ICCV 2019 Papers

Object part parsing in the wild, which requires to simultaneously detect multiple object classes in the scene and accurately segments semantic parts within each class, is challenging for the joint presence of class-level and part-level ambiguities. Despite its importance, however, this problem is not sufficiently explored in existing works. In this paper, we propose a joint parsing framework with boundary and semantic awareness to address this challenging problem. To handle part-level ambiguity, a boundary awareness module is proposed to make mid-level features at multiple scales attend to part boundaries for accurate part localization, which are then fused with high-level features for effective part recognition. For class-level ambiguity, we further present a semantic awareness module that selects discriminative part features relevant to a category to prevent irrelevant features being merged together. The proposed modules are lightweight and implementation friendly, improving the performance substantially when plugged into various baseline architectures. Without bells and whistles, the full model sets new state-of-the-art results on the Pascal-Part dataset, in both multi-class and the conventional single-class setting, while running substantially faster than recent high-performance approaches.

Unsupervised feature learning for point clouds has been vital for large-scale point cloud understanding. Recent deep learning based methods depend on learning global geometry from self-reconstruction. However, these methods are still suffering from ineffective learning of local geometry, which significantly limits the discriminability of learned features. To resolve this issue, we propose MAP-VAE to enable the learning of global and local geometry by jointly leveraging global and local self-supervision. To enable effective local self-supervision, we introduce multi-angle analysis for point clouds. In a multi-angle scenario, we first split a point cloud into a front half and a back half from each angle, and then, train MAP-VAE to learn to predict a back half sequence from the corresponding front half sequence. MAP-VAE performs this half-to-half prediction using RNN to simultaneously learn each local geometry and the spatial relationship among them. In addition, MAP-VAE also learns global geometry via self-reconstruction, where we employ a variational constraint to facilitate novel shape generation. The outperforming results in four shape analysis tasks show that MAP-VAE can learn more discriminative global or local features than the state-of-the-art methods.

Video Recognition has drawn great research interest and great progress has been made. A suitable frame sampling strategy can improve the accuracy and efficiency of recognition. However, mainstream solutions generally adopt hand-crafted frame sampling strategies for recognition. It could degrade the performance, especially in untrimmed videos, due to the variation of frame-level saliency. To this end, we concentrate on improving untrimmed video classification via developing a learning-based frame sampling strategy. We intuitively formulate the frame sampling procedure as multiple parallel Markov decision processes, each of which aims at picking out a frame/clip by gradually adjusting an initial sampling. Then we propose to solve the problems with multi-agent reinforcement learning (MARL). Our MARL framework is composed of a novel RNN-based context-aware observation network which jointly models context information among nearby agents and historical states of a specific agent, a policy network which generates the probability distribution over a predefined action space at each step and a classification network for reward calculation as well as final recognition. Extensive experimental results show that our MARL-based scheme remarkably outperforms hand-crafted strategies with various 2D and 3D baseline methods. Our single RGB model achieves a comparable performance of ActivityNet v1.3 champion submission with multi-modal multi-model fusion and new state-of-the-art results on YouTube Birds and YouTube Cars.

Conventional object detection methods essentially suppose that the training and testing data are collected from a restricted target domain with expensive labeling cost. For alleviating the problem of domain dependency and cumbersome labeling, this paper proposes to detect objects in unrestricted environment by leveraging domain knowledge trained from an auxiliary source domain with sufficient labels. Specifically, we propose a multi-adversarial Faster-RCNN (MAF) framework for unrestricted object detection, which inherently addresses domain disparity minimization for domain adaptation in feature representation. The paper merits are in three-fold: 1) With the idea that object detectors often becomes domain incompatible when image distribution resulted domain disparity appears, we propose a hierarchical domain feature alignment module, in which multiple adversarial domain classifier submodules for layer-wise domain feature confusion are designed; 2) An information invariant scale reduction module (SRM) for hierarchical feature map resizing is proposed for promoting the training efficiency of adversarial domain adaptation; 3) In order to improve the domain adaptability, the aggregated proposal features with detection results are feed into a proposed weighted gradient reversal layer (WGRL) for characterizing hard confused domain samples. We evaluate our MAF on unrestricted tasks including Cityscapes, KITTI, Sim10k, etc. and the experiments show the state-of-the-art performance over the existing detectors.

