CVPR 2020 论文列表

Referring expression comprehension (REF) aims at identifying a particular object in a scene by a natural language expression. It requires joint reasoning over the textual and visual domains to solve the problem. Some popular referring expression datasets, however, fail to provide an ideal test bed for evaluating the reasoning ability of the models, mainly because 1) their expressions typically describe only some simple distinctive properties of the object and 2) their images contain limited distracting information. To bridge the gap, we propose a new dataset for visual reasoning in context of referring expression comprehension with two main features. First, we design a novel expression engine rendering various reasoning logics that can be flexibly combined with rich visual properties to generate expressions with varying compositionality. Second, to better exploit the full reasoning chain embodied in an expression, we propose a new test setting by adding additional distracting images containing objects sharing similar properties with the referent, thus minimising the success rate of reasoning-free cross-domain alignment. We evaluate several state-of-the-art REF models, but find none of them can achieve promising performance. A proposed modular hard mining strategy performs the best but still leaves substantial room for improvement.

Adversarial attack of CNN aims at deceiving models to misbehave by adding imperceptible perturbations to images. This feature facilitates to understand neural networks deeply and to improve the robustness of deep learning models. Although several works have focused on attacking image classifiers and object detectors, an effective and efficient method for attacking single object trackers of any target in a model-free way remains lacking. In this paper, a cooling-shrinking attack method is proposed to deceive state-of-the-art SiameseRPN-based trackers. An effective and efficient perturbation generator is trained with a carefully designed adversarial loss, which can simultaneously cool hot regions where the target exists on the heatmaps and force the predicted bounding box to shrink, making the tracked target invisible to trackers. Numerous experiments on OTB100, VOT2018, and LaSOT datasets show that our method can effectively fool the state-of-the-art SiameseRPN++ tracker by adding small perturbations to the template or the search regions. Besides, our method has good transferability and is able to deceive other top-performance trackers such as DaSiamRPN, DaSiamRPN-UpdateNet, and DiMP. The source codes are available at https://github.com/MasterBin-IIAU/CSA.

This paper addresses the problem of text-to-image synthesis from a new perspective, i.e., the cause-and-effect chain in image generation. Causality is a common phenomenon in cooking. The dish appearance changes depending on the cooking actions and ingredients. The challenge of synthesis is that a generated image should depict the visual result of action-on-object. This paper presents a new network architecture, CookGAN, that mimics visual effect in causality chain, preserves fine-grained details and progressively upsamples image. Particularly, a cooking simulator sub-network is proposed to incrementally make changes to food images based on the interaction between ingredients and cooking methods over a series of steps. Experiments on Recipe1M verify that CookGAN manages to generate food images with reasonably impressive inception score. Furthermore, the images are semantically interpretable and manipulable.

Point clouds are among the popular geometry representations for 3D vision applications. However, without regular structures like 2D images, processing and summarizing information over these unordered data points are very challenging. Although a number of previous works attempt to analyze point clouds and achieve promising performances, their performances would degrade significantly when data variations like shift and scale changes are presented. In this paper, we propose 3D Graph Convolution Networks (3D-GCN), which is designed to extract local 3D features from point clouds across scales, while shift and scale-invariance properties are introduced. The novelty of our 3D-GCN lies in the definition of learnable kernels with a graph max-pooling mechanism. We show that 3D-GCN can be applied to 3D classification and segmentation tasks, with ablation studies and visualizations verifying the design of 3D-GCN.

This paper proposes a novel message passing neural (MPN) architecture Conv-MPN, which reconstructs an outdoor building as a planar graph from a single RGB image. Conv-MPN is specifically designed for cases where nodes of a graph have explicit spatial embedding. In our problem, nodes correspond to building edges in an image. Conv-MPN is different from MPN in that 1) the feature associated with a node is represented as a feature volume instead of a 1D vector; and 2) convolutions encode messages instead of fully connected layers. Conv-MPN learns to select a true subset of nodes (i.e., building edges) to reconstruct a building planar graph. Our qualitative and quantitative evaluations over 2,000 buildings show that Conv-MPN makes significant improvements over the existing fully neural solutions. We believe that the paper has a potential to open a new line of graph neural network research for structured geometry reconstruction.

