CVPR 2020 论文列表

We present a new method to learn video representations from large-scale unlabeled video data. Ideally, this representation will be generic and transferable, directly usable for new tasks such as action recognition and zero or few-shot learning. We formulate unsupervised representation learning as a multi-modal, multi-task learning problem, where the representations are shared across different modalities via distillation. Further, we introduce the concept of loss function evolution by using an evolutionary search algorithm to automatically find optimal combination of loss functions capturing many (self-supervised) tasks and modalities. Thirdly, we propose an unsupervised representation evaluation metric using distribution matching to a large unlabeled dataset as a prior constraint, based on Zipf's law. This unsupervised constraint, which is not guided by any labeling, produces similar results to weakly-supervised, task-specific ones. The proposed unsupervised representation learning results in a single RGB network and outperforms previous methods. Notably, it is also more effective than several label-based methods (e.g., ImageNet), with the exception of large, fully labeled video datasets.

The high frame rate is a critical requirement for capturing fast human motions. In this setting, existing markerless image-based methods are constrained by the lighting requirement, the high data bandwidth and the consequent high computation overhead. In this paper, we propose EventCap -- the first approach for 3D capturing of high-speed human motions using a single event camera. Our method combines model-based optimization and CNN-based human pose detection to capture high frequency motion details and to reduce the drifting in the tracking. As a result, we can capture fast motions at millisecond resolution with significantly higher data efficiency than using high frame rate videos. Experiments on our new event-based fast human motion dataset demonstrate the effectiveness and accuracy of our method, as well as its robustness to challenging lighting conditions.

Weakly supervised object localization (WSOL) aims to localize object with only weak supervision like image-level labels. However, a long-standing problem for available techniques based on the classification network is that they often result in highlighting the most discriminative parts rather than the entire extent of object. Nevertheless, trying to explore the integral extent of the object could degrade the performance of image classification on the contrary. To remedy this, we propose a simple yet powerful approach by introducing a novel adversarial erasing technique, erasing integrated learning (EIL). By integrating discriminative region mining and adversarial erasing in a single forward-backward propagation in a vanilla CNN, the proposed EIL explores the high response class-specific area and the less discriminative region simultaneously, thus could maintain high performance in classification and jointly discover the full extent of the object. Furthermore, we apply multiple EIL (MEIL) modules at different levels of the network in a sequential manner, which for the first time integrates semantic features of multiple levels and multiple scales through adversarial erasing learning. In particular, the proposed EIL and advanced MEIL both achieve a new state-of-the-art performance in CUB-200-2011 and ILSVRC 2016 benchmark, making significant improvement in localization while advancing high performance in image classification.

In many real-world face restoration applications, e.g., smartphone photo albums and old films, multiple high-quality (HQ) images of the same person usually are available for a given degraded low-quality (LQ) observation. However, most existing guided face restoration methods are based on single HQ exemplar image, and are limited in properly exploiting guidance for improving the generalization ability to unknown degradation process. To address these issues, this paper suggests to enhance blind face restoration performance by utilizing multi-exemplar images and adaptive fusion of features from guidance and degraded images. First, given a degraded observation, we select the optimal guidance based on the weighted affine distance on landmark sets, where the landmark weights are learned to make the guidance image optimized to HQ image reconstruction. Second, moving least-square and adaptive instance normalization are leveraged for spatial alignment and illumination translation of guidance image in the feature space. Finally, for better feature fusion, multiple adaptive spatial feature fusion (ASFF) layers are introduced to incorporate guidance features in an adaptive and progressive manner, resulting in our ASFFNet. Experiments show that our ASFFNet performs favorably in terms of quantitative and qualitative evaluation, and is effective in generating photo-realistic results on real-world LQ images. The source code and models are available at https://github.com/csxmli2016/ASFFNet.

