CVPR 2020 Papers

Towards 3D object tracking in point clouds, a novel point-to-box network termed P2B is proposed in an end-to-end learning manner. Our main idea is to first localize potential target centers in 3D search area embedded with target information. Then point-driven 3D target proposal and verification are executed jointly. In this way, the time-consuming 3D exhaustive search can be avoided. Specifically, we first sample seeds from the point clouds in template and search area respectively. Then, we execute permutation-invariant feature augmentation to embed target clues from template into search area seeds and represent them with target-specific features. Consequently, the augmented search area seeds regress the potential target centers via Hough voting. The centers are further strengthened with seed-wise targetness scores. Finally, each center clusters its neighbors to leverage the ensemble power for joint 3D target proposal and verification. We apply PointNet++ as our backbone and experiments on KITTI tracking dataset demonstrate P2B's superiority ( 10%'s improvement over state-of-the-art). Note that P2B can run with 40FPS on a single NVIDIA 1080Ti GPU. Our code and model are available at https://github.com/HaozheQi/P2B.

Solving long-tail large vocabulary object detection with deep learning based models is a challenging and demanding task, which is however under-explored. In this work, we provide the first systematic analysis on the underperformance of state-of-the-art models in front of long-tail distribution. We find existing detection methods are unable to model few-shot classes when the dataset is extremely skewed, which can result in classifier imbalance in terms of parameter magnitude. Directly adapting long-tail classification models to detection frameworks can not solve this problem due to the intrinsic difference between detection and classification. In this work, we propose a novel balanced group softmax (BAGS) module for balancing the classifiers within the detection frameworks through group-wise training. It implicitly modulates the training process for the head and tail classes and ensures they are both sufficiently trained, without requiring any extra sampling for the instances from the tail classes. Extensive experiments on the very recent long-tail large vocabulary object recognition benchmark LVIS show that our proposed BAGS significantly improves the performance of detectors with various backbones and frameworks on both object detection and instance segmentation. It beats all state-of-the-art methods transferred from long-tail image classification and establishes new state-of-the-art. Code is available at https://github.com/FishYuLi/BalancedGroupSoftmax.

Rank-based metrics are some of the most widely used criteria for performance evaluation of computer vision models. Despite years of effort, direct optimization for these metrics remains a challenge due to their non-differentiable and non-decomposable nature. We present an efficient, theoretically sound, and general method for differentiating rank-based metrics with mini-batch gradient descent. In addition, we address optimization instability and sparsity of the supervision signal that both arise from using rank-based metrics as optimization targets. Resulting losses based on recall and Average Precision are applied to image retrieval and object detection tasks. We obtain performance that is competitive with state-of-the-art on standard image retrieval datasets and consistently improve performance of near state-of-the-art object detectors.

A typical domain adaptation approach is to adapt models trained on the annotated data in a source domain (e.g., sunny weather) for achieving high performance on the test data in a target domain (e.g., rainy weather). Whether the target contains a single homogeneous domain or multiple heterogeneous domains, existing works always assume that there exist clear distinctions between the domains, which is often not true in practice (e.g., changes in weather). We study an open compound domain adaptation (OCDA) problem, in which the target is a compound of multiple homogeneous domains without domain labels, reflecting realistic data collection from mixed and novel situations. We propose a new approach based on two technical insights into OCDA: 1) a curriculum domain adaptation strategy to bootstrap generalization across domains in a data-driven self-organizing fashion and 2) a memory module to increase the model's agility towards novel domains. Our experiments on digit classification, facial expression recognition, semantic segmentation, and reinforcement learning demonstrate the effectiveness of our approach.

This work presents an efficient yet effective online Knowledge Distillation method via Collaborative Learning, termed KDCL, which is able to consistently improve the generalization ability of deep neural networks (DNNs) that have different learning capacities. Unlike existing two-stage knowledge distillation approaches that pre-train a DNN with large capacity as the "teacher" and then transfer the teacher's knowledge to another "student" DNN unidirectionally (i.e. one-way), KDCL treats all DNNs as "students" and collaboratively trains them in a single stage (knowledge is transferred among arbitrary students during collaborative training), enabling parallel computing, fast computations, and appealing generalization ability. Specifically, we carefully design multiple methods to generate soft target as supervisions by effectively ensembling predictions of students and distorting the input images. Extensive experiments show that KDCL consistently improves all the "students" on different datasets, including CIFAR-100 and ImageNet. For example, when trained together by using KDCL, ResNet-50 and MobileNetV2 achieve 78.2% and 74.0% top-1 accuracy on ImageNet, outperforming the original results by 1.4% and 2.0% respectively. We also verify that models pre-trained with KDCL transfer well to object detection and semantic segmentation on MS COCO dataset. For instance, the FPN detector is improved by 0.9% mAP.

