CVPR 2020 论文列表

We address a core problem of computer vision: Detection and description of 2D feature points for image matching. For a long time, hand-crafted designs, like the seminal SIFT algorithm, were unsurpassed in accuracy and efficiency. Recently, learned feature detectors emerged that implement detection and description using neural networks. Training these networks usually resorts to optimizing low-level matching scores, often pre-defining sets of image patches which should or should not match, or which should or should not contain key points. Unfortunately, increased accuracy for these low-level matching scores does not necessarily translate to better performance in high-level vision tasks. We propose a new training methodology which embeds the feature detector in a complete vision pipeline, and where the learnable parameters are trained in an end-to-end fashion. We overcome the discrete nature of key point selection and descriptor matching using principles from reinforcement learning. As an example, we address the task of relative pose estimation between a pair of images. We demonstrate that the accuracy of a state-of-the-art learning-based feature detector can be increased when trained for the task it is supposed to solve at test time. Our training methodology poses little restrictions on the task to learn, and works for any architecture which predicts key point heat maps, and descriptors for key point locations.

Online learning is crucial to robust visual object tracking as it can provide high discrimination power in the presence of background distractors. However, there are two contradictory factors affecting its successful deployment on the real visual tracking platform: the discrimination issue due to the challenges in vanilla gradient descent, which does not guarantee good convergence; the robustness issue due to over-fitting resulting from excessive update with limited memory size (the oldest samples are discarded). Despite many dedicated techniques proposed to somehow treat those issues, in this paper we take a new way to strike a compromise between them based on the recursive least-squares estimation (LSE) algorithm. After connecting each fully-connected layer with LSE separately via normal equations, we further propose an improved mini-batch stochastic gradient descent algorithm for fully-connected network learning with memory retention in a recursive fashion. This characteristic can spontaneously reduce the risk of over-fitting resulting from catastrophic forgetting in excessive online learning. Meanwhile, it can effectively improve convergence though the cost function is computed over all the training samples that the algorithm has ever seen. We realize this recursive LSE-aided online learning technique in the state-of-the-art RT-MDNet tracker, and the consistent improvements on four challenging benchmarks prove its efficiency without additional offline training and too much tedious work on parameter adjusting.

We observe that the human trajectory is not only forward predictable, but also backward predictable. Both forward and backward trajectories follow the same social norms and obey the same physical constraints with the only difference in their time directions. Based on this unique property, we develop a new approach, called reciprocal learning, for human trajectory prediction. Two networks, forward and backward prediction networks, are tightly coupled, satisfying the reciprocal constraint, which allows them to be jointly learned. Based on this constraint, we borrow the concept of adversarial attacks of deep neural networks, which iteratively modifies the input of the network to match the given or forced network output, and develop a new method for network prediction, called reciprocal attack for matched prediction. It further improves the prediction accuracy. Our experimental results on benchmark datasets demonstrate that our new method outperforms the state-of-the-art methods for human trajectory prediction.

Panoptic segmentation is a complex full scene parsing task requiring simultaneous instance and semantic segmentation at high resolution. Current state-of-the-art approaches cannot run in real-time, and simplifying these architectures to improve efficiency severely degrades their accuracy. In this paper, we propose a new single-shot panoptic segmentation network that leverages dense detections and a global self-attention mechanism to operate in real-time with performance approaching the state of the art. We introduce a novel parameter-free mask construction method that substantially reduces computational complexity by efficiently reusing information from the object detection and semantic segmentation sub-tasks. The resulting network has a simple data flow that requires no feature map re-sampling, enabling significant hardware acceleration. Our experiments on the Cityscapes and COCO benchmarks show that our network works at 30 FPS on 1024x2048 resolution, trading a 3% relative performance degradation from the current state of the art for up to 440% faster inference.

The success of Convolutional Neural Networks (CNNs) in various applications is accompanied by a significant increase in computation and training time. In this work, we focus on accelerating training by observing that about 90% of gradients are reusable during training. Leveraging this observation, we propose a new algorithm, Reuse-Sparse-Backprop (ReSprop), as a method to sparsify gradient vectors during CNN training. ReSprop maintains state-of-the-art accuracy on CIFAR-10, CIFAR-100, and ImageNet datasets with less than 1.1% accuracy loss while enabling a reduction in back-propagation computations by a factor of 10x resulting in a 2.7x overall speedup in training. As the computation reduction introduced by Re-Sprop is accomplished by introducing fine-grained sparsity that reduces computation efficiency on GPUs, we introduce a generic sparse convolution neural network accelerator (GSCN), which is designed to accelerate sparse back-propagation convolutions. When combined with ReSprop, GSCN achieves 8.0x and 7.2x speedup in the backward pass on ResNet34 and VGG16 versus a GTX 1080 Ti GPU.