We propose an end-to-end variational generative model for scene layout synthesis conditioned on scene graphs. Unlike unconditional scene layout generation, we use scene graphs as an abstract but general representation to guide the synthesis of diverse scene layouts that satisfy relationships included in the scene graph. This gives rise to more flexible control over the synthesis process, allowing various forms of inputs such as scene layouts extracted from sentences or inferred from a single color image. Using our conditional layout synthesizer, we can generate various layouts that share the same structure of the input example. In addition to this conditional generation design, we also integrate a differentiable rendering module that enables layout refinement using only 2D projections of the scene. Given a depth and a semantics map, the differentiable rendering module enables optimizing over the synthesized layout to fit the given input in an analysis-by-synthesis fashion. Experiments suggest that our model achieves higher accuracy and diversity in conditional scene synthesis and allows exemplar-based scene generation from various input forms.

Annotating videos is cumbersome, expensive and not scalable. Yet, many strong video models still rely on manually annotated data. With the recent introduction of the HowTo100M dataset, narrated videos now offer the possibility of learning video representations without manual supervision. In this work we propose a new learning approach, MIL-NCE, capable of addressing mis- alignments inherent in narrated videos. With this approach we are able to learn strong video representations from scratch, without the need for any manual annotation. We evaluate our representations on a wide range of four downstream tasks over eight datasets: action recognition (HMDB-51, UCF-101, Kinetics-700), text-to- video retrieval (YouCook2, MSR-VTT), action localization (YouTube-8M Segments, CrossTask) and action segmentation (COIN). Our method outperforms all published self-supervised approaches for these tasks as well as several fully supervised baselines.

First-person video naturally brings the use of a physical environment to the forefront, since it shows the camera wearer interacting fluidly in a space based on his intentions. However, current methods largely separate the observed actions from the persistent space itself. We introduce a model for environment affordances that is learned directly from egocentric video. The main idea is to gain a human-centric model of a physical space (such as a kitchen) that captures (1) the primary spatial zones of interaction and (2) the likely activities they support. Our approach decomposes a space into a topological map derived from first-person activity, organizing an ego-video into a series of visits to the different zones. Further, we show how to link zones across multiple related environments (e.g., from videos of multiple kitchens) to obtain a consolidated representation of environment functionality. On EPIC-Kitchens and EGTEA+, we demonstrate our approach for learning scene affordances and anticipating future actions in long-form video.

Despite the substantial progress made in deep learning in recent years, advanced approaches remain computationally intensive. The trade-off between accuracy and computation time and energy limits their use in real-time applications on low power and other resource-constrained systems. In this paper, we tackle this fundamental challenge by introducing a hybrid optical-digital implementation of a convolutional neural network (CNN) based on engineering of the point spread function (PSF) of an optical imaging system. This is done by coding an imaging aperture such that its PSF replicates a large convolution kernel of the first layer of a pre-trained CNN. As the convolution takes place in the optical domain, it has zero cost in terms of energy consumption and has zero latency independent of the kernel size. Experimental results on two datasets demonstrate that our approach yields more than two orders of magnitude reduction in the computational cost while achieving near-state-of-the-art accuracy, or equivalently, better accuracy at the same computational cost.

Continuous-time trajectory representation has recently gained popularity for tasks where the fusion of high-frame-rate sensors and multiple unsynchronized devices is required. Lie group cumulative B-splines are a popular way of representing continuous trajectories without singularities. They have been used in near real-time SLAM and odometry systems with IMU, LiDAR, regular, RGB-D and event cameras, as well as for offline calibration. These applications require efficient computation of time derivatives (velocity, acceleration), but all prior works rely on a computationally suboptimal formulation. In this work we present an alternative derivation of time derivatives based on recurrence relations that needs O(k) instead of O(k^2) matrix operations (for a spline of order k) and results in simple and elegant expressions. While producing the same result, the proposed approach significantly speeds up the trajectory optimization and allows for computing simple analytic derivatives with respect to spline knots. The results presented in this paper pave the way for incorporating continuous-time trajectory representations into more applications where real-time performance is required.

Our goal is to use overhead imagery to understand patterns in traffic flow, for instance answering questions such as how fast could you traverse Times Square at 3am on a Sunday. A traditional approach for solving this problem would be to model the speed of each road segment as a function of time. However, this strategy is limited in that a significant amount of data must first be collected before a model can be used and it fails to generalize to new areas. Instead, we propose an automatic approach for generating dynamic maps of traffic speeds using convolutional neural networks. Our method operates on overhead imagery, is conditioned on location and time, and outputs a local motion model that captures likely directions of travel and corresponding travel speeds. To train our model, we take advantage of historical traffic data collected from New York City. Experimental results demonstrate that our method can be applied to generate accurate city-scale traffic models.