We present a novel deep compression algorithm to reduce the memory footprint of LiDAR point clouds. Our method exploits the sparsity and structural redundancy between points to reduce the bitrate. Towards this goal, we first encode the point cloud into an octree, a data-efficient structure suitable for sparse point clouds. We then design a tree-structured conditional entropy model that can be directly applied to octree structures to predict the probability of a symbol's occurrence. We validate the effectiveness of our method over two large-scale datasets. The results demonstrate that our approach reduces the bitrate by 10- 20% at the same reconstruction quality, compared to the previous state-of-the-art. Importantly, we also show that for the same bitrate, our approach outperforms other compression algorithms when performing downstream 3D segmentation and detection tasks using compressed representations. This helps advance the feasibility of using point cloud compression to reduce the onboard and offboard storage for safety-critical applications such as self-driving cars, where a single vehicle captures 84 billion points per day.

Occlusions between human and objects, especially for the activities of human-object interactions, are very common in practical applications. However, most of the existing approaches for 3D human shape and pose estimation require human bodies are well captured without occlusions or with minor self-occlusions. In this paper, we focus on the problem of directly estimating the object-occluded human shape and pose from single color images. Our key idea is to utilize a partial UV map to represent an object-occluded human body, and the full 3D human shape estimation is ultimately converted as an image inpainting problem. We propose a novel two-branch network architecture to train an end-to-end regressor via the latent feature supervision, which also includes a novel saliency map sub-net to extract the human information from object-occluded color images. To supervise the network training, we further build a novel dataset named as 3DOH50K. Several experiments are conducted to reveal the effectiveness of the proposed method. Experimental results demonstrate that the proposed method achieves the state-of-the-art comparing with previous methods. The dataset, codes are publicly available at https://www.yangangwang.com.

Articulated motion analysis often utilizes strong prior knowledge such as a known or trained parts model for humans. Yet, the world contains a variety of articulating objects--mammals, insects, mechanized structures--where the number and configuration of parts for a particular object is unknown in advance. Here, we relax such strong assumptions via an unsupervised, Bayesian nonparametric parts model that infers an unknown number of parts with motions coupled by a body dynamic and parameterized by SE(D), the Lie group of rigid transformations. We derive an inference procedure that utilizes short observation sequences (image, depth, point cloud or mesh) of an object in motion without need for markers or learned body models. Efficient Gibbs decompositions for inference over distributions on SE(D) demonstrate robust part decompositions of moving objects under both 3D and 2D observation models. The inferred representation permits novel analysis, such as object segmentation by relative part motion, and transfers to new observations of the same object type.

We propose a joint albedo-normal approach to non-line-of-sight (NLOS) surface reconstruction using the directional light-cone transform (D-LCT). While current NLOS imaging methods reconstruct either the albedo or surface normals of the hidden scene, the two quantities provide complementary information of the scene, so an efficient method to estimate both simultaneously is desirable. We formulate the recovery of the two quantities as a vector deconvolution problem, and solve it via Cholesky-Wiener decomposition. We demonstrate that surfaces fitted non-parametrically using our recovered normals are more accurate than those produced with NLOS surface reconstruction methods recently proposed, and are 1,000 times faster to compute than using inverse rendering.

Deciding the amount of neurons during the design of a deep neural network to maximize performance is not intuitive. In this work, we attempt to search for the neuron (filter) configuration of a fixed network architecture that maximizes accuracy. Using iterative pruning methods as a proxy, we parametrize the change of the neuron (filter) number of each layer with respect to the change in parameters, allowing us to efficiently scale an architecture across arbitrary sizes. We also introduce architecture descent which iteratively refines the parametrized function used for model scaling. The combination of both proposed methods is coined as NeuralScale. To prove the efficiency of NeuralScale in terms of parameters, we show empirical simulations on VGG11, MobileNetV2 and ResNet18 using CIFAR10, CIFAR100 and TinyImageNet as benchmark datasets. Our results show an increase in accuracy of 3.04%, 8.56% and 3.41% for VGG11, MobileNetV2 and ResNet18 on CIFAR10, CIFAR100 and TinyImageNet respectively under a parameter-constrained setting (output neurons (filters) of default configuration with scaling factor of 0.25).

We study how to train a student deep neural network for visual recognition by distilling knowledge from a blackbox teacher model in a data-efficient manner. Progress on this problem can significantly reduce the dependence on large-scale datasets for learning high-performing visual recognition models. There are two major challenges. One is that the number of queries into the teacher model should be minimized to save computational and/or financial costs. The other is that the number of images used for the knowledge distillation should be small; otherwise, it violates our expectation of reducing the dependence on large-scale datasets. To tackle these challenges, we propose an approach that blends mixup and active learning. The former effectively augments the few unlabeled images by a big pool of synthetic images sampled from the convex hull of the original images, and the latter actively chooses from the pool hard examples for the student neural network and query their labels from the teacher model. We validate our approach with extensive experiments.

