CVPR 2019 论文列表

We study two important concepts in adversarial deep learning---adversarial training and generative adversarial network (GAN). Adversarial training is the technique used to improve the robustness of discriminator by combining adversarial attacker and discriminator in the training phase. GAN is commonly used for image generation by jointly optimizing discriminator and generator. We show these two concepts are indeed closely related and can be used to strengthen each other---adding a generator to the adversarial training procedure can improve the robustness of discriminators, and adding an adversarial attack to GAN training can improve the convergence speed and lead to better generators. Combining these two insights, we develop a framework called Rob-GAN to jointly optimize generator and discriminator in the presence of adversarial attacks---the generator generates fake images to fool discriminator; the adversarial attacker perturbs real images to fool discriminator, and the discriminator wants to minimize loss under fake and adversarial images. Through this end-to-end training procedure, we are able to simultaneously improve the convergence speed of GAN training, the quality of synthetic images, and the robustness of discriminator under strong adversarial attacks. Experimental results demonstrate that the obtained classifier is more robust than the state-of-the-art adversarial training approach (Madry 2017), and the generator outperforms SN-GAN on ImageNet-143.

Unsupervised visual representation learning remains a largely unsolved problem in computer vision research. Among a big body of recently proposed approaches for unsupervised learning of visual representations, a class of self-supervised techniques achieves superior performance on many challenging benchmarks. A large number of the pretext tasks for self-supervised learning have been studied, but other important aspects, such as the choice of convolutional neural networks (CNN), has not received equal attention. Therefore, we revisit numerous previously proposed self-supervised models, conduct a thorough large scale study and, as a result, uncover multiple crucial insights. We challenge a number of common practices in self-supervised visual representation learning and observe that standard recipes for CNN design do not always translate to self-supervised representation learning. As part of our study, we drastically boost the performance of previously proposed techniques and outperform previously published state-of-the-art results by a large margin. We will release the code for reproducing our experiments when the anonymity requirements are lifted.

Crowd counting is the task of estimating people numbers in crowd images. Modern crowd counting methods employ deep neural networks to estimate crowd counts via crowd density regressions. A major challenge of this task lies in the perspective distortion, which results in drastic person scale change in an image. Density regression on the small person area is in general very hard. In this work, we propose a perspective-aware convolutional neural network (PACNN) for efficient crowd counting, which integrates the perspective information into density regression to provide additional knowledge of the person scale change in an image. Ground truth perspective maps are firstly generated for training; PACNN is then specifically designed to predict multi-scale perspective maps and encode them as perspective-aware weighting layers in the network to adaptively combine the outputs of multi-scale density maps. The weights are learned at every pixel of the maps such that the final density combination is robust to the perspective distortion. We conduct extensive experiments on the ShanghaiTech, WorldExpo'10, UCF_CC_50, and UCSD datasets, and demonstrate the effectiveness and efficiency of PACNN over the state-of-the-art.

Few-shot learning in image classification aims to learn a classifier to classify images when only few training examples are available for each class. Recent work has achieved promising classification performance, where an image-level feature based measure is usually used. In this paper, we argue that a measure at such a level may not be effective enough in light of the scarcity of examples in few-shot learning. Instead, we think a local descriptor based image-to-class measure should be taken, inspired by its surprising success in the heydays of local invariant features. Specifically, building upon the recent episodic training mechanism, we propose a Deep Nearest Neighbor Neural Network (DN4 in short) and train it in an end-to-end manner. Its key difference from the literature is the replacement of the image-level feature based measure in the final layer by a local descriptor based image-to-class measure. This measure is conducted online via a k-nearest neighbor search over the deep local descriptors of convolutional feature maps. The proposed DN4 not only learns the optimal deep local descriptors for the image-to-class measure, but also utilizes the higher efficiency of such a measure in the case of example scarcity, thanks to the exchangeability of visual patterns across the images in the same class. Our work leads to a simple, effective, and computationally efficient framework for few-shot learning. Experimental study on benchmark datasets consistently shows its superiority over the related state-of-the-art, with the largest absolute improvement of 17% over the next best. The source code can be available from https://github.com/WenbinLee/DN4.git.