This paper introduces the Attribute-Decomposed GAN, a novel generative model for controllable person image synthesis, which can produce realistic person images with desired human attributes (e.g., pose, head, upper clothes and pants) provided in various source inputs. The core idea of the proposed model is to embed human attributes into the latent space as independent codes and thus achieve flexible and continuous control of attributes via mixing and interpolation operations in explicit style representations. Specifically, a new architecture consisting of two encoding pathways with style block connections is proposed to decompose the original hard mapping into multiple more accessible subtasks. In source pathway, we further extract component layouts with an off-the-shelf human parser and feed them into a shared global texture encoder for decomposed latent codes. This strategy allows for the synthesis of more realistic output images and automatic separation of un-annotated attributes. Experimental results demonstrate the proposed method's superiority over the state of the art in pose transfer and its effectiveness in the brand-new task of component attribute transfer.

Orthogonality is widely used for training deep neural networks (DNNs) due to its ability to maintain all singular values of the Jacobian close to 1 and reduce redundancy in representation. This paper proposes a computationally efficient and numerically stable orthogonalization method using Newton's iteration (ONI), to learn a layer-wise orthogonal weight matrix in DNNs. ONI works by iteratively stretching the singular values of a weight matrix towards 1. This property enables it to control the orthogonality of a weight matrix by its number of iterations. We show that our method improves the performance of image classification networks by effectively controlling the orthogonality to provide an optimal tradeoff between optimization benefits and representational capacity reduction. We also show that ONI stabilizes the training of generative adversarial networks (GANs) by maintaining the Lipschitz continuity of a network, similar to spectral normalization (SN), and further outperforms SN by providing controllable orthogonality.

Scene text detection has witnessed rapid development in recent years. However, there still exists two main challenges: 1) many methods suffer from false positives in their text representations; 2) the large scale variance of scene texts makes it hard for network to learn samples. In this paper, we propose the ContourNet, which effectively handles these two problems taking a further step toward accurate arbitrary-shaped text detection. At first, a scale-insensitive Adaptive Region Proposal Network (Adaptive-RPN) is proposed to generate text proposals by only focusing on the Intersection over Union (IoU) values between predicted and ground-truth bounding boxes. Then a novel Local Orthogonal Texture-aware Module (LOTM) models the local texture information of proposal features in two orthogonal directions and represents text region with a set of contour points. Considering that the strong unidirectional or weakly orthogonal activation is usually caused by the monotonous texture characteristic of false-positive patterns (e.g. streaks.), our method effectively suppresses these false positives by only outputting predictions with high response value in both orthogonal directions. This gives more accurate description of text regions. Extensive experiments on three challenging datasets (Total-Text, CTW1500 and ICDAR2015) verify that our method achieves the state-of-the-art performance. Code is available at https://github.com/wangyuxin87/ContourNet.

We propose a quadratic penalty method for continual learning of neural networks that contain batch normalization (BN) layers. The Hessian of a loss function represents the curvature of the quadratic penalty function, and a Kronecker-factored approximate curvature (K-FAC) is used widely to practically compute the Hessian of a neural network. However, the approximation is not valid if there is dependence between examples, typically caused by BN layers in deep network architectures. We extend the K-FAC method so that the inter-example relations are taken into account and the Hessian of deep neural networks can be properly approximated under practical assumptions. We also propose a method of weight merging and reparameterization to properly handle statistical parameters of BN, which plays a critical role for continual learning with BN, and a method that selects hyperparameters without source task data. Our method shows better performance than baselines in the permuted MNIST task with BN layers and in sequential learning from the ImageNet classification task to fine-grained classification tasks with ResNet-50, without any explicit or implicit use of source task data for hyperparameter selection.

Recently data-driven image inpainting methods have made inspiring progress, impacting fundamental image editing tasks such as object removal and damaged image repairing. These methods are more effective than classic approaches, however, due to memory limitations they can only handle low-resolution inputs, typically smaller than 1K. Meanwhile, the resolution of photos captured with mobile devices increases up to 8K. Naive up-sampling of the low-resolution inpainted result can merely yield a large yet blurry result. Whereas, adding a high-frequency residual image onto the large blurry image can generate a sharp result, rich in details and textures. Motivated by this, we propose a Contextual Residual Aggregation (CRA) mechanism that can produce high-frequency residuals for missing contents by weighted aggregating residuals from contextual patches, thus only requiring a low-resolution prediction from the network. Since convolutional layers of the neural network only need to operate on low-resolution inputs and outputs, the cost of memory and computing power is thus well suppressed. Moreover, the need for high-resolution training datasets is alleviated. In our experiments, we train the proposed model on small images with resolutions 512 x 512 and perform inference on high-resolution images, achieving compelling inpainting quality. Our model can inpaint images as large as 8K with considerable hole sizes, which is intractable with previous learning-based approaches. We further elaborate on the light-weight design of the network architecture, achieving real-time performance on 2K images on a GTX 1080 Ti GPU. Codes are available at: https://github. com/Ascend-Huawei/Ascend-Canada/tree/ master/Models/Research_HiFIll_Model