Object detection has achieved remarkable progress in the past decade. However, the detection of oriented and densely packed objects remains challenging because of following inherent reasons: (1) receptive fields of neurons are all axis-aligned and of the same shape, whereas objects are usually of diverse shapes and align along various directions; (2) detection models are typically trained with generic knowledge and may not generalize well to handle specific objects at test time; (3) the limited dataset hinders the development on this task. To resolve the first two issues, we present a dynamic refinement network that consists of two novel components, i.e., a feature selection module (FSM) and a dynamic refinement head (DRH). Our FSM enables neurons to adjust receptive fields in accordance with the shapes and orientations of target objects, whereas the DRH empowers our model to refine the prediction dynamically in an object-aware manner. To address the limited availability of related benchmarks, we collect an extensive and fully annotated dataset, namely, SKU110K-R, which is relabeled with oriented bounding boxes based on SKU110K. We perform quantitative evaluations on several publicly available benchmarks including DOTA, HRSC2016, SKU110K, and our own SKU110K-R dataset. Experimental results show that our method achieves consistent and substantial gains compared with baseline approaches. Our source code and dataset will be released to encourage follow-up research.

We propose novel dynamic multiscale graph neural networks (DMGNN) to predict 3D skeleton-based human motions. The core idea of DMGNN is to use a multiscale graph to comprehensively model the internal relations of a human body for motion feature learning. This multiscale graph is adaptive during training and dynamic across network layers. Based on this graph, we propose a multiscale graph computational unit (MGCU) to extract features at individual scales and fuse features across scales. The entire model is action-category-agnostic and follows an encoder-decoder framework. The encoder consists of a sequence of MGCUs to learn motion features. The decoder uses a proposed graph-based gate recurrent unit to generate future poses. Extensive experiments show that the proposed DMGNN outperforms state-of-the-art methods in both short and long-term predictions on the datasets of Human 3.6M and CMU Mocap. We further investigate the learned multiscale graphs for the interpretability. The codes could be downloaded from https://github.com/limaosen0/DMGNN.

Modelling long-range dependencies is critical for scene understanding tasks in computer vision. Although CNNs have excelled in many vision tasks, they are still limited in capturing long-range structured relationships as they typically consist of layers of local kernels. A fully-connected graph is beneficial for such modelling, however, its computational overhead is prohibitive. We propose a dynamic graph message passing network, that significantly reduces the computational complexity compared to related works modelling a fully-connected graph. This is achieved by adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically predict node-dependent filter weights and the affinity matrix for propagating information between them. Using this model, we show significant improvements with respect to strong, state-of-the-art baselines on three different tasks and backbone architectures. Our approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters.

Recovering the dynamic fluid surface is a long-standing challenging problem in computer vision. Most existing image-based methods require multiple views or a dedicated imaging system. Here we present a learning-based single-image approach for 3D fluid surface reconstruction. Specifically, we design a deep neural network that estimates the depth and normal maps of a fluid surface by analyzing the refractive distortion of a reference background image. Due to the dynamic nature of fluid surfaces, our network uses recurrent layers that carry temporal information from previous frames to achieve spatio-temporally consistent reconstruction given a video input. Due to the lack of fluid data, we synthesize a large fluid dataset using physics-based fluid modeling and rendering techniques for network training and validation. Through experiments on simulated and real captured fluid images, we demonstrate that our proposed deep neural network trained on our fluid dataset can recover dynamic 3D fluid surfaces with high accuracy.

Light-weight convolutional neural networks (CNNs) suffer performance degradation as their low computational budgets constrain both the depth (number of convolution layers) and the width (number of channels) of CNNs, resulting in limited representation capability. To address this issue, we present Dynamic Convolution, a new design that increases model complexity without increasing the network depth or width. Instead of using a single convolution kernel per layer, dynamic convolution aggregates multiple parallel convolution kernels dynamically based upon their attentions, which are input dependent. Assembling multiple kernels is not only computationally efficient due to the small kernel size, but also has more representation power since these kernels are aggregated in a non-linear way via attention. By simply using dynamic convolution for the state-of-the-art architecture MobileNetV3-Small, the top-1 accuracy of ImageNet classification is boosted by 2.9% with only 4% additional FLOPs and 2.9 AP gain is achieved on COCO keypoint detection.

