CVPR 2020 论文列表

For person re-identification (re-id), attention mechanisms have become attractive as they aim at strengthening discriminative features and suppressing irrelevant ones, which matches well the key of re-id, i.e., discriminative feature learning. Previous approaches typically learn attention using local convolutions, ignoring the mining of knowledge from global structure patterns. Intuitively, the affinities among spatial positions/nodes in the feature map provide clustering-like information and are helpful for inferring semantics and thus attention, especially for person images where the feasible human poses are constrained. In this work, we propose an effective Relation-Aware Global Attention (RGA) module which captures the global structural information for better attention learning. Specifically, for each feature position, in order to compactly grasp the structural information of global scope and local appearance information, we propose to stack the relations, i.e., its pairwise correlations/affinities with all the feature positions (e.g., in raster scan order), and the feature itself together to learn the attention with a shallow convolutional model. Extensive ablation studies demonstrate that our RGA can significantly enhance the feature representation power and help achieve the state-of-the-art performance on several popular benchmarks. The source code is available at https://github.com/microsoft/Relation-Aware-Global-Attention-Networks.

Inspired by the Thomson problem in physics where the distribution of multiple propelling electrons on a unit sphere can be modeled via minimizing some potential energy, hyperspherical energy minimization has demonstrated its potential in regularizing neural networks and improving their generalization power. In this paper, we first study the important role that hyperspherical energy plays in neural network training by analyzing its training dynamics. Then we show that naively minimizing hyperspherical energy suffers from some difficulties due to highly non-linear and non-convex optimization as the space dimensionality becomes higher, therefore limiting the potential to further improve the generalization. To address these problems, we propose the compressive minimum hyperspherical energy (CoMHE) as a more effective regularization for neural networks. Specifically, CoMHE utilizes projection mappings to reduce the dimensionality of neurons and minimizes their hyperspherical energy. According to different designs for the projection mapping, we propose several distinct yet well-performing variants and provide some theoretical guarantees to justify their effectiveness. Our experiments show that CoMHE consistently outperforms existing regularization methods, and can be easily applied to different neural networks.

Semi-supervised generative learning aims to learn the underlying class-conditional distribution of partially labeled data. Generative Adversarial Networks (GANs) have led to promising progress in this task. However, it still needs to further explore the issue of imbalance between real labeled data and fake data in the adversarial learning process. To address this issue, we propose a regularization technique based on Random Regional Replacement (R^3-regularization) to facilitate the generative learning process. Specifically, we construct two types of between-class instances: cross-category ones and real-fake ones. These instances could be closer to the decision boundaries and are important for regularizing the classification and discriminative networks in our class-conditional GANs, which we refer to as R^3-CGAN. Better guidance from these two networks makes the generative network produce instances with class-specific information and high fidelity. We experiment with multiple standard benchmarks, and demonstrate that the R^3-regularization can lead to significant improvement in both classification and class-conditional image synthesis.

Deep neural networks with millions of parameters may suffer from poor generalization due to overfitting. To mitigate the issue, we propose a new regularization method that penalizes the predictive distribution between similar samples. In particular, we distill the predictive distribution between different samples of the same label during training. This results in regularizing the dark knowledge (i.e., the knowledge on wrong predictions) of a single network (i.e., a self-knowledge distillation) by forcing it to produce more meaningful and consistent predictions in a class-wise manner. Consequently, it mitigates overconfident predictions and reduces intra-class variations. Our experimental results on various image classification tasks demonstrate that the simple yet powerful method can significantly improve not only the generalization ability but also the calibration performance of modern convolutional neural networks.

This paper is about regularizing deep convolutional networks (CNNs) based on an adaptive framework for transfer learning with limited training data in the target domain. Recent advances of CNN regularization in this context are commonly due to the use of additional regularization objectives. They guide the training away from the target task using some forms of concrete tasks. Unlike those related approaches, we suggest that an objective without a concrete goal can still serve well as a regularizer. In particular, we demonstrate Pseudo-task Regularization (PtR) which dynamically regularizes a network by simply attempting to regress image representations to pseudo-regression targets during fine-tuning. That is, a CNN is efficiently regularized without additional resources of data or prior domain expertise. In sum, the proposed PtR provides: a) an alternative for network regularization without dependence on the design of concrete regularization objectives or extra annotations; b) a dynamically adjusted and maintained strength of regularization effect by balancing the gradient norms between objectives on-line. Through numerous experiments, surprisingly, the improvements on classification accuracy by PtR are shown greater or on a par to the recent state-of-the-art methods.