The pix2pix and CycleGAN losses have vastly improved the qualitative and quantitative visual quality of results in image-to-image translation tasks. We extend this framework by exploring approximately invertible architectures which are well suited to these losses. These architectures are approximately invertible by design and thus partially satisfy cycle-consistency before training even begins. Furthermore, since invertible architectures have constant memory complexity in depth, these models can be built arbitrarily deep. We are able to demonstrate superior quantitative output on the Cityscapes and Maps datasets at near constant memory budget.

Many 3D vision systems localize cameras within a scene using 3D point clouds. Such point clouds are often obtained using structure from motion (SfM), after which the images are discarded to preserve privacy. In this paper, we show, for the first time, that such point clouds retain enough information to reveal scene appearance and compromise privacy. We present a privacy attack that reconstructs color images of the scene from the point cloud. Our method is based on a cascaded U-Net that takes as input, a 2D multichannel image of the points rendered from a specific viewpoint containing point depth and optionally color and SIFT descriptors and outputs a color image of the scene from that viewpoint. Unlike previous feature inversion methods, we deal with highly sparse and irregular 2D point distributions and inputs where many point attributes are missing, namely keypoint orientation and scale, the descriptor image source and the 3D point visibility. We evaluate our attack algorithm on public datasets and analyze the significance of the point cloud attributes. Finally, we show that novel views can also be generated thereby enabling compelling virtual tours of the underlying scene.

We propose a retrieval-augmented convolutional network (RaCNN) and propose to train it with local mixup, a novel variant of the recently proposed mixup algorithm. The proposed hybrid architecture combining a convolutional network and an off-the-shelf retrieval engine was designed to mitigate the adverse effect of off-manifold adversarial examples, while the proposed local mixup addresses on-manifold ones by explicitly encouraging the classifier to locally behave linearly on the data manifold. Our evaluation of the proposed approach against seven readilyavailable adversarial attacks on three datasets-CIFAR-10, SVHN and ImageNet-demonstrate the improved robustness compared to a vanilla convolutional network, and comparable performance with the state-of-the-art reactive defense approaches.

Video summarization is a technique to create a short skim of the original video while preserving the main stories/content. There exists a substantial interest in automatizing this process due to the rapid growth of the available material. The recent progress has been facilitated by public benchmark datasets, which enable easy and fair comparison of methods. Currently the established evaluation protocol is to compare the generated summary with respect to a set of reference summaries provided by the dataset. In this paper, we will provide in-depth assessment of this pipeline using two popular benchmark datasets. Surprisingly, we observe that randomly generated summaries achieve comparable or better performance to the state-of-the-art. In some cases, the random summaries outperform even the human generated summaries in leave-one-out experiments. Moreover, it turns out that the video segmentation, which is often considered as a fixed pre-processing method, has the most significant impact on the performance measure. Based on our observations, we propose alternative approaches for assessing the importance scores as well as an intuitive visualization of correlation between the estimated scoring and human annotations.

Graph convolutional neural networks have recently shown great potential for the task of zero-shot learning. These models are highly sample efficient as related concepts in the graph structure share statistical strength allowing generalization to new classes when faced with a lack of data. However, multi-layer architectures, which are required to propagate knowledge to distant nodes in the graph, dilute the knowledge by performing extensive Laplacian smoothing at each layer and thereby consequently decrease performance. In order to still enjoy the benefit brought by the graph structure while preventing dilution of knowledge from distant nodes, we propose a Dense Graph Propagation (DGP) module with carefully designed direct links among distant nodes. DGP allows us to exploit the hierarchical graph structure of the knowledge graph through additional connections. These connections are added based on a node's relationship to its ancestors and descendants. A weighting scheme is further used to weigh their contribution depending on the distance to the node to improve information propagation in the graph. Combined with finetuning of the representations in a two-stage training approach our method outperforms state-of-the-art zero-shot learning approaches.

