CVPR 2020 Papers

Image style transfer is an underdetermined problem, where a large number of solutions can satisfy the same constraint (the content and style). Although there have been some efforts to improve the diversity of style transfer by introducing an alternative diversity loss, they have restricted generalization, limited diversity and poor scalability. In this paper, we tackle these limitations and propose a simple yet effective method for diversified arbitrary style transfer. The key idea of our method is an operation called deep feature perturbation (DFP), which uses an orthogonal random noise matrix to perturb the deep image feature maps while keeping the original style information unchanged. Our DFP operation can be easily integrated into many existing WCT (whitening and coloring transform)-based methods, and empower them to generate diverse results for arbitrary styles. Experimental results demonstrate that this learning-free and universal method can greatly increase the diversity while maintaining the quality of stylization.

We introduce a simple but effective unsupervised method for generating diverse images. We train a class-conditional GAN model without using manually annotated class labels. Instead, our model is conditional on labels automatically derived from clustering in the discriminator's feature space. Our clustering step automatically discovers diverse modes, and explicitly requires the generator to cover them. Experiments on standard mode collapse benchmarks show that our method outperforms several competing methods when addressing mode collapse. Our method also performs well on large-scale datasets such as ImageNet and Places365, improving both diversity and standard metrics (e.g., Frechet Inception Distance), compared to previous methods.

Graph representation learning aims to encode all nodes of a graph into low-dimensional vectors that will serve as input of many computer vision tasks. However, most existing algorithms ignore the existence of inherent data distribution and even noises. This may significantly increase the phenomenon of over-fitting and deteriorate the testing accuracy. In this paper, we propose a Distribution-induced Bidirectional Generative Adversarial Network (named DBGAN) for graph representation learning. Instead of the widely used Gaussian assumption, the prior distribution of latent representation in our DBGAN is estimated in a structure-aware way, which implicitly bridges the graph and content spaces by prototype learning. Thus discriminative and robust representations are generated for all nodes. Furthermore, to improve their generalization ability while preserving representation ability, the sample-level and distribution-level consistency are well balanced via a bidirectional adversarial learning framework. An extensive group of experiments is then carefully designed and presented, demonstrating that our DBGAN obtains remarkably more favorable trade-off between representation and robustness, and meanwhile is dimension-efficient, over currently available alternatives in various tasks.

While being the de facto standard coordinate representation for human pose estimation, heatmap has not been investigated in-depth. This work fills this gap. For the first time, we find that the process of decoding the predicted heatmaps into the final joint coordinates in the original image space is surprisingly significant for the performance. We further probe the design limitations of the standard coordinate decoding method, and propose a more principled distributionaware decoding method. Also, we improve the standard coordinate encoding process (i.e. transforming ground-truth coordinates to heatmaps) by generating unbiased/accurate heatmaps. Taking the two together, we formulate a novel Distribution-Aware coordinate Representation of Keypoints (DARK) method. Serving as a model-agnostic plug-in, DARK brings about significant performance boost to existing human pose estimation models. Extensive experiments show that DARK yields the best results on two common benchmarks, MPII and COCO. Besides, DARK achieves the 2nd place entry in the ICCV 2019 COCO Keypoints Challenge. The code is available online.

Watermarking is the process of embedding information into an image that can survive under distortions, while requiring the encoded image to have little or no perceptual difference with the original image. Recently, deep learning-based methods achieved impressive results in both visual quality and message payload under a wide variety of image distortions. However, these methods all require differentiable models for the image distortions at training time, and may generalize poorly to unknown distortions. This is undesirable since the types of distortions applied to watermarked images are usually unknown and non-differentiable. In this paper, we propose a new framework for distortion-agnostic watermarking, where the image distortion is not explicitly modeled during training. Instead, the robustness of our system comes from two sources: adversarial training and channel coding. Compared to training on a fixed set of distortions and noise levels, our method achieves comparable or better results on distortions available during training, and better performance overall on unknown distortions.

