ICCV 2019 论文列表

Neural architecture search (NAS) aims to automate the search procedure of architecture instead of manual design. Even if recent NAS approaches finish the search within days, lengthy training is still required for a specific architecture candidate to get the parameters for its accurate evaluation. Recently one-shot NAS methods are proposed to largely squeeze the tedious training process by sharing parameters across candidates. In this way, the parameters for each candidate can be directly extracted from the shared parameters instead of training them from scratch. However, they have no sense of which candidate will perform better until evaluation so that the candidates to evaluate are randomly sampled and the top-1 candidate is considered the best. In this paper, we propose a Self-Evaluated Template Network (SETN) to improve the quality of the architecture candidates for evaluation so that it is more likely to cover competitive candidates. SETN consists of two components: (1) an evaluator, which learns to indicate the probability of each individual architecture being likely to have a lower validation loss. The candidates for evaluation can thus be selectively sampled according to this evaluator. (2) a template network, which shares parameters among all candidates to amortize the training cost of generated candidates. In experiments, the architecture found by SETN achieves the state-of-the-art performance on CIFAR and ImageNet benchmarks within comparable computation costs.

Modern deep neural networks are often vulnerable to adversarial samples. Based on the first optimization-based attacking method, many following methods are proposed to improve the attacking performance and speed. Recently, generation-based methods have received much attention since they directly use feed-forward networks to generate the adversarial samples, which avoid the time-consuming iterative attacking procedure in optimization-based and gradient-based methods. However, current generation-based methods are only able to attack one specific target (category) within one model, thus making them not applicable to real classification systems that often have hundreds/thousands of categories. In this paper, we propose the first Multi-target Adversarial Network (MAN), which can generate multi-target adversarial samples with a single model. By incorporating the specified category information into the intermediate features, it can attack any category of the target classification model during runtime. Experiments show that the proposed MAN can produce stronger attack results and also have better transferability than previous state-of-the-art methods in both multi-target attack task and single-target attack task. We further use the adversarial samples generated by our MAN to improve the robustness of the classification model. It can also achieve better classification accuracy than other methods when attacked by various methods.

Currently, most deep learning based disparity estimation methods have the problem of over-smoothing at boundaries, which is unfavorable for some applications such as point cloud segmentation, mapping, etc. To address this problem, we first analyze the potential causes and observe that the estimated disparity at edge boundary pixels usually follows multimodal distributions, causing over-smoothing estimation. Based on this observation, we propose a single-modal weighted average operation on the probability distribution during inference, which can alleviate the problem effectively. To integrate the constraint of this inference method into training stage, we further analyze the characteristics of different loss functions and found that using cross entropy with gaussian distribution consistently further improves the performance. For quantitative evaluation, we propose a novel metric that measures the disparity error in the local structure of edge boundaries. Experiments on various datasets using various networks show our method's effectiveness and general applicability. Code will be available at https://github.com/chenchr/otosp.

Adversarial representation learning is a promising paradigm for obtaining data representations that are invariant to certain sensitive attributes while retaining the information necessary for predicting target attributes. Existing approaches solve this problem through iterative adversarial minimax optimization and lack theoretical guarantees. In this paper, we first study the "linear" form of this problem i.e., the setting where all the players are linear functions. We show that the resulting optimization problem is both non-convex and non-differentiable. We obtain an exact closed-form expression for its global optima through spectral learning and provide performance guarantees in terms of analytical bounds on the achievable utility and invariance. We then extend this solution and analysis to non-linear functions through kernel representation. Numerical experiments on UCI, Extended Yale B and CIFAR-100 datasets indicate that, (a) practically, our solution is ideal for "imparting" provable invariance to any biased pre-trained data representation, and (b) the global optima of the "kernel" form can provide a comparable trade-off between utility and invariance in comparison to iterative minimax optimization of existing deep neural network based approaches, but with provable guarantees.

In this paper, we present a thorough evaluation of the efficacy of knowledge distillation and its dependence on student and teacher architectures. Starting with the observation that more accurate teachers often don't make good teachers, we attempt to tease apart the factors that affect knowledge distillation performance. We find crucially that larger models do not often make better teachers. We show that this is a consequence of mismatched capacity, and that small students are unable to mimic large teachers. We find typical ways of circumventing this (such as performing a sequence of knowledge distillation steps) to be ineffective. Finally, we show that this effect can be mitigated by stopping the teacher's training early. Our results generalize across datasets and models.

