CVPR 2020 论文列表

In this work, we propose a unified model compression framework called Multi-Dimensional Pruning (MDP) to simultaneously compress the convolutional neural networks (CNNs) on multiple dimensions. In contrast to the existing model compression methods that only aim to reduce the redundancy along either the spatial/spatial-temporal dimension (e.g., spatial dimension for 2D CNNs, spatial and temporal dimensions for 3D CNNs) or the channel dimension, our newly proposed approach can simultaneously reduce the spatial/spatial-temporal and the channel redundancies for CNNs. Specifically, in order to reduce the redundancy along the spatial/spatial-temporal dimension, we downsample the input tensor of a convolutional layer, in which the scaling factor for the downsampling operation is adaptively selected by our approach. After the convolution operation, the output tensor is upsampled to the original size to ensure the unchanged input size for the subsequent CNN layers. To reduce the channel-wise redundancy, we introduce a gate for each channel of the output tensor as its importance score, in which the gate value is automatically learned. The channels with small importance scores will be removed after the model compression process. Our comprehensive experiments on four benchmark datasets demonstrate that our MDP framework outperforms the existing methods when pruning both 2D CNNs and 3D CNNs.

Deep metric learning leverages carefully designed sampling strategies and loss functions that aid in optimizing the generation of a discriminable embedding space. While effective sampling of pairs is critical for shaping the metric space during training, the relative interactions between pairs, and consequently the forces exerted on these pairs that direct their displacement in the embedding space can significantly impact the formation of well separated clusters. In this work, we identify a shortcoming of existing loss formulations which fail to consider more optimal directions of pair displacements as another criterion for optimization. We propose a novel direction regularization to explicitly account for the layout of sampled pairs and attempt to introduce orthogonality in the representations. The proposed regularization is easily integrated into existing loss functions providing considerable performance improvements. We experimentally validate our hypothesis on the Cars-196, CUB-200 and InShop datasets and outperform existing methods to yield state-of-the-art results on these datasets.

The ability to reliably perceive the environmental states, particularly the existence of objects and their motion behavior, is crucial for autonomous driving. In this work, we propose an efficient deep model, called MotionNet, to jointly perform perception and motion prediction from 3D point clouds. MotionNet takes a sequence of LiDAR sweeps as input and outputs a bird's eye view (BEV) map, which encodes the object category and motion information in each grid cell. The backbone of MotionNet is a novel spatio-temporal pyramid network, which extracts deep spatial and temporal features in a hierarchical fashion. To enforce the smoothness of predictions over both space and time, the training of MotionNet is further regularized with novel spatial and temporal consistency losses. Extensive experiments show that the proposed method overall outperforms the state-of-the-arts, including the latest scene-flow- and 3D-object-detection-based methods. This indicates the potential value of the proposed method serving as a backup to the bounding-box-based system, and providing complementary information to the motion planner in autonomous driving. Code is available at https://www.merl.com/research/license#MotionNet.

Robots and other smart devices need efficient object-based scene representations from their on-board vision systems to reason about contact, physics and occlusion. Recognized precise object models will play an important role alongside non-parametric reconstructions of unrecognized structures. We present a system which can estimate the accurate poses of multiple known objects in contact and occlusion from real-time, embodied multi-view vision. Our approach makes 3D object pose proposals from single RGB-D views, accumulates pose estimates and non-parametric occupancy information from multiple views as the camera moves, and performs joint optimization to estimate consistent, non-intersecting poses for multiple objects in contact. We verify the accuracy and robustness of our approach experimentally on 2 object datasets: YCB-Video, and our own challenging Cluttered YCB-Video. We demonstrate a real-time robotics application where a robot arm precisely and orderly disassembles complicated piles of objects, using only on-board RGB-D vision.

