CVPR 2019 论文列表

In this paper, we propose a new supervision objective named uniform loss to learn deep equidistributed representations for face recognition. Most existing methods aim to learn discriminative face features, encouraging large inter-class distances and small intra-class variations. However, they ignore the distribution of faces in the holistic feature space, which may lead to severe locality and unbalance. With the prior that faces lie on a hypersphere manifold, we impose an equidistributed constraint by uniformly spreading the class centers on the manifold, so that the minimum distance between class centers can be maximized through complete exploitation of the feature space. To this end, we consider the class centers as like charges on the surface of hypersphere with inter-class repulsion, and minimize the total electric potential energy as the uniform loss. Extensive experimental results on the MegaFace Challenge I, IARPA Janus Benchmark A (IJB-A), Youtube Faces (YTF) and Labeled Faces in the Wild (LFW) datasets show the effectiveness of the proposed uniform loss.

We propose the Unified Visual-Semantic Embeddings (Unified VSE) for learning a joint space of visual representation and textual semantics. The model unifies the embeddings of concepts at different levels: objects, attributes, relations, and full scenes. We view the sentential semantics as a combination of different semantic components such as objects and relations; their embeddings are aligned with different image regions. A contrastive learning approach is proposed for the effective learning of this fine-grained alignment from only image-caption pairs. We also present a simple yet effective approach that enforces the coverage of caption embeddings on the semantic components that appear in the sentence. We demonstrate that the Unified VSE outperforms baselines on cross-modal retrieval tasks; the enforcement of the semantic coverage improves the model's robustness in defending text-domain adversarial attacks. Moreover, our model empowers the use of visual cues to accurately resolve word dependencies in novel sentences.

Large-scale face datasets usually exhibit a massive number of classes, a long-tailed distribution, and severe label noise, which undoubtedly aggravate the difficulty of training. In this paper, we propose a training strategy that treats the head data and the tail data in an unequal way, accompanying with noise-robust loss functions, to take full advantage of their respective characteristics. Specifically, the unequal-training framework provides two training data streams: the first stream applies the head data to learn discriminative face representation supervised by Noise Resistance loss; the second stream applies the tail data to learn auxiliary information by gradually mining the stable discriminative information from confusing tail classes. Consequently, both training streams offer complementary information to deep feature learning. Extensive experiments have demonstrated the effectiveness of the new unequal-training framework and loss functions. Better yet, our method could save a significant amount of GPU memory. With our method, we achieve the best result on MegaFace Challenge 2 (MF2) given a large-scale noisy training data set.

Visual localization is the task of accurate camera pose estimation in a known scene. It is a key problem in computer vision and robotics, with applications including self-driving cars, Structure-from-Motion, SLAM, and Mixed Reality. Traditionally, the localization problem has been tackled using 3D geometry. Recently, end-to-end approaches based on convolutional neural networks have become popular. These methods learn to directly regress the camera pose from an input image. However, they do not achieve the same level of pose accuracy as 3D structure-based methods. To understand this behavior, we develop a theoretical model for camera pose regression. We use our model to predict failure cases for pose regression techniques and verify our predictions through experiments. We furthermore use our model to show that pose regression is more closely related to pose approximation via image retrieval than to accurate pose estimation via 3D structure. A key result is that current approaches do not consistently outperform a handcrafted image retrieval baseline. This clearly shows that additional research is needed before pose regression algorithms are ready to compete with structure-based methods.

This paper first answers the question "why do the two most powerful techniques Dropout and Batch Normalization (BN) often lead to a worse performance when they are combined together in many modern neural networks, but cooperate well sometimes as in Wide ResNet (WRN)?" in both theoretical and empirical aspects. Theoretically, we find that Dropout shifts the variance of a specific neural unit when we transfer the state of that network from training to test. However, BN maintains its statistical variance, which is accumulated from the entire learning procedure, in the test phase. The inconsistency of variances in Dropout and BN (we name this scheme "variance shift") causes the unstable numerical behavior in inference that leads to erroneous predictions finally. Meanwhile, the large feature dimension in WRN further reduces the "variance shift" to bring benefits to the overall performance. Thorough experiments on representative modern convolutional networks like DenseNet, ResNet, ResNeXt and Wide ResNet confirm our findings. According to the uncovered mechanism, we get better understandings in the combination of these two techniques and summarize guidelines for better practices.