Robust machine learning is currently one of the most prominent topics which could potentially help shaping a future of advanced AI platforms that not only perform well in average cases but also in worst cases or adverse situations. Despite the long-term vision, however, existing studies on black-box adversarial attacks are still restricted to very specific settings of threat models (e.g., single distortion metric and restrictive assumption on target model's feedback to queries) and/or suffer from prohibitively high query complexity. To push for further advances in this field, we introduce a general framework based on an operator splitting method, the alternating direction method of multipliers (ADMM) to devise efficient, robust black-box attacks that work with various distortion metrics and feedback settings without incurring high query complexity. Due to the black-box nature of the threat model, the proposed ADMM solution framework is integrated with zeroth-order (ZO) optimization and Bayesian optimization (BO), and thus is applicable to the gradient-free regime. This results in two new black-box adversarial attack generation methods, ZO-ADMM and BO-ADMM. Our empirical evaluations on image classification datasets show that our proposed approaches have much lower function query complexities compared to state-of-the-art attack methods, but achieve very competitive attack success rates.

Over the past several years progress in designing better neural network architectures for visual recognition has been substantial. To help sustain this rate of progress, in this work we propose to reexamine the methodology for comparing network architectures. In particular, we introduce a new comparison paradigm of distribution estimates, in which network design spaces are compared by applying statistical techniques to populations of sampled models, while controlling for confounding factors like network complexity. Compared to current methodologies of comparing point and curve estimates of model families, distribution estimates paint a more complete picture of the entire design landscape. As a case study, we examine design spaces used in neural architecture search (NAS). We find significant statistical differences between recent NAS design space variants that have been largely overlooked. Furthermore, our analysis reveals that the design spaces for standard model families like ResNeXt can be comparable to the more complex ones used in recent NAS work. We hope these insights into distribution analysis will enable more robust progress toward discovering better networks for visual recognition.

The premise of training an accurate 3D human pose estimation network is the possession of huge amount of richly annotated training data. Nonetheless, manually obtaining rich and accurate annotations is, even not impossible, tedious and slow. In this paper, we propose to exploit monocular videos to complement the training dataset for the single-image 3D human pose estimation tasks. At the beginning, a baseline model is trained with a small set of annotations. By fixing some reliable estimations produced by the resulting model, our method automatically collects the annotations across the entire video as solving the 3D trajectory completion problem. Then, the baseline model is further trained with the collected annotations to learn the new poses. We evaluate our method on the broadly-adopted Human3.6M and MPI-INF-3DHP datasets. As illustrated in experiments, given only a small set of annotations, our method successfully makes the model to learn new poses from unlabelled monocular videos, promoting the accuracies of the baseline model by about 10%. By contrast with previous approaches, our method does not rely on either multi-view imagery or any explicit 2D keypoint annotations.

In this paper, we propose a novel end-to-end deep neural network model for omnidirectional depth estimation from a wide-baseline multi-view stereo setup. The images captured with ultra wide field-of-view (FOV) cameras on an omnidirectional rig are processed by the feature extraction module, and then the deep feature maps are warped onto the concentric spheres swept through all candidate depths using the calibrated camera parameters. The 3D encoder-decoder block takes the aligned feature volume to produce the omnidirectional depth estimate with regularization on uncertain regions utilizing the global context information. In addition, we present large-scale synthetic datasets for training and testing omnidirectional multi-view stereo algorithms. Our datasets consist of 11K ground-truth depth maps and 45K fisheye images in four orthogonal directions with various objects and environments. Experimental results show that the proposed method generates excellent results in both synthetic and real-world environments, and it outperforms the prior art and the omnidirectional versions of the state-of-the-art conventional stereo algorithms.

As an instance-level recognition problem, person re-identification (ReID) relies on discriminative features, which not only capture different spatial scales but also encapsulate an arbitrary combination of multiple scales. We callse features of both homogeneous and heterogeneous scales omni-scale features. In this paper, a novel deep ReID CNN is designed, termed Omni-Scale Network (OSNet), for omni-scale feature learning. This is achieved by designing a residual block composed of multiple convolutional feature streams, each detecting features at a certain scale. Importantly, a novel unified aggregation gate is introduced to dynamically fuse multi-scale features with input-dependent channel-wise weights. To efficiently learn spatial-channel correlations and avoid overfitting, the building block uses both pointwise and depthwise convolutions. By stacking such blocks layer-by-layer, our OSNet is extremely lightweight and can be trained from scratch on existing ReID benchmarks. Despite its small model size, our OSNet achieves state-of-the-art performance on six person-ReID datasets. Code and models are available at: https://github.com/KaiyangZhou/deep-person-reid.