Recently unsupervised learning of depth from videos has made remarkable progress and the results are comparable to fully supervised methods in outdoor scenes like KITTI. However, there still exist great challenges when directly applying this technology in indoor environments, e.g., large areas of non-texture regions like white wall, more complex ego-motion of handheld camera, transparent glasses and shiny objects. To overcome these problems, we propose a new optical-flow based training paradigm which reduces the difficulty of unsupervised learning by providing a clearer training target and handles the non-texture regions. Our experimental evaluation demonstrates that the result of our method is comparable to fully supervised methods on the NYU Depth V2 benchmark. To the best of our knowledge, this is the first quantitative result of purely unsupervised learning method reported on indoor datasets.

In this paper, we tackle the problem of 3D human shape estimation from single RGB images. While the recent progress in convolutional neural networks has allowed impressive results for 3D human pose estimation, estimating the full 3D shape of a person is still an open issue. Model-based approaches can output precise meshes of naked under-cloth human bodies but fail to estimate details and un-modelled elements such as hair or clothing. On the other hand, non-parametric volumetric approaches can potentially estimate complete shapes but, in practice, they are limited by the resolution of the output grid and cannot produce detailed estimates. In this work, we propose a non-parametric approach that employs a double depth map to represent the 3D shape of a person: a visible depth map and a "hidden" depth map are estimated and combined, to reconstruct the human 3D shape as done with a "mould". This representation through 2D depth maps allows a higher resolution output with a much lower dimension than voxel-based volumetric representations. Additionally, our fully derivable depth-based model allows us to efficiently incorporate a discriminator in an adversarial fashion to improve the accuracy and "humanness" of the 3D output. We train and quantitatively validate our approach on SURREAL and on 3D-HUMANS, a new photorealistic dataset made of semi-synthetic in-house videos annotated with 3D ground truth surfaces.

Video salient object detection aims at discovering the most visually distinctive objects in a video. How to effectively take object motion into consideration during video salient object detection is a critical issue. Existing state-of-the-art methods either do not explicitly model and harvest motion cues or ignore spatial contexts within optical flow images. In this paper, we develop a multi-task motion guided video salient object detection network, which learns to accomplish two sub-tasks using two sub-networks, one sub-network for salient object detection in still images and the other for motion saliency detection in optical flow images. We further introduce a series of novel motion guided attention modules, which utilize the motion saliency sub-network to attend and enhance the sub-network for still images. These two sub-networks learn to adapt to each other by end-to-end training. Experimental results demonstrate that the proposed method significantly outperforms existing state-of-the-art algorithms on a wide range of benchmarks. We hope our simple and effective approach will serve as a solid baseline and help ease future research in video salient object detection. Code and models will be made available.

Moire pattern is a common image quality degradation caused by frequency aliasing between monitors and cameras when taking screen-shot photos. The complex frequency distribution, imbalanced magnitude in colour channels, and diverse appearance attributes of moire pattern make its removal a challenging problem. In this paper, we propose a Moire pattern Removal Neural Network (MopNet) to solve this problem. All core components of MopNet are specially designed for unique properties of moire patterns, including the multi-scale feature aggregation addressing complex frequency, the channel-wise target edge predictor to exploit imbalanced magnitude among colour channels, and the attribute-aware classifier to characterize the diverse appearance for better modelling Moire patterns. Quantitative and qualitative experimental comparison validate the state-of-the-art performance of MopNet.

In this paper, we propose a novel and specially designed method for piecewise dense monocular depth estimation in dynamic scenes. We utilize spatial relations between neighboring superpixels to solve the inherent relative scale ambiguity (RSA) problem and smooth the depth map. However, directly estimating spatial relations is an ill-posed problem. Our core idea is to predict spatial relations based on the corresponding motion relations. Given two or more consecutive frames, we first compute semi-dense (CPM) or dense (optical flow) point matches between temporally neighboring images. Then we develop our method in four main stages: superpixel relations analysis, motion selection, reconstruction, and refinement. The final refinement process helps to improve the quality of the reconstruction at pixel level. Our method does not require per-object segmentation, template priors or training sets, which ensures flexibility in various applications. Extensive experiments on both synthetic and real datasets demonstrate that our method robustly handles different dynamic situations and presents competitive results to the state-of-the-art methods while running much faster than them.

