ICCV 2019 Papers

We propose the l-net, which reconstructs hyperspectral images (e.g., with 24 spectral channels) from a single shot measurement. This task is usually termed snapshot compressive-spectral imaging (SCI), which enjoys low cost, low bandwidth and high-speed sensing rate via capturing the three-dimensional (3D) signal i.e., (x, y, l), using a 2D snapshot. Though proposed more than a decade ago, the poor quality and low-speed of reconstruction algorithms preclude wide applications of SCI. To address this challenge, in this paper, we develop a dual-stage generative model to reconstruct the desired 3D signal in SCI, dubbed l-net. Results on both simulation and real datasets demonstrate the significant advantages of l-net, which leads to >4dB improvement in PSNR for real-mask-in-the-loop simulation data compared to the current state-of-the-art. Furthermore, l-net can finish the reconstruction task within sub-seconds instead of hours taken by the most recently proposed DeSCI algorithm, thus speeding up the reconstruction >1000 times.

Person re-identification (re-ID) is the task of matching person images across camera views, which plays an important role in surveillance and security applications. Inspired by great progress of deep learning, deep re-ID models began to be popular and gained state-of-the-art performance. However, recent works found that deep neural networks (DNNs) are vulnerable to adversarial examples, posing potential threats to DNNs based applications. This phenomenon throws a serious question about whether deep re-ID based systems are vulnerable to adversarial attacks. In this paper, we take the first attempt to implement robust physical-world attacks against deep re-ID. We propose a novel attack algorithm, called advPattern, for generating adversarial patterns on clothes, which learns the variations of image pairs across cameras to pull closer the image features from the same camera, while pushing features from different cameras farther. By wearing our crafted "invisible cloak", an adversary can evade person search, or impersonate a target person to fool deep re-ID models in physical world. We evaluate the effectiveness of our transformable patterns on adversaries' clothes with Market1501 and our established PRCS dataset. The experimental results show that the rank-1 accuracy of re-ID models for matching the adversary decreases from 87.9% to 27.1% under Evading Attack. Furthermore, the adversary can impersonate a target person with 47.1% rank-1 accuracy and 67.9% mAP under Impersonation Attack. The results demonstrate that deep re-ID systems are vulnerable to our physical attacks.

Image emotion recognition attracts much attention in recent years due to its wide applications. It aims to classify the emotional response of humans, where candidate emotion categories are generally defined by specific psychological theories, such as Ekman's six basic emotions. However, with the development of psychological theories, emotion categories become increasingly diverse, fine-grained, and difficult to collect samples. In this paper, we investigate zero-shot learning (ZSL) problem in the emotion recognition task, which tries to recognize the new unseen emotions. Specifically, we propose a novel affective-structural embedding framework, utilizing mid-level semantic representation, i.e., adjective-noun pairs (ANP) features, to construct an affective embedding space. By doing this, the learned intermediate space can narrow the semantic gap between low-level visual and high-level semantic features. In addition, we introduce an affective adversarial constraint to retain the discriminative capacity of visual features and the affective structural information of semantic features during training process. Our method is evaluated on five widely used affective datasets and the perimental results show the proposed algorithm outperforms the state-of-the-art approaches.

How can we teach a robot to predict what will happen next for an activity it has never seen before? We address the problem of zero-shot anticipation by presenting a hierarchical model that generalizes instructional knowledge from large-scale text-corpora and transfers the knowledge to the visual domain. Given a portion of an instructional video, our model predicts coherent and plausible actions multiple steps into the future, all in rich natural language. To demonstrate the anticipation capabilities of our model, we introduce the Tasty Videos dataset, a collection of 2511 recipes for zero-shot learning, recognition and anticipation.

Saliency methods can aid understanding of deep neural networks. Recent years have witnessed many improvements to saliency methods, as well as new ways for evaluating them. In this paper, we 1) present a novel region-based attribution method, XRAI, that builds upon integrated gradients (Sundararajan et al. 2017), 2) introduce evaluation methods for empirically assessing the quality of image-based saliency maps (Performance Information Curves (PICs)), and 3) contribute an axiom-based sanity check for attribution methods. Through empirical experiments and example results, we show that XRAI produces better results than other saliency methods for common models and the ImageNet dataset.