The key challenge for 3D face shape reconstruction is to build the correct dense face correspondence between the deformable mesh and the single input image. Given the ill-posed nature, previous works heavily rely on prior knowledge (such as 3DMM [2]) to reduce depth ambiguity. Although impressive result has been made recently [42, 14, 8], there is still a large room to improve the correspondence so that projected face shape better aligns with the silhouette of each face region (i.e, eye, mouth, nose, cheek, etc.) on the image. To further reduce the ambiguities, we present a novel framework called "Reinforced Differentiable Attributes" ("ReDA") which is more general and effective than previous Differentiable Rendering ("DR"). Specifically, we first extend from color to more broad attributes, including the depth and the face parsing mask. Secondly, unlike the previous Z-buffer rendering, we make the rendering to be more differentiable through a set of convolution operations with multi-scale kernel sizes. In the meanwhile, to make "ReDA" to be more successful for 3D face recon-struction, we further introduce a new free-form deformation layer that sits on top of 3DMM to enjoy both the prior knowledge and out-of-space modeling. Both techniques can be easily integrated into existing 3D face reconstruction pipeline. Extensive experiments on both RGB and RGB-D datasets show that our approach outperforms prior arts.

We study the problem of efficient semantic segmentation for large-scale 3D point clouds. By relying on expensive sampling techniques or computationally heavy pre/post-processing steps, most existing approaches are only able to be trained and operate over small-scale point clouds. In this paper, we introduce RandLA-Net, an efficient and lightweight neural architecture to directly infer per-point semantics for large-scale point clouds. The key to our approach is to use random point sampling instead of more complex point selection approaches. Although remarkably computation and memory efficient, random sampling can discard key features by chance. To overcome this, we introduce a novel local feature aggregation module to progressively increase the receptive field for each 3D point, thereby effectively preserving geometric details. Extensive experiments show that our RandLA-Net can process 1 million points in a single pass with up to 200x faster than existing approaches. Moreover, our RandLA-Net clearly surpasses state-of-the-art approaches for semantic segmentation on two large-scale benchmarks Semantic3D and SemanticKITTI.

Deep neural network based reinforcement learning (RL) can learn appropriate visual representations for complex tasks like vision-based robotic grasping without the need for manually engineering or prior learning a perception system. However, data for RL is collected via running an agent in the desired environment, and for applications like robotics, running a robot in the real world may be extremely costly and time consuming. Simulated training offers an appealing alternative, but ensuring that policies trained in simulation can transfer effectively into the real world requires additional machinery. Simulations may not match reality, and typically bridging the simulation-to-reality gap requires domain knowledge and task-specific engineering. We can automate this process by employing generative models to translate simulated images into realistic ones. However, this sort of translation is typically task-agnostic, in that the translated images may not preserve all features that are relevant to the task. In this paper, we introduce the RL-scene consistency loss for image translation, which ensures that the translation operation is invariant with respect to the Q-values associated with the image. This allows us to learn a task-aware translation. Incorporating this loss into unsupervised domain translation, we obtain the RL-CycleGAN, a new approach for simulation-to-real-world transfer for reinforcement learning. In evaluations of RL-CycleGAN on two vision-based robotics grasping tasks, we show that RL-CycleGAN offers a substantial improvement over a number of prior methods for sim-to-real transfer, attaining excellent real-world performance with only a modest number of real-world observations.

One of the long-term challenges of robotics is to enable robots to interact with humans in the visual world via natural language, as humans are visual animals that communicate through language. Overcoming this challenge requires the ability to perform a wide variety of complex tasks in response to multifarious instructions from humans. In the hope that it might drive progress towards more flexible and powerful human interactions with robots, we propose a dataset of varied and complex robot tasks, described in natural language, in terms of objects visible in a large set of real images. Given an instruction, success requires navigating through a previously-unseen environment to identify an object. This represents a practical challenge, but one that closely reflects one of the core visual problems in robotics. Several state-of-the-art vision-and-language navigation, and referring-expression models are tested to verify the difficulty of this new task, but none of them show promising results because there are many fundamental differences between our task and previous ones. A novel Interactive Navigator-Pointer model is also proposed that provides a strong baseline on the task. The proposed model especially achieves the best performance on the unseen test split, but still leaves substantial room for improvement compared to the human performance. Repository: https://github.com/YuankaiQi/REVERIE.

