ICCV 2017 论文列表

Deep learning has been demonstrated to achieve excellent results for image classification and object detection. However, the impact of deep learning on video analysis (e.g. action detection and recognition) has been limited due to complexity of video data and lack of annotations. Previous convolutional neural networks (CNN) based video action detection approaches usually consist of two major steps: frame-level action proposal detection and association of proposals across frames. Also, these methods employ two-stream CNN framework to handle spatial and temporal feature separately. In this paper, we propose an end-to-end deep network called Tube Convolutional Neural Network (T-CNN) for action detection in videos. The proposed architecture is a unified network that is able to recognize and localize action based on 3D convolution features. A video is first divided into equal length clips and for each clip a set of tube proposals are generated next based on 3D Convolutional Network (ConvNet) features. Finally, the tube proposals of different clips are linked together employing network flow and spatio-temporal action detection is performed using these linked video proposals. Extensive experiments on several video datasets demonstrate the superior performance of T-CNN for classifying and localizing actions in both trimmed and untrimmed videos compared to state-of-the-arts.

In this paper, we propose a binary tree architecture to truncate architecture of wide networks by reducing the width of the networks. More precisely, in the proposed architecture, the width is incrementally reduced from lower layers to higher layers in order to increase the expressive capacity of networks with a less increase on parameter size. Also, in order to ease the gradient vanishing problem, features obtained at different layers are concatenated to form the output of our architecture. By employing the proposed architecture on a baseline wide network, we can construct and train a new network with same depth but considerably less number of parameters. In our experimental analyses, we observe that the proposed architecture enables us to obtain better parameter size and accuracy trade-off compared to baseline networks using various benchmark image classification datasets. The results show that our model can decrease the classification error of a baseline from 20.43% to 19.22% on Cifar-100 using only 28% of parameters that the baseline has.

Manual annotations of temporal bounds for object interactions (i.e. start and end times) are typical training input to recognition, localization and detection algorithms. For three publicly available egocentric datasets, we uncover inconsistencies in ground truth temporal bounds within and across annotators and datasets. We systematically assess the robustness of state-of-the-art approaches to changes in labeled temporal bounds, for object interaction recognition. As boundaries are trespassed, a drop of up to 10% is observed for both Improved Dense Trajectories and Two-Stream Convolutional Neural Network. We demonstrate that such disagreement stems from a limited understanding of the distinct phases of an action, and propose annotating based on the Rubicon Boundaries, inspired by a similarly named cognitive model, for consistent temporal bounds of object interactions. Evaluated on a public dataset, we report a 4% increase in overall accuracy, and an increase in accuracy for 55% of classes when Rubicon Boundaries are used for temporal annotations.

Learning visual representations with self-supervised learning has become popular in computer vision. The idea is to design auxiliary tasks where labels are free to obtain. Most of these tasks end up providing data to learn specific kinds of invariance useful for recognition. In this paper, we propose to exploit different self-supervised approaches to learn representations invariant to (i) inter-instance variations (two objects in the same class should have similar features) and (ii) intra-instance variations (viewpoint, pose, deformations, illumination, etc). Instead of combining two approaches with multi-task learning, we argue to organize and reason the data with multiple variations. Specifically, we propose to generate a graph with millions of objects mined from hundreds of thousands of videos. The objects are connected by two types of edges which correspond to two types of invariance: "different instances but a similar viewpoint and category" and "different viewpoints of the same instance". By applying simple transitivity on the graph with these edges, we can obtain pairs of images exhibiting richer visual invariance. We use this data to train a Triplet-Siamese network with VGG16 as the base architecture and apply the learned representations to different recognition tasks. For object detection, we achieve 63.2% mAP on PASCAL VOC 2007 using Fast R-CNN (compare to 67.3% with ImageNet pre-training). For the challenging COCO dataset, our method is surprisingly close (23.5%) to the ImageNet-supervised counterpart (24.4%) using the Faster R-CNN framework. We also show that our network can perform significantly better than the ImageNet network in the surface normal estimation task.