Visual question answering is the task of returning the answer to a question about an image. A challenge is that different people often provide different answers to the same visual question. To our knowledge, this is the first work that aims to understand why. We propose a taxonomy of nine plausible reasons, and create two labelled datasets consisting of 45,000 visual questions indicating which reasons led to answer differences. We then propose a novel problem of predicting directly from a visual question which reasons will cause answer differences as well as a novel algorithm for this purpose. Experiments demonstrate the advantage of our approach over several related baselines on two diverse datasets. We publicly share the datasets and code at https://vizwiz.org.

Mirrors are everywhere in our daily lives. Existing computer vision systems do not consider mirrors, and hence may get confused by the reflected content inside a mirror, resulting in a severe performance degradation. However, separating the real content outside a mirror from the reflected content inside it is non-trivial. The key challenge is that mirrors typically reflect contents similar to their surroundings, making it very difficult to differentiate the two. In this paper, we present a novel method to segment mirrors from an input image. To the best of our knowledge, this is the first work to address the mirror segmentation problem with a computational approach. We make the following contributions. First, we construct a large-scale mirror dataset that contains mirror images with corresponding manually annotated masks. This dataset covers a variety of daily life scenes, and will be made publicly available for future research. Second, we propose a novel network, called MirrorNet, for mirror segmentation, by modeling both semantical and low-level color/texture discontinuities between the contents inside and outside of the mirrors. Third, we conduct extensive experiments to evaluate the proposed method, and show that it outperforms the carefully chosen baselines from the state-of-the-art detection and segmentation methods.

Scene Parsing is a crucial step to enable autonomous systems to understand and interact with their surroundings. Supervised deep learning methods have made great progress in solving scene parsing problems, however, come at the cost of laborious manual pixel-level annotation. Synthetic data as well as weak supervision have been investigated to alleviate this effort. Nonetheless, synthetically generated data still suffers from severe domain shift while weak labels often lack precision. Moreover, most existing works for weakly supervised scene parsing are limited to salient foreground objects. The aim of this work is hence twofold: Exploit synthetic data where feasible and integrate weak supervision where necessary. More concretely, we address this goal by utilizing depth as transfer domain because its synthetic-to-real discrepancy is much lower than for color. At the same time, we perform weak localization from easily obtainable image level labels and integrate both using a novel contour-based scheme. Our approach is implemented as a teacher-student learning framework to solve the transfer learning problem by generating a pseudo ground truth. Using only depth-based adaptation, this approach already outperforms previous transfer learning approaches on the popular indoor scene parsing SUN RGB-D dataset. Our proposed two-stage integration more than halves the gap towards fully supervised methods when compared to previous state-of-the-art in transfer learning.

There is active research targeting local image manipulations that can fool deep neural networks (DNNs) into producing incorrect results. This paper examines a type of global image manipulation that can produce similar adverse effects. Specifically, we explore how strong color casts caused by incorrectly applied computational color constancy - referred to as white balance (WB) in photography - negatively impact the performance of DNNs targeting image segmentation and classification. In addition, we discuss how existing image augmentation methods used to improve the robustness of DNNs are not well suited for modeling WB errors. To address this problem, a novel augmentation method is proposed that can emulate accurate color constancy degradation. We also explore pre-processing training and testing images with a recent WB correction algorithm to reduce the effects of incorrectly white-balanced images. We examine both augmentation and pre-processing strategies on different datasets and demonstrate notable improvements on the CIFAR-10, CIFAR-100, and ADE20K datasets.

We describe a latent approach that learns to detect actions in long sequences given training videos with only whole-video class labels. Our approach makes use of two innovations to attention-modeling in weakly-supervised learning. First, and most notably, our framework uses an attention model to extract both foreground and background frames who's appearance is explicitly modeled. Most prior work ignores the background, but we show that modeling it allows our system to learn a richer notions of actions and their temporal extents. Second, we combine bottom-up, class-agnostic attention modules with top-down, class-specific activation maps, using the latter as form of self-supervision for the former. Doing so allows our model to learn a more accurate model of attention without explicit temporal supervision. These modifications lead to 10% AP@IoU=0.5 improvement over existing systems on THUMOS14. Our proposed weakly-supervised system outperforms the recent state-of-the-art by at least 4.3% AP@IoU=0.5. Finally, we demonstrate that weakly-supervised learning can be used to aggressively scale-up learning to in-the-wild, uncurated Instagram videos (where relevant frames and videos are automatically selected through attentional processing). This allows our weakly supervised approach to even outperform fully-supervised methods for action detection at some overlap thresholds.

