augmentation method is flexible and controllable. Furthermore, we bridge the gap between the isolated processes of data augmentation and network optimization by joint learning. An agent network learns from the output of the recognition network and controls the fiducial points to generate more proper training samples for the recognition network. Extensive experiments on various benchmarks, including regular scene text, irregular scene text and handwritten text, show that the proposed augmentation and the joint learning methods significantly boost the performance of the recognition networks. A general toolkit for geometric augmentation is available.

Current 6D object pose estimation methods usually require a 3D model for each object. These methods also require additional training in order to incorporate new objects. As a result, they are difficult to scale to a large number of objects and cannot be directly applied to unseen objects. We propose a novel framework for 6D pose estimation of unseen objects. We present a network that reconstructs a latent 3D representation of an object using a small number of reference views at inference time. Our network is able to render the latent 3D representation from arbitrary views. Using this neural renderer, we directly optimize for pose given an input image. By training our network with a large number of 3D shapes for reconstruction and rendering, our network generalizes well to unseen objects. We present a new dataset for unseen object pose estimation--MOPED. We evaluate the performance of our method for unseen object pose estimation on MOPED as well as the ModelNet and LINEMOD datasets. Our method performs competitively to supervised methods that are trained on those objects. Code and data will be available at https://keunhong.com/publications/latentfusion/

Training with more data has always been the most stable and effective way of improving performance in deep learn-ing era. As the largest object detection dataset so far, OpenImages brings great opportunities and challenges for object detection in general and sophisticated scenarios. However, owing to its semi-automatic collecting and labeling pipeline to deal with the huge data scale, Open Images dataset suffers from label-related problems that objects may explicitly or implicitly have multiple labels and the label distribution is extremely imbalanced. In this work, we quantitatively analyze these label problems and provide a simple but effective solution. We design a concurrent softmax to handle the multi-label problems in object detection and propose a soft-sampling methods with hybrid training scheduler to deal with the label imbalance. Overall, our method yields a dramatic improvement of 3.34 points, leading to the best single model with 60.90 mAP on the public object detection test set of Open Images. And our ensembling result achieves 67.17mAP, which is 4.29 points higher than the first place method last year.

Representation learning is widely applied for various tasks on multimedia data, e.g., retrieval and search. One approach for learning useful representation is by utilizing the relationships or similarities between examples. In this work, we explore two promising scalable representation learning approaches on video domain. With hierarchical graph clusters built upon video-to-video similarities, we propose: 1) smart negative sampling strategy that significantly boosts training efficiency with triplet loss, and 2) a pseudo-classification approach using the clusters as pseudo-labels. The embeddings trained with the proposed methods are competitive on multiple video understanding tasks, including related video retrieval and video annotation. Both of these proposed methods are highly scalable, as verified by experiments on large-scale datasets.

Many existing studies reveal that annotation inconsistency widely exists among a variety of facial expression recognition (FER) datasets. The reason might be the subjectivity of human annotators and the ambiguous nature of the expression labels. One promising strategy tackling such a problem is a recently proposed learning paradigm called Label Distribution Learning (LDL), which allows multiple labels with different intensity to be linked to one expression. However, it is often impractical to directly apply label distribution learning because numerous existing datasets only contain one-hot labels rather than label distributions. To solve the problem, we propose a novel approach named Label Distribution Learning on Auxiliary Label Space Graphs(LDL-ALSG) that leverages the topological information of the labels from related but more distinct tasks, such as action unit recognition and facial landmark detection. The underlying assumption is that facial images should have similar expression distributions to their neighbours in the label space of action unit recognition and facial landmark detection. Our proposed method is evaluated on a variety of datasets and outperforms those state-of-the-art methods consistently with a huge margin.

To get more accurate saliency maps, recent methods mainly focus on aggregating multi-level features from fully convolutional network (FCN) and introducing edge information as auxiliary supervision. Though remarkable progress has been achieved, we observe that the closer the pixel is to the edge, the more difficult it is to be predicted, because edge pixels have a very imbalance distribution. To address this problem, we propose a label decoupling framework (LDF) which consists of a label decoupling (LD) procedure and a feature interaction network (FIN). LD explicitly decomposes the original saliency map into body map and detail map, where body map concentrates on center areas of objects and detail map focuses on regions around edges. Detail map works better because it involves much more pixels than traditional edge supervision. Different from saliency map, body map discards edge pixels and only pays attention to center areas. This successfully avoids the distraction from edge pixels during training. Therefore, we employ two branches in FIN to deal with body map and detail map respectively. Feature interaction (FI) is designed to fuse the two complementary branches to predict the saliency map, which is then used to refine the two branches again. This iterative refinement is helpful for learning better representations and more precise saliency maps. Comprehensive experiments on six benchmark datasets demonstrate that LDF outperforms state-of-the-art approaches on different evaluation metrics.

Modern face alignment methods have become quite accurate at predicting the locations of facial landmarks, but they do not typically estimate the uncertainty of their predicted locations nor predict whether landmarks are visible. In this paper, we present a novel framework for jointly predicting landmark locations, associated uncertainties of these predicted locations, and landmark visibilities. We model these as mixed random variables and estimate them using a deep network trained using our proposed Location, Uncertainty, and Visibility Likelihood (LUVLi) loss. In addition, we release an entirely new labeling of a large face alignment dataset with over 19,000 face images in a full range of head poses. Each face is manually labeled with the ground-truth locations of 68 landmarks, with the additional information of whether each landmarks is visible, self-occluded (due to extreme head poses), or externally occluded. Not only does our joint estimation yield accurate estimates of the uncertainty of predicted landmark locations, but it also yields state-of-the-art estimates for the landmark locations themselves on mulitple standard face alignment datasets. Our method's estimates of the uncertainty of predicted landmark locations could be used to automatically identify input images on which face alignment fails, which can be critical for downstream tasks.

We introduce a one-shot segmentation method to alleviate the burden of manual annotation for medical images. The main idea is to treat one-shot segmentation as a classical atlas-based segmentation problem, where voxel-wise correspondence from the atlas to the unlabelled data is learned. Subsequently, segmentation label of the atlas can be transferred to the unlabelled data with the learned correspondence. However, since ground truth correspondence between images is usually unavailable, the learning system must be well-supervised to avoid mode collapse and convergence failure. To overcome this difficulty, we resort to the forward-backward consistency, which is widely used in correspondence problems, and additionally learn the backward correspondences from the warped atlases back to the original atlas. This cycle-correspondence learning design enables a variety of extra, cycle-consistency-based supervision signals to make the training process stable, while also boost the performance. We demonstrate the superiority of our method over both deep learning-based one-shot segmentation methods and a classical multi-atlas segmentation method via thorough experiments.

We study the energy minimization problem in low-level vision tasks from a novel perspective. We replace the heuristic regularization term with a data-driven learnable subspace constraint, and preserve the data term to exploit domain knowledge derived from the first principles of a task. This learning subspace minimization (LSM) framework unifies the network structures and the parameters for many different low-level vision tasks, which allows us to train a single network for multiple tasks simultaneously with shared parameters, and even generalizes the trained network to an unseen task as long as the data term can be formulated. We validate our LSM frame on four low-level tasks including edge detection, interactive segmentation, stereo matching, and optical flow, and validate the network on various datasets. The experiments demonstrate that the proposed LSM generates state-of-the-art results with smaller model size, faster training convergence, and real-time inference.