Deep neural networks for video action recognition frequently require 3D convolutional filters and often encounter overfitting due to a larger number of parameters. In this paper, we propose Random Mean Scaling (RMS), a simple and effective regularization method, to relieve the overfitting problem in 3D residual networks. The key idea of RMS is to randomly vary the magnitude of low-frequency components of the feature to regularize the model. The low-frequency component can be derived by a spatio-temporal mean on the local patch of a feature. We present that selective regularization on this locally smoothed feature makes a model handle the low-frequency and high-frequency component distinctively, resulting in performance improvement. RMS can enhance a model with little additional computation only during training, similar to other regularization methods. RMS also can be incorporated into typical training process without any bells and whistles. Experimental results show the improvement in generalization performance on a popular action recognition datasets demonstrating the effectiveness of RMS as a regularization technique, compared to other state-of-the-art regularization methods.

For images taken through glass, existing methods focus on the restoration of the background scene by regarding the reflection components as noise. However, the scene reflected by glass surface also contains important information to be recovered, especially for the surveillance or criminal investigations. In this paper, instead of removing reflection components from the mixture image, we aim at recovering reflection scenes from the mixture image. We first propose a strategy to obtain such ground truth and its corresponding input images. Then, we propose a two-stage framework to obtain the visible reflection scene from the mixture image. Specifically, we train the network with a shift-invariant loss which is robust to misalignment between the input and output images. The experimental results show that our proposed method achieves promising results.

Referring image segmentation aims at segmenting the foreground masks of the entities that can well match the description given in the natural language expression. Previous approaches tackle this problem using implicit feature interaction and fusion between visual and linguistic modalities, but usually fail to explore informative words of the expression to well align features from the two modalities for accurately identifying the referred entity. In this paper, we propose a Cross-Modal Progressive Comprehension (CMPC) module and a Text-Guided Feature Exchange (TGFE) module to effectively address the challenging task. Concretely, the CMPC module first employs entity and attribute words to perceive all the related entities that might be considered by the expression. Then, the relational words are adopted to highlight the correct entity as well as suppress other irrelevant ones by multimodal graph reasoning. In addition to the CMPC module, we further leverage a simple yet effective TGFE module to integrate the reasoned multimodal features from different levels with the guidance of textual information. In this way, features from multi-levels could communicate with each other and be refined based on the textual context. We conduct extensive experiments on four popular referring segmentation benchmarks and achieve new state-of-the-art performances. Code is available at https://github.com/spyflying/CMPC-Refseg.

This paper tackles the automatic colorization task of a sketch image given an already-colored reference image. Colorizing a sketch image is in high demand in comics, animation, and other content creation applications, but it suffers from information scarcity of a sketch image. To address this, a reference image can render the colorization process in a reliable and user-driven manner. However, it is difficult to prepare for a training data set that has a sufficient amount of semantically meaningful pairs of images as well as the ground truth for a colored image reflecting a given reference (e.g., coloring a sketch of an originally blue car given a reference green car). To tackle this challenge, we propose to utilize the identical image with geometric distortion as a virtual reference, which makes it possible to secure the ground truth for a colored output image. Furthermore, it naturally provides the ground truth for dense semantic correspondence, which we utilize in our internal attention mechanism for color transfer from reference to sketch input. We demonstrate the effectiveness of our approach in various types of sketch image colorization via quantitative as well as qualitative evaluation against existing methods.

Social interaction is an important topic in human trajectory prediction to generate plausible paths. In this paper, we present a novel insight of group-based social interaction model to explore relationships among pedestrians. We recursively extract social representations supervised by group-based annotations and formulate them into a social behavior graph, called Recursive Social Behavior Graph. Our recursive mechanism explores the representation power largely. Graph Convolutional Neural Network then is used to propagate social interaction information in such a graph. With the guidance of Recursive Social Behavior Graph, we surpass state-of-the-art methods on ETH and UCY dataset for 11.1% in ADE and 10.8% in FDE in average, and successfully predict complex social behaviors.