Existing knowledge distillation methods focus on convolutional neural networks (CNNs), where the input samples like images lie in a grid domain, and have largely overlooked graph convolutional networks (GCN) that handle non-grid data. In this paper, we propose to our best knowledge the first dedicated approach to distilling knowledge from a pre-trained GCN model. To enable the knowledge transfer from the teacher GCN to the student, we propose a local structure preserving module that explicitly accounts for the topological semantics of the teacher. In this module, the local structure information from both the teacher and the student are extracted as distributions, and hence minimizing the distance between these distributions enables topology-aware knowledge transfer from the teacher, yielding a compact yet high-performance student model. Moreover, the proposed approach is readily extendable to dynamic graph models, where the input graphs for the teacher and the student may differ. We evaluate the proposed method on two different datasets using GCN models of different architectures, and demonstrate that our method achieves the state-of-the-art knowledge distillation performance for GCN models.

State-of-the-art deep dehazing models are often difficult in training. Knowledge distillation paves a way to train a student network assisted by a teacher network. However, most knowledge distill methods are used for image classification and segmentation as well as object detection, and few investigate distilling image restoration and use different task for knowledge transfer. In this paper, we propose a knowledge-distill dehazing network which distills image dehazing with the heterogeneous task imitation. In our network, the teacher is an off-the-shelf auto-encoder network and is used for image reconstruction. The dehazing network is trained assisted by the teacher network with the process-oriented learning mechanism. The student network imitates the task of image reconstruction in the teacher network. Moreover, we design a spatial-weighted channel-attention residual block for the student image dehazing network to adaptively learn the content-aware channel level attention and pay more attention to the features for dense hazy regions reconstruction. To evaluate the effectiveness of the proposed method, we compare our method with several state-of-the-art methods on two synthetic and real-world datasets, as well as real hazy images.

Collecting large-scale data with clean labels for supervised training of neural networks is practically challenging. Although noisy labels are usually cheap to acquire, existing methods suffer a lot from label noise. This paper targets at the challenge of robust training at high label noise regimes. The key insight to achieve this goal is to wisely leverage a small trusted set to estimate exemplar weights and pseudo labels for noisy data in order to reuse them for supervised training. We present a holistic framework to train deep neural networks in a way that is highly invulnerable to label noise. Our method sets the new state of the art on various types of label noise and achieves excellent performance on large-scale datasets with real-world label noise. For instance, on CIFAR100 with a 40% uniform noise ratio and only 10 trusted labeled data per class, our method achieves 80.2% classification accuracy, where the error rate is only 1.4% higher than a neural network trained without label noise. Moreover, increasing the noise ratio to 80%, our method still maintains a high accuracy of 75.5%, compared to the previous best accuracy 48.2%.

The discriminative knowledge from a high-capacity deep neural network (a.k.a. the "teacher") could be distilled to facilitate the learning efficacy of a shallow counterpart (a.k.a. the "student"). This paper deals with a general scenario reusing the knowledge from a cross-task teacher --- two models are targeting non-overlapping label spaces. We emphasize that the comparison ability between instances acts as an essential factor threading knowledge across domains, and propose the RElationship FacIlitated Local cLassifiEr Distillation (ReFilled) approach, which decomposes the knowledge distillation flow into branches for embedding and the top-layer classifier. In particular, different from reconciling the instance-label confidence between models, ReFilled requires the teacher to reweight the hard triplets push forwarded by the student so that the similarity comparison levels between instances are matched. A local embedding-induced classifier from the teacher further supervises the student's classification confidence. ReFilled demonstrates its effectiveness when reusing cross-task models, and also achieves state-of-the-art performance on the standard knowledge distillation benchmarks. The code of the paper can be accessed at https://github.com/njulus/ReFilled.

Whole understanding of the surroundings is paramount to autonomous systems. Recent works have shown that deep neural networks can learn geometry (depth) and motion (optical flow) from a monocular video without any explicit supervision from ground truth annotations, particularly hard to source for these two tasks. In this paper, we take an additional step toward holistic scene understanding with monocular cameras by learning depth and motion alongside with semantics, with supervision for the latter provided by a pre-trained network distilling proxy ground truth images. We address the three tasks jointly by a) a novel training protocol based on knowledge distillation and self-supervision and b) a compact network architecture which enables efficient scene understanding on both power hungry GPUs and low-power embedded platforms. We thoroughly assess the performance of our framework and show that it yields state-of-the-art results for monocular depth estimation, optical flow and motion segmentation.