Temporal repetition counting aims to estimate the number of cycles of a given repetitive action. Existing deep learning methods assume repetitive actions are performed in a fixed time-scale, which is invalid for the complex repetitive actions in real life. In this paper, we tailor a context-aware and scale-insensitive framework, to tackle the challenges in repetition counting caused by the unknown and diverse cycle-lengths. Our approach combines two key insights: (1) Cycle lengths from different actions are unpredictable that require large-scale searching, but, once a coarse cycle length is determined, the variety between repetitions can be overcome by regression. (2) Determining the cycle length cannot only rely on a short fragment of video but a contextual understanding. The first point is implemented by a coarse-to-fine cycle refinement method. It avoids the heavy computation of exhaustively searching all the cycle lengths in the video, and, instead, it propagates the coarse prediction for further refinement in a hierarchical manner. We secondly propose a bidirectional cycle length estimation method for a context-aware prediction. It is a regression network that takes two consecutive coarse cycles as input, and predicts the locations of the previous and next repetitive cycles. To benefit the training and evaluation of temporal repetition counting area, we construct a new and largest benchmark, which contains 526 videos with diverse repetitive actions. Extensive experiments show that the proposed network trained on a single dataset outperforms state-of-the-art methods on several benchmarks, indicating that the proposed framework is general enough to capture repetition patterns across domains. Code and data are available in https://github.com/Xiaodomgdomg/Deep-Temporal-Repetition-Counting.

The problem of predicting human motion given a sequence of past observations is at the core of many applications in robotics and computer vision. Current state-of-the-art formulates this problem as a sequence-to-sequence task, in which a historical of 3D skeletons feeds a Recurrent Neural Network (RNN) that predicts future movements, typically in the order of 1 to 2 seconds. However, one aspect that has been obviated so far, is the fact that human motion is inherently driven by interactions with objects and/or other humans in the environment. In this paper, we explore this scenario using a novel context-aware motion prediction architecture. We use a semantic-graph model where the nodes parameterize the human and objects in the scene and the edges their mutual interactions. These interactions are iteratively learned through a graph attention layer, fed with the past observations, which now include both object and human body motions. Once this semantic graph is learned, we inject it to a standard RNN to predict future movements of the human/s and object/s. We consider two variants of our architecture, either freezing the contextual interactions in the future of updating them. A thorough evaluation in the Whole-Body Human Motion Database shows that in both cases, our context-aware networks clearly outperform baselines in which the context information is not considered.

While image captioning has progressed rapidly, existing works focus mainly on describing single images. In this paper, we introduce a new task, context-aware group captioning, which aims to describe a group of target images in the context of another group of related reference images. Context-aware group captioning requires not only summarizing information from both the target and reference image group but also contrasting between them. To solve this problem, we propose a framework combining self-attention mechanism with contrastive feature construction to effectively summarize common information from each image group while capturing discriminative information between them. To build the dataset for this task, we propose to group the images and generate the group captions based on single image captions using scene graphs matching. Our datasets are constructed on top of the public Conceptual Captions dataset and our new Stock Captions dataset. Experiments on the two datasets show the effectiveness of our method on this new task.

As a typical cross-modal problem, image-text bi-directional retrieval relies heavily on the joint embedding learning and similarity measure for each image-text pair. It remains challenging because prior works seldom explore semantic correspondences between modalities and semantic correlations in a single modality at the same time. In this work, we propose a unified Context-Aware Attention Network (CAAN), which selectively focuses on critical local fragments (regions and words) by aggregating the global context. Specifically, it simultaneously utilizes global inter-modal alignments and intra-modal correlations to discover latent semantic relations. Considering the interactions between images and sentences in the retrieval process, intra-modal correlations are derived from the second-order attention of region-word alignments instead of intuitively comparing the distance between original features. Our method achieves fairly competitive results on two generic image-text retrieval datasets Flickr30K and MS-COCO.