We propose a method to produce a continuous stream of novel views under fine-grained (e.g., 1 degree step-size) camera control at interactive rates. A novel learning pipeline determines the output pixels directly from the source color. Injecting geometric transformations, including perspective projection, 3D rotation and translation into the network forces implicit reasoning about the underlying geometry. The latent 3D geometry representation is compact and meaningful under 3D transformation, being able to produce geometrically accurate views for both single objects and natural scenes. Our experiments show that both proposed components, the transforming encoder-decoder and depth-guided appearance mapping, lead to significantly improved generalization beyond the training views and in consequence to more accurate view synthesis under continuous 6-DoF camera control. Finally, we show that our method outperforms state-of-the-art baseline methods on public datasets.

Monocular 3D human-pose estimation from static images is a challenging problem, due to the curse of dimensionality and the ill-posed nature of lifting 2D-to-3D. In this paper, we propose a Deep Conditional Variational Autoencoder based model that synthesizes diverse anatomically plausible 3D-pose samples conditioned on the estimated 2D-pose. We show that CVAE-based 3D-pose sample set is consistent with the 2D-pose and helps tackling the inherent ambiguity in 2D-to-3D lifting. We propose two strategies for obtaining the final 3D pose- (a) depth-ordering/ordinal relations to score and weight-average the candidate 3D-poses, referred to as OrdinalScore, and (b) with supervision from an Oracle. We report close to state-of-the-art results on two benchmark datasets using OrdinalScore, and state-of-the-art results using the Oracle. We also show that our pipeline yields competitive results without paired image-to-3D annotations. The training and evaluation code is available at https://github.com/ssfootball04/generative_pose.

We tackle the fundamentally ill-posed problem of 3D human localization from monocular RGB images. Driven by the limitation of neural networks outputting point estimates, we address the ambiguity in the task by predicting confidence intervals through a loss function based on the Laplace distribution. Our architecture is a light-weight feed-forward neural network that predicts 3D locations and corresponding confidence intervals given 2D human poses. The design is particularly well suited for small training data, cross-dataset generalization, and real-time applications. Our experiments show that we (i) outperform state-of-the-art results on KITTI and nuScenes datasets, (ii) even outperform a stereo-based method for far-away pedestrians, and (iii) estimate meaningful confidence intervals. We further share insights on our model of uncertainty in cases of limited observations and out-of-distribution samples.

Existing 3D scene flow estimation methods provide the 3D geometry and 3D motion of a scene and gain a lot of interest, for example in the context of autonomous driving. These methods are traditionally based on a temporal series of stereo images. In this paper, we propose a novel monocular 3D scene flow estimation method, called Mono-SF. Mono-SF jointly estimates the 3D structure and motion of the scene by combining multi-view geometry and single-view depth information. Mono-SF considers that the scene flow should be consistent in terms of warping the reference image in the consecutive image based on the principles of multi-view geometry. For integrating single-view depth in a statistical manner, a convolutional neural network, called ProbDepthNet, is proposed. ProbDepthNet estimates pixel-wise depth distributions from a single image rather than single depth values. Additionally, as part of ProbDepthNet, a novel recalibration technique for regression problems is proposed to ensure well-calibrated distributions. Our experiments show that Mono-SF outperforms state-of-the-art monocular baselines and ablation studies support the Mono-SF approach and ProbDepthNet design.

Conventional unsupervised domain adaptation (UDA) assumes that training data are sampled from a single domain. This neglects the more practical scenario where training data are collected from multiple sources, requiring multi-source domain adaptation. We make three major contributions towards addressing this problem. First, we collect and annotate by far the largest UDA dataset, called DomainNet, which contains six domains and about 0.6 million images distributed among 345 categories, addressing the gap in data availability for multi-source UDA research. Second, we propose a new deep learning approach, Moment Matching for Multi-Source Domain Adaptation (M3SDA), which aims to transfer knowledge learned from multiple labeled source domains to an unlabeled target domain by dynamically aligning moments of their feature distributions. Third, we provide new theoretical insights specifically for moment matching approaches in both single and multiple source domain adaptation. Extensive experiments are conducted to demonstrate the power of our new dataset in benchmarking state-of-the-art multi-source domain adaptation methods, as well as the advantage of our proposed model. Dataset and Code are available at http://ai.bu.edu/M3SDA/

Recognizing visual unseen classes, i.e. for which no training data is available, is known as Zero Shot Learning (ZSL). Some of the best performing methods apply the triplet loss to seen classes to learn a mapping between visual representations of images and attribute vectors that constitute class prototypes. They nevertheless make several implicit assumptions that limit their performance on real use cases, particularly with fine-grained datasets comprising a large number of classes. We identify three of these assumptions and put forward corresponding novel contributions to address them. Our approach consists in taking into account both inter-class and intra-class relations, respectively by being more permissive with confusions between similar classes, and by penalizing visual samples which are atypical to their class. The approach is tested on four datasets, including the large-scale ImageNet, and exhibits performances significantly above recent methods, even generative methods based on more restrictive hypotheses.