Recently, data-driven deep saliency models have achieved high performance and have outperformed classical saliency models, as demonstrated by results on datasets such as the MIT300 and SALICON. Yet, there remains a large gap between the performance of these models and the inter-human baseline. Some outstanding questions include what have these models learned, how and where they fail, and how they can be improved. This article attempts to answer these questions by analyzing the representations learned by individual neurons located at the intermediate layers of deep saliency models. To this end, we follow the steps of existing deep saliency models, that is borrowing a pre-trained model of object recognition to encode the visual features and learning a decoder to infer the saliency. We consider two cases when the encoder is used as a fixed feature extractor and when it is fine-tuned, and compare the inner representations of the network. To study how the learned representations depend on the task, we fine-tune the same network using the same image set but for two different tasks: saliency prediction versus scene classification. Our analyses reveal that: 1) some visual regions (e.g. head, text, symbol, vehicle) are already encoded within various layers of the network pre-trained for object recognition, 2) using modern datasets, we find that fine-tuning pre-trained models for saliency prediction makes them favor some categories (e.g. head) over some others (e.g. text), 3) although deep models of saliency outperform classical models on natural images, the converse is true for synthetic stimuli (e.g. pop-out search arrays), an evidence of significant difference between human and data-driven saliency models, and 4) we confirm that, after-fine tuning, the change in inner-representations is mostly due to the task and not the domain shift in the data

This paper presents a new neural network for enhancing underexposed photos. Instead of directly learning an image-to-image mapping as previous work, we introduce intermediate illumination in our network to associate the input with expected enhancement result, which augments the network's capability to learn complex photographic adjustment from expert-retouched input/output image pairs. Based on this model, we formulate a loss function that adopts constraints and priors on the illumination, prepare a new dataset of 3,000 underexposed image pairs, and train the network to effectively learn a rich variety of adjustment for diverse lighting conditions. By these means, our network is able to recover clear details, distinct contrast, and natural color in the enhancement results. We perform extensive experiments on the benchmark MIT-Adobe FiveK dataset and our new dataset, and show that our network is effective to deal with previously challenging images.

Single image de-raining is an extremely challenging problem since the rainy image may contain rain streaks which may vary in size, direction and density. Previous approaches have attempted to address this problem by leveraging some prior information to remove rain streaks from a single image. One of the major limitations of these approaches is that they do not consider the location information of rain drops in the image. The proposed Uncertainty guided Multi-scale Residual Learning (UMRL) network attempts to address this issue by learning the rain content at different scales and using them to estimate the final de-rained output. In addition, we introduce a technique which guides the network to learn the network weights based on the confidence measure about the estimate. Furthermore, we introduce a new training and testing procedure based on the notion of cycle spinning to improve the final de-raining performance. Extensive experiments on synthetic and real datasets to demonstrate that the proposed method achieves significant improvements over the recent state-of-the-art methods.

In this paper, we propose UnOS, an unified system for unsupervised optical flow and stereo depth estimation using convolutional neural network (CNN) by taking advantages of their inherent geometrical consistency based on the rigid-scene assumption. UnOS significantly outperforms other state-of-the-art (SOTA) unsupervised approaches that treated the two tasks independently. Specifically, given two consecutive stereo image pairs from a video, UnOS estimates per-pixel stereo depth images, camera ego-motion and optical flow with three parallel CNNs. Based on these quantities, UnOS computes rigid optical flow and compares it against the optical flow estimated from the FlowNet, yielding pixels satisfying the rigid-scene assumption. Then, we encourage geometrical consistency between the two estimated flows within rigid regions, from which we derive a rigid-aware direct visual odometry (RDVO) module. We also propose rigid and occlusion-aware flow-consistency losses for the learning of UnOS. We evaluated our results on the popular KITTI dataset over 4 related tasks, i.e. stereo depth, optical flow, visual odometry and motion segmentation.