Visual attention not only improves the performance of image captioners, but also serves as a visual interpretation to qualitatively measure the caption rationality and model transparency. Specifically, we expect that a captioner can fix its attentive gaze on the correct objects while generating the corresponding words. This ability is also known as grounded image captioning. However, the grounding accuracy of existing captioners is far from satisfactory.To improve the grounding accuracy while retaining the captioning quality, it is expensive to collect the word-region alignment as strong supervision.To this end, we propose a Part-of-Speech (POS) enhanced image-text matching model (SCAN): POS-SCAN, as the effective knowledge distillation for more grounded image captioning. The benefits are two-fold: 1) given a sentence and an image, POS-SCAN can ground the objects more accurately than SCAN; 2) POS-SCAN serves as a word-region alignment regularization for the captioner's visual attention module. By showing benchmark experimental results, we demonstrate that conventional image captioners equipped with POS-SCAN can significantly improve the grounding accuracy without strong supervision. Last but not the least, we explore the indispensable Self-Critical Sequence Training (SCST) in the context of grounded image captioning and show that the image-text matching score can serve as a reward for more grounded captioning.

We present a novel method for monocular hand shape and pose estimation at unprecedented runtime performance of 100fps and at state-of-the-art accuracy. This is enabled by a new learning based architecture designed such that it can make use of all the sources of available hand training data: image data with either 2D or 3D annotations, as well as stand-alone 3D animations without corresponding image data. It features a 3D hand joint detection module and an inverse kinematics module which regresses not only 3D joint positions but also maps them to joint rotations in a single feed-forward pass. This output makes the method more directly usable for applications in computer vision and graphics compared to only regressing 3D joint positions. We demonstrate that our architectural design leads to a significant quantitative and qualitative improvement over the state of the art on several challenging benchmarks. We will make our code publicly available for future research.

Monocular 3D object detection is an essential component in autonomous driving while challenging to solve, especially for those occluded samples which are only partially visible. Most detectors consider each 3D object as an independent training target, inevitably resulting in a lack of useful information for occluded samples. To this end, we propose a novel method to improve the monocular 3D object detection by considering the relationship of paired samples. This allows us to encode spatial constraints for partially-occluded objects from their adjacent neighbors. Specifically, the proposed detector computes uncertainty-aware predictions for object locations and 3D distances for the adjacent object pairs, which are subsequently jointly optimized by nonlinear least squares. Finally, the one-stage uncertainty-aware prediction structure and the post-optimization module are dedicatedly integrated for ensuring the run-time efficiency. Experiments demonstrate that our method yields the best performance on KITTI 3D detection benchmark, by outperforming state-of-the-art competitors by wide margins, especially for the hard samples.

We present Momentum Contrast (MoCo) for unsupervised visual representation learning. From a perspective on contrastive learning as dictionary look-up, we build a dynamic dictionary with a queue and a moving-averaged encoder. This enables building a large and consistent dictionary on-the-fly that facilitates contrastive unsupervised learning. MoCo provides competitive results under the common linear protocol on ImageNet classification. More importantly, the representations learned by MoCo transfer well to downstream tasks. MoCo can outperform its supervised pre-training counterpart in 7 detection/segmentation tasks on PASCAL VOC, COCO, and other datasets, sometimes surpassing it by large margins. This suggests that the gap between unsupervised and supervised representation learning has been largely closed in many vision tasks.

Despite their effectiveness in a wide range of tasks, deep architectures suffer from some important limitations. In particular, they are vulnerable to catastrophic forgetting, i.e. they perform poorly when they are required to update their model as new classes are available but the original training set is not retained. This paper addresses this problem in the context of semantic segmentation. Current strategies fail on this task because they do not consider a peculiar aspect of semantic segmentation: since each training step provides annotation only for a subset of all possible classes, pixels of the background class (i.e. pixels that do not belong to any other classes) exhibit a semantic distribution shift. In this work we revisit classical incremental learning methods, proposing a new distillation-based framework which explicitly accounts for this shift. Furthermore, we introduce a novel strategy to initialize classifier's parameters, thus preventing biased predictions toward the background class. We demonstrate the effectiveness of our approach with an extensive evaluation on the Pascal-VOC 2012 and ADE20K datasets, significantly outperforming state of the art incremental learning methods.