This paper addresses the problem of line pattern noise removal from a single image, such as rain streak, hyperspectral stripe and so on. Most of the previous methods model the line pattern noise in original image domain, which fail to explicitly exploit the directional characteristic, thus resulting in a redundant subspace with poor representation ability for those line pattern noise. To achieve a compact subspace for the line pattern structure, in this work, we incorporate a transformation into the image decomposition model so that maps the input image to a domain where the line pattern streak/stripe appearance has an extremely distinct low-rank structure, which naturally allows us to enforce a low-rank prior to extract the line pattern streak/stripe from the noisy image. Moreover, the random noise is usually mixed up with the line pattern noise, which makes the challenging problem much more difficult. While previous methods resort to the spectral or temporal correlation of the multi-images, we give a detailed analysis between the noisy and clean image in both local gradient and nonlocal domain, and propose a compositional directional total variational and low-rank prior for the image layer, thus to simultaneously accommodate both types of noise. The proposed method has been evaluated on two different tasks, including remote sensing image mixed random stripe noise removal and rain streak removal, all of which obtain very impressive performances.

Transferring objects from one place to another place is a common task performed by human in daily life. During this process, it is usually intuitive for humans to choose an object as a proper container and to use an efficient pose to carry objects; yet, it is non-trivial for current computer vision and machine learning algorithms. In this paper, we propose an approach to jointly infer container and human pose for transferring objects by minimizing the costs associated both object and pose candidates. Our approach predicts which object to choose as a container while reasoning about how humans interact with physical surroundings to accomplish the task of transferring objects given visual input. In the learning phase, the presented method learns how humans make rational choices of containers and poses for transferring different objects, as well as the physical quantities required by the transfer task (e.g., compatibility between container and containee, energy cost of carrying pose) via a structured learning approach. In the inference phase, given a scanned 3D scene with different object candidates and a dictionary of human poses, our approach infers the best object as a container together with human pose for transferring a given object.

In many computer vision tasks, for example saliency prediction or semantic segmentation, the desired output is a foreground map that predicts pixels where some criteria is satisfied. Despite the inherently spatial nature of this task commonly used learning objectives do not incorporate the spatial relationships between misclassified pixels and the underlying ground truth. The Weighted F-measure, a recently proposed evaluation metric, does reweight errors spatially, and has been shown to closely correlate with human evaluation of quality, and stably rank predictions with respect to noisy ground truths (such as a sloppy human annotator might generate). However it suffers from computational complexity which makes it intractable as an optimization objective for gradient descent, which must be evaluated thousands or millions of times while learning a model's parameters. We propose a differentiable and efficient approximation of this metric. By incorporating spatial information into the objective we can use a simpler model than competing methods without sacrificing accuracy, resulting in faster inference speeds and alleviating the need for pre/post-processing. We match (or improve) performance on several tasks compared to prior state of the art by traditional metrics, and in many cases significantly improve performance by the weighted F-measure.

The majority of existing solutions to the Multi-Target Tracking (MTT) problem do not combine cues over a long period of time in a coherent fashion. In this paper, we present an online method that encodes long-term temporal dependencies across multiple cues. One key challenge of tracking methods is to accurately track occluded targets or those which share similar appearance properties with surrounding objects. To address this challenge, we present a structure of Recurrent Neural Networks (RNN) that jointly reasons on multiple cues over a temporal window. Our method allows to correct data association errors and recover observations from occluded states. We demonstrate the robustness of our data-driven approach by tracking multiple targets using their appearance, motion, and even interactions. Our method outperforms previous works on multiple publicly available datasets including the challenging MOT benchmark.

We formulate tracking as an online decision-making process, where a tracking agent must follow an object despite ambiguous image frames and a limited computational budget. Crucially, the agent must decide where to look in the upcoming frames, when to reinitialize because it believes the target has been lost, and when to update its appearance model for the tracked object. Such decisions are typically made heuristically. Instead, we propose to learn an optimal decision-making policy by formulating tracking as a partially observable decision-making process (POMDP). We learn policies with deep reinforcement learning algorithms that need supervision (a reward signal) only when the track has gone awry. We demonstrate that sparse rewards allow us to quickly train on massive datasets, several orders of magnitude more than past work. Interestingly, by treating the data source of Internet videos as unlimited streams, we both learn and evaluate our trackers in a single, unified computational stream.