Imperfect data (noise, outliers and partial overlap) and high degrees of freedom make non-rigid registration a classical challenging problem in computer vision. Existing methods typically adopt the l_p type robust estimator to regularize the fitting and smoothness, and the proximal operator is used to solve the resulting non-smooth problem. However, the slow convergence of these algorithms limits its wide applications. In this paper, we propose a formulation for robust non-rigid registration based on a globally smooth robust estimator for data fitting and regularization, which can handle outliers and partial overlaps. We apply the majorization-minimization algorithm to the problem, which reduces each iteration to solving a simple least-squares problem with L-BFGS. Extensive experiments demonstrate the effectiveness of our method for non-rigid alignment between two shapes with outliers and partial overlap. with quantitative evaluation showing that it outperforms state-of-the-art methods in terms of registration accuracy and computational speed. The source code is available at https://github.com/Juyong/Fast_RNRR.

Autonomous vehicles commonly rely on highly detailed birds-eye-view maps of their environment, which capture both static elements of the scene such as road layout as well as dynamic elements such as other cars and pedestrians. Generating these map representations on the fly is a complex multi-stage process which incorporates many important vision-based elements, including ground plane estimation, road segmentation and 3D object detection. In this work we present a simple, unified approach for estimating these map representations directly from monocular images using a single end-to-end deep learning architecture. For the maps themselves we adopt a semantic Bayesian occupancy grid framework, allowing us to trivially accumulate information over multiple cameras and timesteps. We demonstrate the effectiveness of our approach by evaluating against several challenging baselines on the NuScenes and Argoverse datasets, and show that we are able to achieve a relative improvement of 9.1% and 22.3% respectively compared to the best-performing existing method.

Human gaze behavior prediction is important for behavioral vision and for computer vision applications. Most models mainly focus on predicting free-viewing behavior using saliency maps, but do not generalize to goal-directed behavior, such as when a person searches for a visual target object. We propose the first inverse reinforcement learning (IRL) model to learn the internal reward function and policy used by humans during visual search. We modeled the viewer's internal belief states as dynamic contextual belief maps of object locations. These maps were learned and then used to predict behavioral scanpaths for multiple target categories. To train and evaluate our IRL model we created COCO-Search18, which is now the largest dataset of high-quality search fixations in existence. COCO-Search18 has 10 participants searching for each of 18 target-object categories in 6202 images, making about 300,000 goal-directed fixations. When trained and evaluated on COCO-Search18, the IRL model outperformed baseline models in predicting search fixation scanpaths, both in terms of similarity to human search behavior and search efficiency. Finally, reward maps recovered by the IRL model reveal distinctive target-dependent patterns of object prioritization, which we interpret as a learned object context.

In this paper, we introduce an anchor-box free and single shot instance segmentation method, which is conceptually simple, fully convolutional and can be used by easily embedding it into most off-the-shelf detection methods. Our method, termed PolarMask, formulates the instance segmentation problem as predicting contour of instance through instance center classification and dense distance regression in a polar coordinate. Moreover, we propose two effective approaches to deal with sampling high-quality center examples and optimization for dense distance regression, respectively, which can significantly improve the performance and simplify the training process. Without any bells and whistles, PolarMask achieves 32.9% in mask mAP with single-model and single-scale training/testing on the challenging COCO dataset. For the first time, we show that the complexity of instance segmentation, in terms of both design and computation complexity, can be the same as bounding box object detection and this much simpler and flexible instance segmentation framework can achieve competitive accuracy. We hope that the proposed PolarMask framework can serve as a fundamental and strong baseline for single shot instance segmentation task.

We present a new method for efficient high-quality image segmentation of objects and scenes. By analogizing classical computer graphics methods for efficient rendering with over- and undersampling challenges faced in pixel labeling tasks, we develop a unique perspective of image segmentation as a rendering problem. From this vantage, we present the PointRend (Point-based Rendering) neural network module: a module that performs point-based segmentation predictions at adaptively selected locations based on an iterative subdivision algorithm. PointRend can be flexibly applied to both instance and semantic segmentation tasks by building on top of existing state-of-the-art models. While many concrete implementations of the general idea are possible, we show that a simple design already achieves excellent results. Qualitatively, PointRend outputs crisp object boundaries in regions that are over-smoothed by previous methods. Quantitatively, PointRend yields significant gains on COCO and Cityscapes, for both instance and semantic segmentation. PointRend's efficiency enables output resolutions that are otherwise impractical in terms of memory or computation compared to existing approaches. Code has been made available at https://github.com/facebookresearch/detectron2/tree/master/projects/PointRend.