In this paper, we propose a unified panoptic segmentation network (UPSNet) for tackling the newly proposed panoptic segmentation task. On top of a single backbone residual network, we first design a deformable convolution based semantic segmentation head and a Mask R-CNN style instance segmentation head which solve these two subtasks simultaneously. More importantly, we introduce a parameter-free panoptic head which solves the panoptic segmentation via pixel-wise classification. It first leverages the logits from the previous two heads and then innovatively expands the representation for enabling prediction of an extra unknown class which helps better resolving the conflicts between semantic and instance segmentation. Besides, it handles the challenge caused by the varying number of instances and permits back propagation to the bottom modules in an end-to-end manner. Extensive experimental results on Cityscapes, COCO and our internal dataset demonstrate that our UPSNet achieves state-of-the-art performance with much faster inference. Code has been made available at: https://github.com/uber-research/UPSNet

Text effects transfer can dramatically make the text visually pleasing. In this paper, we present a novel framework to stylize the text with exquisite decor, which is ignored by the previous text stylization methods. Decorative elements pose a challenge to spontaneously handle basal text effects and decor, which are two different styles. To address this issue, our key idea is to learn to separate, transfer and recombine the decors and the basal text effect. A novel text effect transfer network is proposed to infer the styled version of the target text. The stylized text is finally embellished with decor where the placement of the decor is carefully determined by a novel structure-aware strategy. Furthermore, we propose a domain adaptation strategy for decor detection and a one-shot training strategy for text effects transfer, which greatly enhance the robustness of our network to new styles. We base our experiments on our collected topography dataset including 59,000 professionally styled text and demonstrate the superiority of our method over other state-of-the-art style transfer methods.

In skeleton-based action recognition, graph convolutional networks (GCNs), which model the human body skeletons as spatiotemporal graphs, have achieved remarkable performance. However, in existing GCN-based methods, the topology of the graph is set manually, and it is fixed over all layers and input samples. This may not be optimal for the hierarchical GCN and diverse samples in action recognition tasks. In addition, the second-order information (the lengths and directions of bones) of the skeleton data, which is naturally more informative and discriminative for action recognition, is rarely investigated in existing methods. In this work, we propose a novel two-stream adaptive graph convolutional network (2s-AGCN) for skeleton-based action recognition. The topology of the graph in our model can be either uniformly or individually learned by the BP algorithm in an end-to-end manner. This data-driven method increases the flexibility of the model for graph construction and brings more generality to adapt to various data samples. Moreover, a two-stream framework is proposed to model both the first-order and the second-order information simultaneously, which shows notable improvement for the recognition accuracy. Extensive experiments on the two large-scale datasets, NTU-RGBD and Kinetics-Skeleton, demonstrate that the performance of our model exceeds the state-of-the-art with a significant margin.

Collaboration is a necessary skill to perform tasks that are beyond one agent's capabilities. Addressed extensively in both conventional and modern AI, multi-agent collaboration has often been studied in the context of simple grid worlds. We argue that there are inherently visual aspects to collaboration which should be studied in visually rich environments. A key element in collaboration is communication that can be either explicit, through messages, or implicit, through perception of the other agents and the visual world. Learning to collaborate in a visual environment entails learning (1) to perform the task, (2) when and what to communicate, and (3) how to act based on these communications and the perception of the visual world. In this paper we study the problem of learning to collaborate directly from pixels in AI2-THOR and demonstrate the benefits of explicit and implicit modes of communication to perform visual tasks. Refer to our project page for more details: https://prior.allenai.org/projects/two-body-problem

Short-wave infrared (SWIR) imaging has a wide range of applications for both industry and civilian. However, the InGaAs sensors commonly used for SWIR imaging suffer from a variety of drawbacks, including high price, low resolution, unstable quality, and so on. In this paper, we propose a novel solution for SWIR imaging using a common Silicon sensor, which has cheaper price, higher resolution and better technical maturity compared with the specialized InGaAs sensor. Our key idea is to approximate the response of the InGaAs sensor by exploiting the largely ignored sensitivity of a Silicon sensor, weak as it is, in the SWIR range. To this end, we build a multi-channel optical system to collect a new SWIR dataset and present a physically meaningful three-stage image processing algorithm on the basis of CNN. Both qualitative and quantitative experiments show promising experimental results, which demonstrate the effectiveness of the proposed method.