Weakly-supervised temporal action localization (WS-TAL) is a promising but challenging task with only video-level action categorical labels available during training. Without requiring temporal action boundary annotations in training data, WS-TAL could possibly exploit automatically retrieved video tags as video-level labels. However, such coarse video-level supervision inevitably incurs confusions, especially in untrimmed videos containing multiple action instances. To address this challenge, we propose the Contrast-based Localization EvaluAtioN Network (CleanNet) with our new action proposal evaluator, which provides pseudo-supervision by leveraging the temporal contrast in snippet-level action classification predictions. Essentially, the new action proposal evaluator enforces an additional temporal contrast constraint so that high-evaluation-score action proposals are more likely to coincide with true action instances. Moreover, the new action localization module is an integral part of CleanNet which enables end-to-end training. This is in contrast to many existing WS-TAL methods where action localization is merely a post-processing step. Experiments on THUMOS14 and ActivityNet datasets validate the efficacy of CleanNet against existing state-ofthe- art WS-TAL algorithms.

Weakly supervised object detection aims at learning precise object detectors, given image category labels. In recent prevailing works, this problem is generally formulated as a multiple instance learning module guided by an image classification loss. The object bounding box is assumed to be the one contributing most to the classification among all proposals. However, the region contributing most is also likely to be a crucial part or the supporting context of an object. To obtain a more accurate detector, in this work we propose a novel end-to-end weakly supervised detection approach, where a newly introduced generative adversarial segmentation module interacts with the conventional detection module in a collaborative loop. The collaboration mechanism takes full advantages of the complementary interpretations of the weakly supervised localization task, namely detection and segmentation tasks, forming a more comprehensive solution. Consequently, our method obtains more precise object bounding boxes, rather than parts or irrelevant surroundings. Expectedly, the proposed method achieves an accuracy of 53.7% on the PASCAL VOC 2007 dataset, outperforming the state-of-the-arts and demonstrating its superiority for weakly supervised object detection.

Multispectral pedestrian detection has shown great advantages under poor illumination conditions, since the thermal modality provides complementary information for the color image. However, real multispectral data suffers from the position shift problem, i.e. the color-thermal image pairs are not strictly aligned, making one object has different positions in different modalities. In deep learning based methods, this problem makes it difficult to fuse the feature maps from both modalities and puzzles the CNN training. In this paper, we propose a novel Aligned Region CNN (AR-CNN) to handle the weakly aligned multispectral data in an end-to-end way. Firstly, we design a Region Feature Alignment (RFA) module to capture the position shift and adaptively align the region features of the two modalities. Secondly, we present a new multimodal fusion method, which performs feature re-weighting to select more reliable features and suppress the useless ones. Besides, we propose a novel RoI jitter strategy to improve the robustness to unexpected shift patterns of different devices and system settings. Finally, since our method depends on a new kind of labelling: bounding boxes that match each modality, we manually relabel the KAIST dataset by locating bounding boxes in both modalities and building their relationships, providing a new KAIST-Paired Annotation. Extensive experimental validations on existing datasets are performed, demonstrating the effectiveness and robustness of the proposed method. Code and data are available at https://github.com/luzhang16/AR-CNN.

Multi-modal learning, particularly among imaging and linguistic modalities, has made amazing strides in many high-level fundamental visual understanding problems, ranging from language grounding to dense event captioning. However, much of the research has been limited to approaches that either do not take audio corresponding to video into account at all, or those that model the audio-visual correlations in service of sound or sound source localization. In this paper, we present the evidence, that audio signals can carry surprising amount of information when it comes to high-level visual-lingual tasks. Specifically, we focus on the problem of weakly-supervised dense event captioning in videos and show that audio on its own can nearly rival performance of a state-of-the-art visual model and, combined with video, can improve on the state-of-the-art performance. Extensive experiments on the ActivityNet Captions dataset show that our proposed multi-modal approach outperforms state-of-the-art unimodal methods, as well as validate specific feature representation and architecture design choices.