Understanding images beyond salient actions involves reasoning about scene context, objects, and the roles they play in the captured event. Situation recognition has recently been introduced as the task of jointly reasoning about the verbs (actions) and a set of semantic-role and entity (noun) pairs in the form of action frames. Labeling an image with an action frame requires an assignment of values (nouns) to the roles based on the observed image content. Among the inherent challenges are the rich conditional structured dependencies between the output role assignments and the overall semantic sparsity. In this paper, we propose a novel mixture-kernel attention graph neural network (GNN) architecture designed to address these challenges. Our GNN enables dynamic graph structure during training and inference, through the use of a graph attention mechanism, and context-aware interactions between role pairs. We illustrate the efficacy of our model and design choices by conducting experiments on imSitu benchmark dataset, with accuracy improvements of up to 10% over the state-of-the-art.

Attention has become more attractive in person re-identification (ReID) as it is capable of biasing the allocation of available resources towards the most informative parts of an input signal. However, state-of-the-art works concentrate only on coarse or first-order attention design, e.g. spatial and channels attention, while rarely exploring higher-order attention mechanism. We take a step towards addressing this problem. In this paper, we first propose the High-Order Attention (HOA) module to model and utilize the complex and high-order statistics information in attention mechanism, so as to capture the subtle differences among pedestrians and to produce the discriminative attention proposals. Then, rethinking person ReID as a zero-shot learning problem, we propose the Mixed High-Order Attention Network (MHN) to further enhance the discrimination and richness of attention knowledge in an explicit manner. Extensive experiments have been conducted to validate the superiority of our MHN for person ReID over a wide variety of state-of-the-art methods on three large-scale datasets, including Market-1501, DukeMTMC-ReID and CUHK03-NP. Code is available at http://www.bhchen.cn.

A key challenge of infrared small object segmentation (ISOS) is to balance miss detection (MD) and false alarm (FA). This usually needs "opposite" strategies to suppress the two terms, and has not been well resolved in the literature. In this paper, we propose a deep adversarial learning framework to improve this situation. Departing from the tradition of jointly reducing MD and FA via a single objective, we decompose this difficult task into two sub-tasks handled by two models trained adversarially, with each focusing on reducing either MD or FA. Such a new design brings forth at least three advantages. First, as each model focuses on a relatively simpler sub-task, the overall difficulty of ISOS is somehow decreased. Second, the adversarial training of the two models naturally produces a delicate balance of MD and FA, and low rates for both MD and FA could be achieved at Nash equilibrium. Third, this MD-FA detachment gives us more flexibility to develop specific models dedicated to each sub-task. To realize the above design, we propose a conditional Generative Adversarial Network comprising of two generators and one discriminator. Each generator strives for one sub-task, while the discriminator differentiates the three segmentation results from the two generators and the ground truth. Moreover, in order to better serve the sub-tasks, the two generators, based on context aggregation networks, utilzse different size of receptive fields, providing both local and global views of objects for segmentation. As verified on multiple infrared image data sets, our method consistently achieves better segmentation than many state-of-the-art ISOS methods.

We consider the problem of detecting objects, as they come into view, from videos in an online fashion. We provide the first real-time solution that is guaranteed to minimize the delay, i.e., the time between when the object comes in view and the declared detection time, subject to acceptable levels of detection accuracy. The method leverages modern CNN-based object detectors that operate on a single frame, to aggregate detection results over frames to provide reliable detection at a rate, specified by the user, in guaranteed minimal delay. To do this, we formulate the problem as a Quickest Detection problem, which provides the aforementioned guarantees. We derive our algorithms from this theory. We show in experiments, that with an overhead of just 50 fps, we can increase the number of correct detections and decrease the overall computational cost compared to running a modern single-frame detector.