While deep learning continues to permeate through all fields of signal processing and machine learning, a critical exploit in these frameworks exists and remains unsolved. These exploits, or adversarial examples, are a type of signal attack that can change the output class of a classifier by perturbing the stimulus signal by an imperceptible amount. The attack takes advantage of statistical irregularities within the training data, where the added perturbations can move the image across deep learning decision boundaries. What is even more alarming is the transferability of these attacks to different deep learning models and architectures. This means a successful attack on one model has adversarial effects on other, unrelated models. In a general sense, adversarial attack through perturbations is not a machine learning vulnerability. Human and biological vision can also be fooled by various methods, i.e. mixing high and low frequency images together, by altering semantically related signals, or by sufficiently distorting the input signal. However, the amount and magnitude of such a distortion required to alter biological perception is at a much larger scale. In this work, we explored this gap through the lens of biology and neuroscience in order to understand the robustness exhibited in human perception. Our experiments show that by leveraging sparsity and modeling the biological mechanisms at a cellular level, we are able to mitigate the effect of adversarial alterations to the signal that have no perceptible meaning. Furthermore, we present and illustrate the effects of top-down functional processes that contribute to the inherent immunity in human perception in the context of exploiting these properties to make a more robust machine vision system.

In this paper, we investigate a challenging unsupervised domain adaptation setting --- unsupervised model adaptation. We aim to explore how to rely only on unlabeled target data to improve performance of an existing source prediction model on the target domain, since labeled source data may not be available in some real-world scenarios due to data privacy issues. For this purpose, we propose a new framework, which is referred to as collaborative class conditional generative adversarial net to bypass the dependence on the source data. Specifically, the prediction model is to be improved through generated target-style data, which provides more accurate guidance for the generator. As a result, the generator and the prediction model can collaborate with each other without source data. Furthermore, due to the lack of supervision from source data, we propose a weight constraint that encourages similarity to the source model. A clustering-based regularization is also introduced to produce more discriminative features in the target domain. Compared to conventional domain adaptation methods, our model achieves superior performance on multiple adaptation tasks with only unlabeled target data, which verifies its effectiveness in this challenging setting.

This paper considers a network referred to as Modality Shifting Attention Network (MSAN) for Multimodal Video Question Answering (MVQA) task. MSAN decomposes the task into two sub-tasks: (1) localization of temporal moment relevant to the question, and (2) accurate prediction of the answer based on the localized moment. The modality required for temporal localization may be different from that for answer prediction, and this ability to shift modality is essential for performing the task. To this end, MSAN is based on (1) the moment proposal network (MPN) that attempts to locate the most appropriate temporal moment from each of the modalities, and also on (2) the heterogeneous reasoning network (HRN) that predicts the answer using an attention mechanism on both modalities. MSAN is able to place importance weight on the two modalities for each sub-task using a component referred to as Modality Importance Modulation (MIM). Experimental results show that MSAN outperforms previous state-of-the-art by achieving 71.13% test accuracy on TVQA benchmark dataset. Extensive ablation studies and qualitative analysis are conducted to validate various components of the network.

Multi-Class Incremental Learning (MCIL) aims to learn new concepts by incrementally updating a model trained on previous concepts. However, there is an inherent trade-off to effectively learning new concepts without catastrophic forgetting of previous ones. To alleviate this issue, it has been proposed to keep around a few examples of the previous concepts but the effectiveness of this approach heavily depends on the representativeness of these examples. This paper proposes a novel and automatic framework we call mnemonics, where we parameterize exemplars and make them optimizable in an end-to-end manner. We train the framework through bilevel optimizations, i.e., model-level and exemplar-level. We conduct extensive experiments on three MCIL benchmarks, CIFAR-100, ImageNet-Subset and ImageNet, and show that using mnemonics exemplars can surpass the state-of-the-art by a large margin. Interestingly and quite intriguingly, the mnemonics exemplars tend to be on the boundaries between different classes.

Despite the blooming success of architecture search for vision tasks in resource-constrained environments, the design of on-device object detection architectures have mostly been manual. The few automated search efforts are either centered around non-mobile-friendly search spaces or not guided by on-device latency. We propose MnasFPN, a mobile-friendly search space for the detection head, and combine it with latency-aware architecture search to produce efficient object detection models. The learned MnasFPN head, when paired with MobileNetV2 body, outperforms MobileNetV3+SSDLite by 1.8 mAP at similar latency on Pixel. It is both 1 mAP more accurate and 10% faster than NAS-FPNLite. Ablation studies show that the majority of the performance gain comes from innovations in the search space. Further explorations reveal an interesting coupling between the search space design and the search algorithm, for which the complexity of MnasFPN search space is opportune.