Deep learning based 3D reconstruction techniques have recently achieved impressive results. However, while state-of-the-art methods are able to output complex 3D geometry, it is not clear how to extend these results to time-varying topologies. Approaches treating each time step individually lack continuity and exhibit slow inference, while traditional 4D reconstruction methods often utilize a template model or discretize the 4D space at fixed resolution. In this work, we present Occupancy Flow, a novel spatio-temporal representation of time-varying 3D geometry with implicit correspondences. Towards this goal, we learn a temporally and spatially continuous vector field which assigns a motion vector to every point in space and time. In order to perform dense 4D reconstruction from images or sparse point clouds, we combine our method with a continuous 3D representation. Implicitly, our model yields correspondences over time, thus enabling fast inference while providing a sound physical description of the temporal dynamics. We show that our method can be used for interpolation and reconstruction tasks, and demonstrate the accuracy of the learned correspondences. We believe that Occupancy Flow is a promising new 4D representation which will be useful for a variety of spatio-temporal reconstruction tasks.

Occlusion relationship reasoning demands closed contour to express the object, and orientation of each contour pixel to describe the order relationship between objects. Current CNN-based methods neglect two critical issues of the task: (1) simultaneous existence of the relevance and distinction for the two elements, i.e, occlusion edge and occlusion orientation; and (2) inadequate exploration to the orientation features. For the reasons above, we propose the Occlusion-shared and Feature-separated Network (OFNet). On one hand, considering the relevance between edge and orientation, two sub-networks are designed to share the occlusion cue. On the other hand, the whole network is split into two paths to learn the high semantic features separately. Moreover, a contextual feature for orientation prediction is extracted, which represents the bilateral cue of the foreground and background areas. The bilateral cue is then fused with the occlusion cue to precisely locate the object regions. Finally, a stripe convolution is designed to further aggregate features from surrounding scenes of the occlusion edge. The proposed OFNet remarkably advances the state-of-the-art approaches on PIOD and BSDS ownership dataset.

Occlusion is a key problem in 3D human pose estimation from a monocular video. To address this problem, we introduce an occlusion-aware deep-learning framework. By employing estimated 2D confidence heatmaps of keypoints and an optical-flow consistency constraint, we filter out the unreliable estimations of occluded keypoints. When occlusion occurs, we have incomplete 2D keypoints and feed them to our 2D and 3D temporal convolutional networks (2D and 3D TCNs) that enforce temporal smoothness to produce a complete 3D pose. By using incomplete 2D keypoints, instead of complete but incorrect ones, our networks are less affected by the error-prone estimations of occluded keypoints. Training the occlusion-aware 3D TCN requires pairs of a 3D pose and a 2D pose with occlusion labels. As no such a dataset is available, we introduce a "Cylinder Man Model" to approximate the occupation of body parts in 3D space. By projecting the model onto a 2D plane in different viewing angles, we obtain and label the occluded keypoints, providing us plenty of training data. In addition, we use this model to create a pose regularization constraint, preferring the 2D estimations of unreliable keypoints to be occluded. Our method outperforms state-of-the-art methods on Human 3.6M and HumanEva-I datasets.

Deep Convolutional Neural Networks (CNNs) have been pushing the frontier of face recognition over past years. However, existing CNN models are far less accurate when handling partially occluded faces. These general face models generalize poorly for occlusions on variable facial areas. Inspired by the fact that human visual system explicitly ignores the occlusion and only focuses on the non-occluded facial areas, we propose a mask learning strategy to find and discard corrupted feature elements from recognition. A mask dictionary is firstly established by exploiting the differences between the top conv features of occluded and occlusion-free face pairs using innovatively designed pairwise differential siamese network (PDSN). Each item of this dictionary captures the correspondence between occluded facial areas and corrupted feature elements, which is named Feature Discarding Mask (FDM). When dealing with a face image with random partial occlusions, we generate its FDM by combining relevant dictionary items and then multiply it with the original features to eliminate those corrupted feature elements from recognition. Comprehensive experiments on both synthesized and realistic occluded face datasets show that the proposed algorithm significantly outperforms the state-of-the-art systems.