Instance segmentation is an important task for scene understanding. Compared to the fully-developed 2D, 3D instance segmentation for point clouds have much room to improve. In this paper, we present PointGroup, a new end-to-end bottom-up architecture, specifically focused on better grouping the points by exploring the void space between objects. We design a two-branch network to extract point features and predict semantic labels and offsets, for shifting each point towards its respective instance centroid. A clustering component is followed to utilize both the original and offset-shifted point coordinate sets, taking advantage of their complementary strength. Further, we formulate the ScoreNet to evaluate the candidate instances, followed by the Non-Maximum Suppression (NMS) to remove duplicates. We conduct extensive experiments on two challenging datasets, ScanNet v2 and S3DIS, on which our method achieves the highest performance, 63.6% and 64.0%, compared to 54.9% and 54.4% achieved by former best solutions in terms of mAP with IoU threshold 0.5.

We present PointAugment, a new auto-augmentation framework that automatically optimizes and augments point cloud samples to enrich the data diversity when we train a classification network. Different from existing auto-augmentation methods for 2D images, PointAugment is sample-aware and takes an adversarial learning strategy to jointly optimize an augmentor network and a classifier network, such that the augmentor can learn to produce augmented samples that best fit the classifier. Moreover, we formulate a learnable point augmentation function with a shape-wise transformation and a point-wise displacement, and carefully design loss functions to adopt the augmented samples based on the learning progress of the classifier. Extensive experiments also confirm PointAugment's effectiveness and robustness to improve the performance of various networks on shape classification and retrival.

Snapshot compressive imaging (SCI) aims to capture the high-dimensional (usually 3D) images using a 2D sensor (detector) in a single snapshot. Though enjoying the advantages of low-bandwidth, low-power and low-cost, applying SCI to large-scale problems (HD or UHD videos) in our daily life is still challenging. The bottleneck lies in the reconstruction algorithms; they are either too slow (iterative optimization algorithms) or not flexible to the encoding process (deep learning based end-to-end networks). In this paper, we develop fast and flexible algorithms for SCI based on the plug-and-play (PnP) framework. In addition to the widely used PnP-ADMM method, we further propose the PnP-GAP (generalized alternating projection) algorithm with a lower computational workload and prove the global convergence of PnP-GAP under the SCI hardware constraints. By employing deep denoising priors, we first time show that PnP can recover a UHD color video (3840x1644x48 with PNSR above 30dB) from a snapshot 2D measurement. Extensive results on both simulation and real datasets verify the superiority of our proposed algorithm.

We address the challenging task of occlusion-aware indoor 3D scene understanding. We represent scenes by a set of planes, where each one is defined by its normal, offset and two masks outlining (i) the extent of the visible part and (ii) the full region that consists of both visible and occluded parts of the plane. We infer these planes from a single input image with a novel neural network architecture. It consists of a two-branch category-specific module that aims to predict layout and objects of the scene separately so that different types of planes can be handled better. We also introduce a novel loss function based on plane warping that can leverage multiple views at training time for improved occlusion-aware reasoning. In order to train and evaluate our occlusion-reasoning model, we use the ScanNet dataset and propose (i) a strategy to automatically extract ground truth for both visible and hidden regions and (ii) a new evaluation metric that specifically focuses on the prediction in hidden regions. We empirically demonstrate that our proposed approach can achieve higher accuracy for occlusion reasoning compared to competitive baselines on the ScanNet dataset, e.g. 42.65% relative improvement on hidden regions.

In this paper, we address the makeup transfer task, which aims to transfer the makeup from a reference image to a source image. Existing methods have achieved promising progress in constrained scenarios, but transferring between images with large pose and expression differences is still challenging. Besides, they cannot realize customizable transfer that allows a controllable shade of makeup or specifies the part to transfer, which limits their applications. To address these issues, we propose Pose and expression robust Spatial-aware GAN (PSGAN). It first utilizes Makeup Distill Network to disentangle the makeup of the reference image as two spatial-aware makeup matrices. Then, Attentive Makeup Morphing module is introduced to specify how the makeup of a pixel in the source image is morphed from the reference image. With the makeup matrices and the source image, Makeup Apply Network is used to perform makeup transfer. Our PSGAN not only achieves state-of-the-art results even when large pose and expression differences exist but also is able to perform partial and shade-controllable makeup transfer. Both the code and a newly collected dataset containing facial images with various poses and expressions will be available at https://github.com/wtjiang98/PSGAN.