Filter level pruning is an effective method to accelerate the inference speed of deep CNN models. Although numerous pruning algorithms have been proposed, there are still two open issues. The first problem is how to prune residual connections. We propose to prune both channels inside and outside the residual connections via a KL-divergence based criterion. The second issue is pruning with limited data. We observe an interesting phenomenon: directly pruning on a small dataset is usually worse than fine-tuning a small model which is pruned or trained from scratch on the large dataset. Knowledge distillation is an effective approach to compensate for the weakness of limited data. However, the logits of a teacher model may be noisy. In order to avoid the influence of label noise, we propose a label refinement approach to solve this problem. Experiments have demonstrated the effectiveness of our method (CURL, Compression Using Residual-connections and Limited-data). CURL significantly outperforms previous state-of-the-art methods on ImageNet. More importantly, when pruning on small datasets, CURL achieves comparable or much better performance than fine-tuning a pretrained small model.

Transfer learning has proven to be a successful technique to train deep learning models in the domains where little training data is available. The dominant approach is to pretrain a model on a large generic dataset such as ImageNet and finetune its weights on the target domain. However, in the new era of an ever-increasing number of massive datasets, selecting the relevant data for pretraining is a critical issue. We introduce Neural Data Server (NDS), a large-scale search engine for finding the most useful transfer learning data to the target domain. NDS consists of a dataserver which indexes several large popular image datasets, and aims to recommend data to a client, an end-user with a target application with its own small labeled dataset. The dataserver represents large datasets with a much more compact mixture-of-experts model, and employs it to perform data search in a series of dataserver-client transactions at a low computational cost. We show the effectiveness of NDS in various transfer learning scenarios, demonstrating state-of-the-art performance on several target datasets and tasks such as image classification, object detection and instance segmentation. Neural Data Server is available as a web-service at http://aidemo s.cs.toronto.edu/nds/.

We propose a novel learnable representation for detail preserving shape deformation. The goal of our method is to warp a source shape to match the general structure of a target shape, while preserving the surface details of the source. Our method extends a traditional cage-based deformation technique, where the source shape is enclosed by a coarse control mesh termed cage, and translations prescribed on the cage vertices are interpolated to any point on the source mesh via special weight functions. The use of this sparse cage scaffolding enables preserving surface details regardless of the shape's intricacy and topology. Our key contribution is a novel neural network architecture for predicting deformations by controlling the cage. We incorporate a differentiable cage-based deformation module in our architecture, and train our network end-to-end. Our method can be trained with common collections of 3D models in an unsupervised fashion, without any cage-specific annotations. We demonstrate the utility of our method for synthesizing shape variations and deformation transfer.

We propose a new method for cancer subtype classification from histopathological images, which can automatically detect tumor-specific features in a given whole slide image (WSI). The cancer subtype should be classified by referring to a WSI, i.e., a large-sized image (typically 40,000x40,000 pixels) of an entire pathological tissue slide, which consists of cancer and non-cancer portions. One difficulty arises from the high cost associated with annotating tumor regions in WSIs. Furthermore, both global and local image features must be extracted from the WSI by changing the magnifications of the image. In addition, the image features should be stably detected against the differences of staining conditions among the hospitals/specimens. In this paper, we develop a new CNN-based cancer subtype classification method by effectively combining multiple-instance, domain adversarial, and multi-scale learning frameworks in order to overcome these practical difficulties. When the proposed method was applied to malignant lymphoma subtype classifications of 196 cases collected from multiple hospitals, the classification performance was significantly better than the standard CNN or other conventional methods, and the accuracy compared favorably with that of standard pathologists.

Referring expression comprehension (REC) and segmentation (RES) are two highly-related tasks, which both aim at identifying the referent according to a natural language expression. In this paper, we propose a novel Multi-task Collaborative Network (MCN) to achieve a joint learning of REC and RES for the first time. In MCN, RES can help REC to achieve better language-vision alignment, while REC can help RES to better locate the referent. In addition, we address a key challenge in this multi-task setup, i.e., the prediction conflict, with two innovative designs namely, Consistency Energy Maximization (CEM) and Adaptive Soft Non-Located Suppression (ASNLS). Specifically, CEM enables REC and RES to focus on similar visual regions by maximizing the consistency energy between two tasks. ASNLS supresses the response of unrelated regions in RES based on the prediction of REC. To validate our model, we conduct extensive experiments on three benchmark datasets of REC and RES, i.e., RefCOCO, RefCOCO+ and RefCOCOg. The experimental results report the significant performance gains of MCN over all existing methods, i.e., up to +7.13% for REC and +11.50% for RES over SOTA, which well confirm the validity of our model for joint REC and RES learning.

Fine-grained action recognition datasets exhibit environmental bias, where multiple video sequences are captured from a limited number of environments. Training a model in one environment and deploying in another results in a drop in performance due to an unavoidable domain shift. Unsupervised Domain Adaptation (UDA) approaches have frequently utilised adversarial training between the source and target domains. However, these approaches have not explored the multi-modal nature of video within each domain. In this work we exploit the correspondence of modalities as a self-supervised alignment approach for UDA in addition to adversarial alignment (Fig. 1). We test our approach on three kitchens from the large-scale EPIC-Kitchens dataset, using two modalities commonly employed for action recognition: RGB and Optical Flow. We show that multi-modal self-supervision alone improves the performance over source-only training by 2.4% on average. We then combine adversarial training with multi-modal self-supervision, showing that our approach outperforms other UDA methods by 3%.

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.

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.