Wasserstein GANs are increasingly used in Computer Vision applications as they are easier to train. Previous WGAN variants mainly use the l_1 transport cost to compute the Wasserstein distance between the real and synthetic data distributions. The l_1 transport cost restricts the discriminator to be 1-Lipschitz. However, WGANs with l_1 transport cost were recently shown to not always converge. In this paper, we propose WGAN-QC, a WGAN with quadratic transport cost. Based on the quadratic transport cost, we propose an Optimal Transport Regularizer (OTR) to stabilize the training process of WGAN-QC. We prove that the objective of the discriminator during each generator update computes the exact quadratic Wasserstein distance between real and synthetic data distributions. We also prove that WGAN-QC converges to a local equilibrium point with finite discriminator updates per generator update. We show experimentally on a Dirac distribution that WGAN-QC converges, when many of the l_1 cost WGANs fail to [22]. Qualitative and quantitative results on the CelebA, CelebA-HQ, LSUN and the ImageNet dog datasets show that WGAN-QC is better than state-of-art GAN methods. WGAN-QC has much faster runtime than other WGAN variants.

We study on weakly-supervised object detection (WSOD) which plays a vital role in relieving human involvement from object-level annotations. Predominant works integrate region proposal mechanisms with convolutional neural networks (CNN). Although CNN is proficient in extracting discriminative local features, grand challenges still exist to measure the likelihood of a bounding box containing a complete object (i.e., "objectness"). In this paper, we propose a novel WSOD framework with Objectness Distillation (i.e., WSOD2) by designing a tailored training mechanism for weakly-supervised object detection. Multiple regression targets are specifically determined by jointly considering bottom-up (BU) and top-down (TD) objectness from low-level measurement and CNN confidences with an adaptive linear combination. As bounding box regression can facilitate a region proposal learning to approach its regression target with high objectness during training, deep objectness representation learned from bottom-up evidences can be gradually distilled into CNN by optimization. We explore different adaptive training curves for BU/TD objectness, and show that the proposed WSOD2 can achieve state-of-the-art results.

Relationships encode the interactions among individual instances and play a critical role in deep visual scene understanding. Suffering from the high predictability with non-visual information, relationship models tend to fit the statistical bias rather than "learning" to infer the relationships from images. To encourage further development in visual relationships, we propose a novel method to mine more valuable relationships by automatically pruning visually-irrelevant relationships. We construct a new scene graph dataset named Visually-Relevant Relationships Dataset (VrR-VG) based on Visual Genome. Compared with existing datasets, the performance gap between learnable and statistical method is more significant in VrR-VG, and frequency-based analysis does not work anymore. Moreover, we propose to learn a relationship-aware representation by jointly considering instances, attributes and relationships. By applying the representation-aware feature learned on VrR-VG, the performances of image captioning and visual question answering are systematically improved, which demonstrates the effectiveness of both our dataset and features embedding schema. Both our VrR-VG dataset and representation-aware features will be made publicly available soon.

In this paper, we propose a novel iterative multi-task framework to complete the segmentation mask of an occluded vehicle and recover the appearance of its invisible parts. In particular, firstly, to improve the quality of the segmentation completion, we present two coupled discriminators that introduce an auxiliary 3D model pool for sampling authentic silhouettes as adversarial samples. In addition, we propose a two-path structure with a shared network to enhance the appearance recovery capability. By iteratively performing the segmentation completion and the appearance recovery, the results will be progressively refined. To evaluate our method, we present a dataset, Occluded Vehicle dataset, containing synthetic and real-world occluded vehicle images. Based on this dataset, we conduct comparison experiments and demonstrate that our model outperforms the state-of-the-arts in both tasks of recovering segmentation mask and appearance for occluded vehicles. Moreover, we also demonstrate that our appearance recovery approach can benefit the occluded vehicle tracking in real-world videos.

Recently, convolutional neural networks (CNNs) have shown great success on the task of monocular depth estimation. A fundamental yet unanswered question is: how CNNs can infer depth from a single image. Toward answering this question, we consider visualization of inference of a CNN by identifying relevant pixels of an input image to depth estimation. We formulate it as an optimization problem of identifying the smallest number of image pixels from which the CNN can estimate a depth map with the minimum difference from the estimate from the entire image. To cope with a difficulty with optimization through a deep CNN, we propose to use another network to predict those relevant image pixels in a forward computation. In our experiments, we first show the effectiveness of this approach, and then apply it to different depth estimation networks on indoor and outdoor scene datasets. The results provide several findings that help exploration of the above question.