Under stereo settings, the problems of disparity estimation, stereo magnification and stereo-view synthesis have gathered wide attention. However, the limited image quality brings non-negligible difficulties in developing related applications and becomes the main bottleneck of stereo images. To the best of our knowledge, stereo image restoration is rarely studied. Towards this end, this paper analyses how to effectively explore disparity information, and proposes a unified stereo image restoration framework. The proposed framework explicitly learn the inherent pixel correspondence between stereo views and restores stereo image with the cross-view information at image and feature level. A Feature Modulation Dense Block (FMDB) is introduced to insert disparity prior throughout the whole network. The experiments in terms of efficiency, objective and perceptual quality, and the accuracy of depth estimation demonstrates the superiority of the proposed framework on various stereo image restoration tasks.

In this paper, we propose a novel system named Disp R-CNN for 3D object detection from stereo images. Many recent works solve this problem by first recovering a point cloud with disparity estimation and then apply a 3D detector. The disparity map is computed for the entire image, which is costly and fails to leverage category-specific prior. In contrast, we design an instance disparity estimation network (iDispNet) that predicts disparity only for pixels on objects of interest and learns a category-specific shape prior for more accurate disparity estimation. To address the challenge from scarcity of disparity annotation in training, we propose to use a statistical shape model to generate dense disparity pseudo-ground-truth without the need of LiDAR point clouds, which makes our system more widely applicable. Experiments on the KITTI dataset show that, even when LiDAR ground-truth is not available at training time, Disp R-CNN achieves competitive performance and outperforms previous state-of-the-art methods by 20% in terms of average precision. The code will be available at https://github.com/zju3dv/disprcnn.

Spatial-temporal graphs have been widely used by skeleton-based action recognition algorithms to model human action dynamics. To capture robust movement patterns from these graphs, long-range and multi-scale context aggregation and spatial-temporal dependency modeling are critical aspects of a powerful feature extractor. However, existing methods have limitations in achieving (1) unbiased long-range joint relationship modeling under multi-scale operators and (2) unobstructed cross-spacetime information flow for capturing complex spatial-temporal dependencies. In this work, we present (1) a simple method to disentangle multi-scale graph convolutions and (2) a unified spatial-temporal graph convolutional operator named G3D. The proposed multi-scale aggregation scheme disentangles the importance of nodes in different neighborhoods for effective long-range modeling. The proposed G3D module leverages dense cross-spacetime edges as skip connections for direct information propagation across the spatial-temporal graph. By coupling these proposals, we develop a powerful feature extractor named MS-G3D based on which our model outperforms previous state-of-the-art methods on three large-scale datasets: NTU RGB+D 60, NTU RGB+D 120, and Kinetics Skeleton 400.

Leveraging physical knowledge described by partial differential equations (PDEs) is an appealing way to improve unsupervised video forecasting models. Since physics is too restrictive for describing the full visual content of generic video sequences, we introduce PhyDNet, a two-branch deep architecture, which explicitly disentangles PDE dynamics from unknown complementary information. A second contribution is to propose a new recurrent physical cell (PhyCell), inspired from data assimilation techniques, for performing PDE-constrained prediction in latent space. Extensive experiments conducted on four various datasets show the ability of PhyDNet to outperform state-of-the-art methods. Ablation studies also highlight the important gain brought out by both disentanglement and PDE-constrained prediction. Finally, we show that PhyDNet presents interesting features for dealing with missing data and long-term forecasting.

We propose an approach for face image generation of virtual people with disentangled, precisely-controllable latent representations for identity of non-existing people, expression, pose, and illumination. We embed 3D priors into adversarial learning and train the network to imitate the image formation of an analytic 3D face deformation and rendering process. To deal with the generation freedom induced by the domain gap between real and rendered faces, we further introduce contrastive learning to promote disentanglement by comparing pairs of generated images. Experiments show that through our imitative-contrastive learning, the factor variations are very well disentangled and the properties of a generated face can be precisely controlled. We also analyze the learned latent space and present several meaningful properties supporting factor disentanglement. Our method can also be used to embed real images into the disentangled latent space. We hope our method could provide new understandings of the relationship between physical properties and deep image synthesis.