Crowd counting is a challenging task due to factors such as large variations in crowdedness and severe occlusions. Although recent deep learning based counting algorithms have achieved a great progress, the correlation knowledge among samples and the semantic prior have not yet been fully exploited. In this paper, a residual regression framework is proposed for crowd counting utilizing the correlation information among samples. By incorporating such information into our network, we discover that more intrinsic characteristics can be learned by the network which thus generalizes better to unseen scenarios. Besides, we show how to effectively leverage the semantic prior to improve the performance of crowd counting. We also observe that the adversarial loss can be used to improve the quality of predicted density maps, thus leading to an improvement in crowd counting. Experiments on public datasets demonstrate the effectiveness and generalization ability of the proposed method.

Light field cameras are considered to have many potential applications since angular and spatial information is captured simultaneously. However, the limited spatial resolution has brought lots of difficulties in developing related applications and becomes the main bottleneck of light field cameras. In this paper, a learning-based method using residual convolutional networks is proposed to reconstruct light fields with higher spatial resolution. The view images in one light field are first grouped into different image stacks with consistent sub-pixel offsets and fed into different network branches to implicitly learn inherent corresponding relations. The residual information in different spatial directions is then calculated from each branch and further integrated to supplement high-frequency details for the view image. Finally, a flexible solution is proposed to super-resolve entire light field images with various angular resolutions. Experimental results on synthetic and real-world datasets demonstrate that the proposed method outperforms other state-of-the-art methods by a large margin in both visual and numerical evaluations. Furthermore, the proposed method shows good performances in preserving the inherent epipolar property in light field images.

Transfer learning is widely used in deep neural network models when there are few labeled examples available. The common approach is to take a pre-trained network in a similar task and finetune the model parameters. This is usually done blindly without a pre-selection from a set of pre-trained models, or by finetuning a set of models trained on different tasks and selecting the best performing one by cross-validation. We address this problem by proposing an approach to assess the relationship between visual tasks and their task-specific models. Our method uses Representation Similarity Analysis (RSA), which is commonly used to find a correlation between neuronal responses from brain data and models. With RSA we obtain a similarity score among tasks by computing correlations between models trained on different tasks. Our method is efficient as it requires only pre-trained models, and a few images with no further training. We demonstrate the effectiveness and efficiency of our method to generating task taxonomy on Taskonomy dataset. We next evaluate the relationship of RSA with the transfer learning performance on Taskonomy tasks and a new task: Pascal VOC semantic segmentation. Our results reveal that models trained on tasks with higher similarity score show higher transfer learning performance. Surprisingly, the best transfer learning result for Pascal VOC semantic segmentation is not obtained from the pre-trained model on semantic segmentation, probably due to the domain differences, and our method successfully selects the high performing models.

In this paper, we propose a convolutional layer inspired by optical flow algorithms to learn motion representations. Our representation flow layer is a fully-differentiable layer designed to capture the `flow' of any representation channel within a convolutional neural network for action recognition. Its parameters for iterative flow optimization are learned in an end-to-end fashion together with the other CNN model parameters, maximizing the action recognition performance. Furthermore, we newly introduce the concept of learning `flow of flow' representations by stacking multiple representation flow layers. We conducted extensive experimental evaluations, confirming its advantages over previous recognition models using traditional optical flows in both computational speed and performance. The code is publicly available.

This paper addresses the problem of 3D human pose estimation from single images. While for a long time human skeletons were parameterized and fitted to the observation by satisfying a reprojection error, nowadays researchers directly use neural networks to infer the 3D pose from the observations. However, most of these approaches ignore the fact that a reprojection constraint has to be satisfied and are sensitive to overfitting. We tackle the overfitting problem by ignoring 2D to 3D correspondences. This efficiently avoids a simple memorization of the training data and allows for a weakly supervised training. One part of the proposed reprojection network (RepNet) learns a mapping from a distribution of 2D poses to a distribution of 3D poses using an adversarial training approach. Another part of the network estimates the camera. This allows for the definition of a network layer that performs the reprojection of the estimated 3D pose back to 2D which results in a reprojection loss function. Our experiments show that RepNet generalizes well to unknown data and outperforms state-of-the-art methods when applied to unseen data. Moreover, our implementation runs in real-time on a standard desktop PC.