Existing inpainting methods have achieved promising performance for recovering regular or small image defects. However, filling in large continuous holes remains difficult due to the lack of constraints for the hole center. In this paper, we devise a Recurrent Feature Reasoning (RFR) network which is mainly constructed by a plug-and-play Recurrent Feature Reasoning module and a Knowledge Consistent Attention (KCA) module. Analogous to how humans solve puzzles (i.e., first solve the easier parts and then use the results as additional information to solve difficult parts), the RFR module recurrently infers the hole boundaries of the convolutional feature maps and then uses them as clues for further inference. The module progressively strengthens the constraints for the hole center and the results become explicit. To capture information from distant places in the feature map for RFR, we further develop KCA and incorporate it in RFR. Empirically, we first compare the proposed RFR-Net with existing backbones, demonstrating that RFR-Net is more efficient (e.g., a 4% SSIM improvement for the same model size). We then place the network in the context of the current state-of-the-art, where it exhibits improved performance. The corresponding source code is available at: https://github.com/jingyuanli001/RFR-Inpainting

Six degree-of-freedom pose estimation of a known object in a single image is a long-standing computer vision objective. It is classically posed as a correspondence problem between a known geometric model, such as a CAD model, and image locations. If a CAD model is not available, it is possible to use multi-view visual reconstruction methods to create a geometric model, and use this in the same manner. Instead, we propose a learning-based method whose input is a collection of images of a target object, and whose output is the pose of the object in a novel view. At inference time, our method maps from the RoI features of the input image to a dense collection of object-centric 3D coordinates, one per pixel. This dense 2D-3D mapping is then used to determine 6dof pose using standard PnP plus RANSAC. The model that maps 2D to object 3D coordinates is established at training time by automatically discovering and matching image landmarks that are consistent across multiple views. We show that this method eliminates the requirement for a 3D CAD model (needed by classical geometry-based methods and state-of-the-art learning-based methods alike) but still achieves performance on a par with the prior art.

In this paper, we tackle an important task in computer vision: any view object recognition. In both training and testing, for each object instance, we are only given its 2D image viewed from an unknown angle. We propose a computational framework by designing object and viewer-centered neural networks (OVCNet) to recognize an object instance viewed from an arbitrary unknown angle. OVCNet consists of three branches that respectively implement object-centered, 3D viewer-centered, and in-plane viewer-centered recognition. We evaluate our proposed OVCNet using two metrics with unseen views from both seen and novel object instances. Experimental results demonstrate the advantages of OVCNet over classic 2D-image-based CNN classifiers, 3D-object (inferred from 2D image) classifiers, and competing multi-view based approaches. It gives rise to a viable and practical computing framework that combines both viewpoint-dependent and viewpoint-independent features for object recognition from any view.

Person re-identification (Re-ID) in real-world scenarios usually suffers from various degradation factors, e.g., low-resolution, weak illumination, blurring and adverse weather. On the one hand, these degradations lead to severe discriminative information loss, which significantly obstructs identity representation learning; on the other hand, the feature mismatch problem caused by low-level visual variations greatly reduces retrieval performance. An intuitive solution to this problem is to utilize low-level image restoration methods to improve the image quality. However, existing restoration methods cannot directly serve to real-world Re-ID due to various limitations, e.g., the requirements of reference samples, domain gap between synthesis and reality, and incompatibility between low-level and high-level methods. In this paper, to solve the above problem, we propose a degradation invariance learning framework for real-world person Re-ID. By introducing a self-supervised disentangled representation learning strategy, our method is able to simultaneously extract identity-related robust features and remove real-world degradations without extra supervision. We use low-resolution images as the main demonstration, and experiments show that our approach is able to achieve state-of-the-art performance on several Re-ID benchmarks. In addition, our framework can be easily extended to other real-world degradation factors, such as weak illumination, with only a few modifications.

We introduce an information-theoretic loss function, RankMI, and an associated training algorithm for deep representation learning for image retrieval. Our proposed framework consists of alternating updates to a network that estimates the divergence between distance distributions of matching and non-matching pairs of learned embeddings, and an embedding network that maximizes this estimate via sampled negatives. In addition, under this information-theoretic lens we draw connections between RankMI and commonly-used ranking losses, e.g., triplet loss. We extensively evaluate RankMI on several standard image retrieval datasets, namely, CUB-200-2011, CARS-196, and Stanford Online Products. Our method achieves competitive results or significant improvements over previous reported results on all datasets.