We explore different design choices for injecting noise into generative adversarial networks (GANs) with the goal of disentangling the latent space. Instead of traditional approaches, we propose feeding multiple noise codes through separate fully-connected layers respectively. The aim is restricting the influence of each noise code to specific parts of the generated image. We show that disentanglement in the first layer of the generator network leads to disentanglement in the generated image. Through a grid-based structure, we achieve several aspects of disentanglement without complicating the network architecture and without requiring labels. We achieve spatial disentanglement, scale-space disentanglement, and disentanglement of the foreground object from the background style allowing fine-grained control over the generated images. Examples include changing facial expressions in face images, changing beak length in bird images, and changing car dimensions in car images. This empirically leads to better disentanglement scores than state-of-the-art methods on the FFHQ dataset.

Vision is often used as a complementary modality for audio speech recognition (ASR), especially in the noisy environment where performance of solo audio modality significantly deteriorates. After combining visual modality, ASR is upgraded to the multi-modality speech recognition (MSR). In this paper, we propose a two-stage speech recognition model. In the first stage, the target voice is separated from background noises with help from the corresponding visual information of lip movements, making the model 'listen' clearly. At the second stage, the audio modality combines visual modality again to better understand the speech by a MSR sub-network, further improving the recognition rate. There are some other key contributions: we introduce a pseudo-3D residual convolution (P3D)-based visual front-end to extract more discriminative features; we upgrade the temporal convolution block from 1D ResNet with the temporal convolutional network (TCN), which is more suitable for the temporal tasks; the MSR sub-network is built on the top of Element-wise-Attention Gated Recurrent Unit (EleAtt-GRU), which is more effective than Transformer in long sequences. We conducted extensive experiments on the LRS3-TED and the LRW datasets. Our two-stage model (audio enhanced multi-modality speech recognition, AE-MSR) consistently achieves the state-of-the-art performance by a significant margin, which demonstrates the necessity and effectiveness of AE-MSR.

In this paper, we target to address the problem of compression and acceleration of Convolutional Neural Networks (CNNs). Specifically, we propose a novel structural pruning method to obtain a compact CNN with strong discriminative power. To find such networks, we propose an efficient discrete optimization method to directly optimize channel-wise differentiable discrete gate under resource constraint while freezing all the other model parameters. Although directly optimizing discrete variables is a complex non-smooth, non-convex and NP-hard problem, our optimization method can circumvent these difficulties by using the straight-through estimator. Thus, our method is able to ensure that the sub-network discovered within the training process reflects the true sub-network. We further extend the discrete gate to its stochastic version in order to thoroughly explore the potential sub-networks. Unlike many previous methods requiring per-layer hyper-parameters, we only require one hyper-parameter to control FLOPs budget. Moreover, our method is globally discrimination-aware due to the discrete setting. The experimental results on CIFAR-10 and ImageNet show that our method is competitive with state-of-the-art methods.

Finding the largest subset of sequences (i.e., time series) that are correlated above a certain threshold, within large datasets, is of significant interest for computer vision and pattern recognition problems across domains, including behavior analysis, computational biology, neuroscience, and finance. Maximal clique algorithms can be used to solve this problem, but they are not scalable. We present an approximate, but highly efficient and scalable, method that represents the search space as a union of sets called epsilon-expanded clusters, one of which is theoretically guaranteed to contain the largest subset of synchronized sequences. The method finds synchronized sets by fitting a Euclidean ball on epsilon-expanded clusters, using Jung's theorem. We validate the method on data from the three distinct domains of facial behavior analysis, finance, and neuroscience, where we respectively discover the synchrony among pixels of face videos, stock market item prices, and dynamic brain connectivity data. Experiments show that our method produces results comparable to, but up to 300 times faster than, maximal clique algorithms, with speed gains increasing exponentially with the number of input sequences.