Distance metric learning (DML) has been successfully applied to object classification, both in the standard regime of rich training data and in the few-shot scenario, where each category is represented by only a few examples. In this work, we propose a new method for DML that simultaneously learns the backbone network parameters, the embedding space, and the multi-modal distribution of each of the training categories in that space, in a single end-to-end training process. Our approach outperforms state-of-the-art methods for DML-based object classification on a variety of standard fine-grained datasets. Furthermore, we demonstrate the effectiveness of our approach on the problem of few-shot object detection, by incorporating the proposed DML architecture as a classification head into a standard object detection model. We achieve the best results on the ImageNet-LOC dataset compared to strong baselines, when only a few training examples are available. We also offer the community a new episodic benchmark based on the ImageNet dataset for the few-shot object detection task.

Image cropping aims to improve the composition as well as aesthetic quality of an image by removing extraneous content from it. Existing image cropping databases provide only one or several human-annotated bounding boxes as the groundtruth, which cannot reflect the non-uniqueness and flexibility of image cropping in practice. The employed evaluation metrics such as intersection-over-union cannot reliably reflect the real performance of cropping models, either. This work revisits the problem of image cropping, and presents a grid anchor based formulation by considering the special properties and requirements (e.g., local redundancy, content preservation, aspect ratio) of image cropping. Our formulation reduces the searching space of candidate crops from millions to less than one hundred. Consequently, a grid anchor based cropping benchmark is constructed, where all crops of each image are annotated and more reliable evaluation metrics are defined. We also design an effective and lightweight network module, which simultaneously considers the region of interest and region of discard for more accurate image cropping. Our model can stably output visually pleasing crops for images of different scenes and run at a speed of 125 FPS.

Knowledge distillation aims at transferring knowledge acquired in one model (a teacher) to another model (a student) that is typically smaller. Previous approaches can be expressed as a form of training the student to mimic output activations of individual data examples represented by the teacher. We introduce a novel approach, dubbed relational knowledge distillation (RKD), that transfers mutual relations of data examples instead. For concrete realizations of RKD, we propose distance-wise and angle-wise distillation losses that penalize structural differences in relations. Experiments conducted on different tasks show that the proposed method improves educated student models with a significant margin. In particular for metric learning, it allows students to outperform their teachers' performance, achieving the state of the arts on standard benchmark datasets.

This paper focuses on multi-person action forecasting in videos. More precisely, given a history of H previous frames, the goal is to detect actors and to predict their future actions for the next T frames. Our approach jointly models temporal and spatial interactions among different actors by constructing a recurrent graph, using actor proposals obtained with Faster R-CNN as nodes. Our method learns to select a subset of discriminative relations without requiring explicit supervision, thus enabling us to tackle challenging visual data. We refer to our model as Discriminative Relational Recurrent Network (DRRN). Evaluation of action prediction on AVA demonstrates the effectiveness of our proposed method compared to simpler baselines. Furthermore, we significantly improve performance on the task of early action classification on J-HMDB, from the previous SOTA of 48% to 60%.

Point cloud analysis is very challenging, as the shape implied in irregular points is difficult to capture. In this paper, we propose RS-CNN, namely, Relation-Shape Convolutional Neural Network, which extends regular grid CNN to irregular configuration for point cloud analysis. The key to RS-CNN is learning from relation, i.e., the geometric topology constraint among points. Specifically, the convolutional weight for local point set is forced to learn a high-level relation expression from predefined geometric priors, between a sampled point from this point set and the others. In this way, an inductive local representation with explicit reasoning about the spatial layout of points can be obtained, which leads to much shape awareness and robustness. With this convolution as a basic operator, RS-CNN, a hierarchical architecture can be developed to achieve contextual shape-aware learning for point cloud analysis. Extensive experiments on challenging benchmarks across three tasks verify RS-CNN achieves the state of the arts.