Iterative Closest Point (ICP) solves the rigid point cloud registration problem iteratively in two steps: (1) make hard assignments of spatially closest point correspondences, and then (2) find the least-squares rigid transformation. The hard assignments of closest point correspondences based on spatial distances are sensitive to the initial rigid transformation and noisy/outlier points, which often cause ICP to converge to wrong local minima. In this paper, we propose the RPM-Net -- a less sensitive to initialization and more robust deep learning-based approach for rigid point cloud registration. To this end, our network uses the differentiable Sinkhorn layer and annealing to get soft assignments of point correspondences from hybrid features learned from both spatial coordinates and local geometry. To further improve registration performance, we introduce a secondary network to predict optimal annealing parameters. Unlike some existing methods, our RPM-Net handles missing correspondences and point clouds with partial visibility. Experimental results show that our RPM-Net achieves state-of-the-art performance compared to existing non-deep learning and recent deep learning methods. Our source code is available at the project website (https://github.com/yewzijian/RPMNet).

In this paper, we design a tracking model consisting of response generation and bounding box regression, where the first component produces a heat map to indicate the presence of the object at different positions and the second part regresses the relative bounding box shifts to anchors mounted on sliding-window locations. Thanks to the resizable convolutional filters used in both components to adapt to the shape changes of objects, our tracking model does not need to enumerate different sized anchors, thus saving model parameters. To effectively adapt the model to appearance variations, we propose to offline train a recurrent neural optimizer to update tracking model in a meta-learning setting, which can converge the model in a few gradient steps. This improves the convergence speed of updating the tracking model while achieving better performance. We extensively evaluate our trackers, ROAM and ROAM++, on the OTB, VOT, LaSOT, GOT-10K and TrackingNet benchmark and our methods perform favorably against state-of-the-art algorithms.

Spiking Neural Networks (SNNs) have recently attracted significant research interest as the third generation of artificial neural networks that can enable low-power event-driven data analytics. The best performing SNNs for image recognition tasks are obtained by converting a trained Analog Neural Network (ANN), consisting of Rectified Linear Units (ReLU), to SNN composed of integrate-and-fire neurons with "proper" firing thresholds. The converted SNNs typically incur loss in accuracy compared to that provided by the original ANN and require sizable number of inference time-steps to achieve the best accuracy. We find that performance degradation in the converted SNN stems from using "hard reset" spiking neuron that is driven to fixed reset potential once its membrane potential exceeds the firing threshold, leading to information loss during SNN inference. We propose ANN-SNN conversion using "soft reset" spiking neuron model, referred to as Residual Membrane Potential (RMP) spiking neuron, which retains the "residual" membrane potential above threshold at the firing instants. We demonstrate near loss-less ANN-SNN conversion using RMP neurons for VGG-16, ResNet-20, and ResNet-34 SNNs on challenging datasets including CIFAR-10 (93.63% top-1), CIFAR-100 (70.93% top-1), and ImageNet (73.09% top-1 accuracy). Our results also show that RMP-SNN surpasses the best inference accuracy provided by the converted SNN with "hard reset" spiking neurons using 2-8 times fewer inference time-steps across network architectures and datasets.

The automatic extraction of animal 3D pose from images without markers is of interest in a range of scientific fields. Most work to date predicts animal pose from RGB images, based on 2D labelling of joint positions. However, due to the difficult nature of obtaining training data, no ground truth dataset of 3D animal motion is available to quantitatively evaluate these approaches. In addition, a lack of 3D animal pose data also makes it difficult to train 3D pose-prediction methods in a similar manner to the popular field of body-pose prediction. In our work, we focus on the problem of 3D canine pose estimation from RGBD images, recording a diverse range of dog breeds with several Microsoft Kinect v2s, simultaneously obtaining the 3D ground truth skeleton via a motion capture system. We generate a dataset of synthetic RGBD images from this data. A stacked hourglass network is trained to predict 3D joint locations, which is then constrained using prior models of shape and pose. We evaluate our model on both synthetic and real RGBD images and compare our results to previously published work fitting canine models to images. Finally, despite our training set consisting only of dog data, visual inspection implies that our network can produce good predictions for images of other quadrupeds - e.g. horses or cats - when their pose is similar to that contained in our training set.