Mixture models are well-established learning approaches that, in computer vision, have mostly been applied to inverse or ill-defined problems. However, they are general-purpose divide-and-conquer techniques, splitting the input space into relatively homogeneous subsets in a data-driven manner. Not only ill-defined but also well-defined complex problems should benefit from them. To this end, we devise a framework for spatial regression using mixture density networks. We realize the framework for object detection and human pose estimation. For both tasks, a mixture model yields higher accuracy and divides the input space into interpretable modes. For object detection, mixture components focus on object scale, with the distribution of components closely following that of ground truth the object scale. This practically alleviates the need for multi-scale testing, providing a superior speed-accuracy trade-off. For human pose estimation, a mixture model divides the data based on viewpoint and uncertainty -- namely, front and back views, with back view imposing higher uncertainty. We conduct experiments on the MS COCO dataset and do not face any mode collapse.

We present MixNMatch, a conditional generative model that learns to disentangle and encode background, object pose, shape, and texture from real images with minimal supervision, for mix-and-match image generation. We build upon FineGAN, an unconditional generative model, to learn the desired disentanglement and image generator, and leverage adversarial joint image-code distribution matching to learn the latent factor encoders. MixNMatch requires bounding boxes during training to model background, but requires no other supervision. Through extensive experiments, we demonstrate MixNMatch's ability to accurately disentangle, encode, and combine multiple factors for mix-and-match image generation, including sketch2color, cartoon2img, and img2gif applications. Our code/models/demo can be found at https://github.com/Yuheng-Li/MixNMatch

Racial equality is an important theme of international human rights law, but it has been largely obscured when the overall face recognition accuracy is pursued blindly. More facts indicate racial bias indeed degrades the fairness of recognition system and the error rates on non-Caucasians are usually much higher than Caucasians. To encourage fairness, we introduce the idea of adaptive margin to learn balanced performance for different races based on large margin losses. A reinforcement learning based race balance network (RL-RBN) is proposed. We formulate the process of finding the optimal margins for non-Caucasians as a Markov decision process and employ deep Q-learning to learn policies for an agent to select appropriate margin by approximating the Q-value function. Guided by the agent, the skewness of feature scatter between races can be reduced. Besides, we provide two ethnicity aware training datasets, called BUPT-Globalface and BUPT-Balancedface dataset, which can be utilized to study racial bias from both data and algorithm aspects. Extensive experiments on RFW database show that RL-RBN successfully mitigates racial bias and learns more balanced performance.

The curvature regularities have received growing attention with the advantage of providing strong priors in the continuity of edges in image processing applications. However, owing to the non-convex and non-smooth properties of the high-order regularizer, the numerical solution becomes challenging in real-time tasks. In this paper, we propose a novel curvature regularity, the total curvature (TC), by minimizing the normal curvatures along different directions. We estimate the normal curvatures discretely in the local neighborhood according to differential geometry theory. The resulting curvature regularity can be regarded as a re-weighted total variation (TV) minimization problem, which can be efficiently solved by the alternating direction method of multipliers (ADMM) based algorithm. By comparing with TV and Euler's elastica energy, we demonstrate the effectiveness and superiority of the total curvature regularity for various image processing applications.

We explore the possibility of using line intersection constraints for 3D scan registration. Typical 3D registration algorithms exploit point and plane correspondences, while line intersection constraints have not been used in the context of 3D scan registration before. Constraints from a match of pairs of intersecting lines in two 3D scans can be seen as two 3D line intersections, a plane correspondence, and a point correspondence. In this paper, we present minimal solvers that combine these different type of constraints: 1) three line intersections and one point match; 2) one line intersection and two point matches; 3) three line intersections and one plane match; 4) one line intersection and two plane matches; and 5) one line intersection, one point match, and one plane match. To use all the available solvers, we present a hybrid RANSAC loop. We propose a non-linear refinement technique using all the inliers obtained from the RANSAC. Vast experiments with simulated data and two real-data data-sets show that the use of these features and the combined solvers improve the accuracy. The code is available.