We propose DualConvMesh-Nets (DCM-Net) a family of deep hierarchical convolutional networks over 3D geometric data that combines two types of convolutions. The first type, Geodesic convolutions, defines the kernel weights over mesh surfaces or graphs. That is, the convolutional kernel weights are mapped to the local surface of a given mesh. The second type, Euclidean convolutions, is independent of any underlying mesh structure. The convolutional kernel is applied on a neighborhood obtained from a local affinity representation based on the Euclidean distance between 3D points. Intuitively, geodesic convolutions can easily separate objects that are spatially close but have disconnected surfaces, while Euclidean convolutions can represent interactions between nearby objects better, as they are oblivious to object surfaces. To realize a multi-resolution architecture, we borrow well-established mesh simplification methods from the geometry processing domain and adapt them to define mesh-preserving pooling and unpooling operations. We experimentally show that combining both types of convolutions in our architecture leads to significant performance gains for 3D semantic segmentation, and we report competitive results on three scene segmentation benchmarks. Models and code will be made publicly available.

Current state-of-the-art semantic segmentation methods often apply high-resolution input to attain high performance, which brings large computation budgets and limits their applications on resource-constrained devices. In this paper, we propose a simple and flexible two-stream framework named Dual Super-Resolution Learning (DSRL) to effectively improve the segmentation accuracy without introducing extra computation costs. Specifically, the proposed method consists of three parts: Semantic Segmentation Super-Resolution (SSSR), Single Image Super-Resolution (SISR) and Feature Affinity (FA) module, which can keep high-resolution representations with low-resolution input while simultaneously reducing the model computation complexity. Moreover, it can be easily generalized to other tasks, e.g., human pose estimation. This simple yet effective method leads to strong representations and is evidenced by promising performance on both semantic segmentation and human pose estimation. Specifically, for semantic segmentation on CityScapes, we can achieve \geq2% higher mIoU with similar FLOPs, and keep the performance with 70% FLOPs. For human pose estimation, we can gain \geq2% mAP with the same FLOPs and maintain mAP with 30% fewer FLOPs. Code and models are available at https://github.com/wanglixilinx/DSRL.

We introduce DeepInversion, a new method for synthesizing images from the image distribution used to train a deep neural network. We "invert" a trained network (teacher) to synthesize class-conditional input images starting from random noise, without using any additional information about the training dataset. Keeping the teacher fixed, our method optimizes the input while regularizing the distribution of intermediate feature maps using information stored in the batch normalization layers of the teacher. Further, we improve the diversity of synthesized images using Adaptive DeepInversion, which maximizes the Jensen-Shannon divergence between the teacher and student network logits. The resulting synthesized images from networks trained on the CIFAR-10 and ImageNet datasets demonstrate high fidelity and degree of realism, and help enable a new breed of data-free applications - ones that do not require any real images or labeled data. We demonstrate the applicability of our proposed method to three tasks of immense practical importance - (i) data-free network pruning, (ii) data-free knowledge transfer, and (iii) data-free continual learning.

Existing models often leverage co-occurrences between objects and their context to improve recognition accuracy. However, strongly relying on context risks a model's generalizability, especially when typical co-occurrence patterns are absent. This work focuses on addressing such contextual biases to improve the robustness of the learnt feature representations. Our goal is to accurately recognize a category in the absence of its context, without compromising on performance when it co-occurs with context. Our key idea is to decorrelate feature representations of a category from its co-occurring context. We achieve this by learning a feature subspace that explicitly represents categories occurring in the absence of context along side a joint feature subspace that represents both categories and context. Our very simple yet effective method is extensible to two multi-label tasks -- object and attribute classification. On 4 challenging datasets, we demonstrate the effectiveness of our method in reducing contextual bias.