In static monitoring cameras, useful contextual information can stretch far beyond the few seconds typical video understanding models might see: subjects may exhibit similar behavior over multiple days, and background objects remain static. Due to power and storage constraints, sampling frequencies are low, often no faster than one frame per second, and sometimes are irregular due to the use of a motion trigger. In order to perform well in this setting, models must be robust to irregular sampling rates. In this paper we propose a method that leverages temporal context from the unlabeled frames of a novel camera to improve performance at that camera. Specifically, we propose an attention-based approach that allows our model, Context R-CNN, to index into a long term memory bank constructed on a per-camera basis and aggregate contextual features from other frames to boost object detection performance on the current frame. We apply Context R-CNN to two settings: (1) species detection using camera traps, and (2) vehicle detection in traffic cameras, showing in both settings that Context R-CNN leads to performance gains over strong baselines. Moreover, we show that increasing the contextual time horizon leads to improved results. When applied to camera trap data from the Snapshot Serengeti dataset, Context R-CNN with context from up to a month of images outperforms a single-frame baseline by 17.9% mAP, and outperforms S3D (a 3d convolution based baseline) by 11.2% mAP.

Recent works have widely explored the contextual dependencies to achieve more accurate segmentation results. However, most approaches rarely distinguish different types of contextual dependencies, which may pollute the scene understanding. In this work, we directly supervise the feature aggregation to distinguish the intra-class and interclass context clearly. Specifically, we develop a Context Prior with the supervision of the Affinity Loss. Given an input image and corresponding ground truth, Affinity Loss constructs an ideal affinity map to supervise the learning of Context Prior. The learned Context Prior extracts the pixels belonging to the same category, while the reversed prior focuses on the pixels of different classes. Embedded into a conventional deep CNN, the proposed Context Prior Layer can selectively capture the intra-class and inter-class contextual dependencies, leading to robust feature representation. To validate the effectiveness, we design an effective Context Prior Network (CPNet). Extensive quantitative and qualitative evaluations demonstrate that the proposed model performs favorably against state-of-the-art semantic segmentation approaches. More specifically, our algorithm achieves 46.3% mIoU on ADE20K, 53.9% mIoU on PASCAL-Context, and 81.3% mIoU on Cityscapes. Code is available at https://git.io/ContextPrior.

Graph convolutional models have gained impressive successes on skeleton based human action recognition task. As graph convolution is a local operation, it cannot fully investigate non-local joints that could be vital to recognizing the action. For example, actions like typing and clapping request the cooperation of two hands, which are distant from each other in a human skeleton graph. Multiple graph convolutional layers thus tend to be stacked together to increase receptive field, which brings in computational inefficiency and optimization difficulty. But there is still no guarantee that distant joints (e.g. two hands) can be well integrated. In this paper, we propose a context aware graph convolutional network (CA-GCN). Besides the computation of localized graph convolution, CA-GCN considers a context term for each vertex by integrating information of all other vertices. Long range dependencies among joints are thus naturally integrated in context information, which then eliminates the need of stacking multiple layers to enlarge receptive field and greatly simplifies the network. Moreover, we further propose an advanced CA-GCN, in which asymmetric relevance measurement and higher level representation are utilized to compute context information for more flexibility and better performance. Besides the joint features, our CA-GCN could also be extended to handle graphs with edge (limb) features. Extensive experiments on two real-world datasets demonstrate the importance of context information and the effectiveness of the proposed CA-GCN in skeleton based action recognition.

Converting point clouds generated by Laser scanning, multiview stereo imagery or depth cameras into compact polygon meshes is a challenging problem in vision. Existing methods are either robust to imperfect data or scalable, but rarely both. In this paper, we address this issue with an hybrid method that successively connects and slices planes detected from 3D data. The core idea consists in constructing an efficient and compact partitioning data structure. The later is i) spatially-adaptive in the sense that a plane slices a restricted number of relevant planes only, and ii) composed of components with different structural meaning resulting from a preliminary analysis of the plane connectivity. Our experiments on a variety of objects and sensors show the versatility of our approach as well as its competitiveness with respect to existing methods.

Deep neural networks have achieved state-of-the-art performance in a wide range of recognition/classification tasks. However, when applying deep learning to real-world applications, there are still multiple challenges. A typical challenge is that unknown samples may be fed into the system during the testing phase and traditional deep neural networks will wrongly recognize the unknown sample as one of the known classes. Open set recognition is a potential solution to overcome this problem, where the open set classifier should have the ability to reject unknown samples as well as maintain high classification accuracy on known classes. The variational auto-encoder (VAE) is a popular model to detect unknowns, but it cannot provide discriminative representations for known classification. In this paper, we propose a novel method, Conditional Gaussian Distribution Learning (CGDL), for open set recognition. In addition to detecting unknown samples, this method can also classify known samples by forcing different latent features to approximate different Gaussian models. Meanwhile, to avoid information hidden in the input vanishing in the middle layers, we also adopt the probabilistic ladder architecture to extract high-level abstract features. Experiments on several standard image datasets reveal that the proposed method significantly outperforms the baseline method and achieves new state-of-the-art results.