Deep Neural Networks are quite vulnerable to adversarial perturbations. Current state-of-the-art adversarial attack methods typically require very time consuming hyper-parameter tuning, or require many iterations to solve an optimization based adversarial attack. To address this problem, we present a new family of trust region based adversarial attacks, with the goal of computing adversarial perturbations efficiently. We propose several attacks based on variants of the trust region optimization method. We test the proposed methods on Cifar-10 and ImageNet datasets using several different models including AlexNet, ResNet-50, VGG-16, and DenseNet-121 models. Our methods achieve comparable results with the Carlini-Wagner (CW) attack, but with significant speed up of up to 37x, for the VGG-16 model on a Titan Xp GPU. For the case of ResNet-50 on ImageNet, we can bring down its classification accuracy to less than 0.1% with at most 1.5% relative L_infinity (or L_2) perturbation requiring only 1.02 seconds as compared to 27.04 seconds for the CW attack. We have open sourced our method which can be accessed at [??].

Joint object detection and semantic segmentation is essential in many fields such as self-driving cars. An initial attempt towards this goal is to simply share a single network for multi-task learning. We argue that it does not make full use of the fact that detection and segmentation are mutually beneficial. In this paper, we propose a framework called TripleNet to deeply boost these two tasks. On the one hand, to deeply join the two tasks at different scales, triple supervisions including detection-oriented supervision and class-aware/agnostic segmentation supervisions are imposed on each layer of the decoder. Class-agnostic segmentation provides an objectness prior to detection and segmentation. On the other hand, to further intercross the two tasks and refine the features in each scale, two light-weight modules (i.e., the inner-connected module and the attention skip-layer fusion) are incorporated. Because segmentation supervision on each decoder layer are not performed at the test stage and two added modules are light-weight, the proposed TripleNet can run at a real-time speed (16 fps). Experiments on the VOC 2007/2012 and COCO datasets show that TripleNet outperforms all the other one-stage methods on both two tasks (e.g., 81.9% mAP and 83.3% mIoU on VOC 2012, and 37.1% mAP and 59.6% mIoU on COCO) by a single network.

In this paper, we study the problem of 3D object detection from stereo images, in which the key challenge is how to effectively utilize stereo information. Different from previous methods using pixel-level depth maps, we propose to employ 3D anchors to explicitly construct object-level correspondences between the regions of interest in stereo images, from which the deep neural network learns to detect and triangulate the targeted object in 3D space. We also introduce a cost-efficient channel reweighting strategy that enhances representational features and weakens noisy signals to facilitate the learning process. All of these are flexibly integrated into a solid baseline detector that inputs monocular images. We demonstrate that both the monocular baseline and the stereo triangulation learning network outperform the prior state-of-the-arts in 3D object detection and localization on the challenging KITTI dataset.

Cross-modal transfer is helpful to enhance modality-specific discriminative power for scene recognition. To this end, this paper presents a unified framework to integrate the tasks of cross-modal translation and modality-specific recognition, termed as Translate-to-Recognize Network TRecgNet. Specifically, both translation and recognition tasks share the same encoder network, which allows to explicitly regularize the training of recognition task with the help of translation, and thus improve its final generalization ability. For translation task, we place a decoder module on top of the encoder network and it is optimized with a new layer-wise semantic loss, while for recognition task, we use a linear classifier based on the feature embedding from encoder and its training is guided by the standard cross-entropy loss. In addition, our TRecgNet allows to exploit large numbers of unlabeled RGB-D data to train the translation task and thus improve the representation power of encoder network. Empirically, we verify that this new semi-supervised setting is able to further enhance the performance of recognition network. We perform experiments on two RGB-D scene recognition benchmarks: NYU Depth v2 and SUN RGB-D, demonstrating that TRecgNet achieves superior performance to the existing state-of-the-art methods, especially for recognition solely based on a single modality.

In this paper, we introduce a new idea for unsupervised domain adaptation via a remold of Prototypical Networks, which learn an embedding space and perform classification via a remold of the distances to the prototype of each class. Specifically, we present Transferrable Prototypical Networks (TPN) for adaptation such that the prototypes for each class in source and target domains are close in the embedding space and the score distributions predicted by prototypes separately on source and target data are similar. Technically, TPN initially matches each target example to the nearest prototype in the source domain and assigns an example a "pseudo" label. The prototype of each class could then be computed on source-only, target-only and source-target data, respectively. The optimization of TPN is end-to-end trained by jointly minimizing the distance across the prototypes on three types of data and KL-divergence of score distributions output by each pair of the prototypes. Extensive experiments are conducted on the transfers across MNIST, USPS and SVHN datasets, and superior results are reported when comparing to state-of-the-art approaches. More remarkably, we obtain an accuracy of 80.4% of single model on VisDA 2017 dataset.