In this paper, we introduce a new large-scale object detection dataset, Objects365, which has 365 object categories over 600K training images. More than 10 million, high-quality bounding boxes are manually labeled through a three-step, carefully designed annotation pipeline. It is the largest object detection dataset (with full annotation) so far and establishes a more challenging benchmark for the community. Objects365 can serve as a better feature learning dataset for localization-sensitive tasks like object detection and semantic segmentation. The Objects365 pre-trained models significantly outperform ImageNet pre-trained models with 5.6 points gain (42 vs 36.4) based on the standard setting of 90K iterations on COCO benchmark. Even compared with much long training time like 540K iterations, our Objects365 pretrained model with 90K iterations still have 2.7 points gain (42 vs 39.3). Meanwhile, the finetuning time can be greatly reduced (up to 10 times) when reaching the same accuracy. Better generalization ability of Object365 has also been verified on CityPersons, VOC segmentation, and ADE tasks. The dataset as well as the pretrained-models have been released at www.objects365.org.

We present a method that tackles the challenge of predicting color and depth behind the visible content of an image. Our approach aims at building up a Layered Depth Image (LDI) from a single RGB input, which is an efficient representation that arranges the scene in layers, including originally occluded regions. Unlike previous work, we enable an adaptive scheme for the number of layers and incorporate semantic encoding for better hallucination of partly occluded objects. Additionally, our approach is object-driven, which especially boosts the accuracy for the occluded intermediate objects. The framework consists of two steps. First, we individually complete each object in terms of color and depth, while estimating the scene layout. Second, we rebuild the scene based on the regressed layers and enforce the recomposed image to resemble the structure of the original input. The learned representation enables various applications, such as 3D photography and diminished reality, all from a single RGB image.

Video object detection is more challenging than image object detection because of the deteriorated frame quality. To enhance the feature representation, state-of-the-art methods propagate temporal information into the deteriorated frame by aligning and aggregating entire feature maps from multiple nearby frames. However, restricted by feature map's low storage-efficiency and vulnerable content-address allocation, long-term temporal information is not fully stressed by these methods. In this work, we propose the first object guided external memory network for online video object detection. Storage-efficiency is handled by object guided hard-attention to selectively store valuable features, and long-term information is protected when stored in an addressable external data matrix. A set of read/write operations are designed to accurately propagate/allocate and delete multi-level memory feature under object guidance. We evaluate our method on the ImageNet VID dataset and achieve state-of-the-art performance as well as good speed-accuracy tradeoff. Furthermore, by visualizing the external memory, we show the detailed object-level reasoning process across frames.

This paper proposes a novel noisy label detection approach, named O2U-net, for deep neural networks without human annotations. Different from prior work which requires specifically designed noise-robust loss functions or networks, O2U-net is easy to implement but effective. It only requires adjusting the hyper-parameters of the deep network to make its status transfer from overfitting to underfitting (O2U) cyclically. The losses of each sample are recorded during iterations. The higher the normalized average loss of a sample, the higher the probability of being noisy labels. O2U-net is naturally compatible with active learning and other human annotation approaches. This introduces extra flexibility for learning with noisy labels. We conduct sufficient experiments on multiple datasets in various settings. The experimental results prove the state-of-the-art of O2S-net.

Understanding proper distance measures between distributions is at the core of several learning tasks such as generative models, domain adaptation, clustering, etc. In this work, we focus on mixture distributions that arise naturally in several application domains where the data contains different sub-populations. For mixture distributions, established distance measures such as the Wasserstein distance do not take into account imbalanced mixture proportions. Thus, even if two mixture distributions have identical mixture components but different mixture proportions, the Wasserstein distance between them will be large. This often leads to undesired results in distance-based learning methods for mixture distributions. In this paper, we resolve this issue by introducing the Normalized Wasserstein measure. The key idea is to introduce mixture proportions as optimization variables, effectively normalizing mixture proportions in the Wasserstein formulation. Using the proposed normalized Wasserstein measure leads to significant performance gains for mixture distributions with imbalanced mixture proportions compared to the vanilla Wasserstein distance. We demonstrate the effectiveness of the proposed measure in GANs, domain adaptation and adversarial clustering in several benchmark datasets.