With an increasing number of users sharing information online, privacy implications entailing such actions are a major concern. For explicit content, such as user profile or GPS data, devices (e.g. mobile phones) as well as web services (e.g. facebook) offer to set privacy settings in order to enforce the users' privacy preferences. We propose the first approach that extends this concept to image content in the spirit of a Visual Privacy Advisor. First, we categorize personal information in images into 68 image attributes and collect a dataset, which allows us to train models that predict such information directly from images. Second, we run a user study to understand the privacy preferences of different users w.r.t. such attributes. Third, we propose models that predict user specific privacy score from images in order to enforce the users' privacy preferences. Our model is trained to predict the user specific privacy risk and even outperforms the judgment of the users, who often fail to follow their own privacy preferences on image data.

Compositional models represent visual patterns as hierarchies of meaningful and reusable parts. They are attractive to vision modeling due to their ability to decompose complex patterns into simpler ones and resolve the low-level ambiguities in high-level image interpretations. However, current compositional models separate structure and part discovery from parameter estimation, which generally leads to suboptimal learning and fitting of the model. Moreover, the commonly adopted latent structural learning is not scalable for deep architectures. To address these difficult issues for compositional models, this paper quests for a unified framework for compositional pattern modeling, inference and learning. Represented by And-Or graphs (AOGs), it jointly models the compositional structure, parts, features, and composition/sub-configuration relationships. We show that the inference algorithm of the proposed framework is equivalent to a feed-forward network. Thus, all the parameters can be learned efficiently via the highly-scalable back-propagation (BP) in an end-to-end fashion. We validate the model via the task of handwritten digit recognition. By visualizing the processes of bottom-up composition and top-down parsing, we show that our model is fully interpretable, being able to learn the hierarchical compositions from visual primitives to visual patterns at increasingly higher levels. We apply this new compositional model to natural scene character recognition and generic object detection. Experimental results have demonstrated its effectiveness.

Despite recent advances in face recognition using deep learning, severe accuracy drops are observed for large pose variations in unconstrained environments. Learning pose-invariant features is one solution, but needs expensively labeled large-scale data and carefully designed feature learning algorithms. In this work, we focus on frontalizing faces in the wild under various head poses, including extreme profile views. We propose a novel deep 3D Morphable Model (3DMM) conditioned Face Frontalization Generative Adversarial Network (GAN), termed as FF-GAN, to generate neutral head pose face images. Our framework differs from both traditional GANs and 3DMM based modeling. Incorporating 3DMM into the GAN structure provides shape and appearance priors for fast convergence with less training data, while also supporting end-to-end training. The 3DMM-conditioned GAN employs not only the discriminator and generator loss but also a new masked symmetry loss to retain visual quality under occlusions, besides an identity loss to recover high frequency information. Experiments on face recognition, landmark localization and 3D reconstruction consistently show the advantage of our frontalization method on faces in the wild datasets.

Recognizing how objects interact with each other is a crucial task in visual recognition. If we define the context of the interaction to be the objects involved, then most current methods can be categorized as either: (i) training a single classifier on the combination of the interaction and its context; or (ii) aiming to recognize the interaction independently of its explicit context. Both methods suffer limitations: the former scales poorly with the number of combinations and fails to generalize to unseen combinations, while the latter often leads to poor interaction recognition performance due to the difficulty of designing a context-independent interaction classifier. To mitigate those drawbacks, this paper proposes an alternative, context-aware interaction recognition framework. The key to our method is to explicitly construct an interaction classifier which combines the context, and the interaction. The context is encoded via word2vec into a semantic space, and is used to derive a classification result for the interaction. The proposed method still builds one classifier for one interaction (as per type (ii) above), but the classifier built is adaptive to context via weights which are context dependent. The benefit of using the semantic space is that it naturally leads to zero-shot generalizations in which semantically similar contexts (subject-object pairs) can be recognized as suitable contexts for an interaction, even if they were not observed in the training set. Our method also scales with the number of interaction-context pairs since our model parameters do not increase with the number of interactions. Thus our method avoids the limitation of both approaches. We demonstrate experimentally that the proposed framework leads to improved performance for all investigated interaction representations and datasets.

In this paper, we study the task of 3D human pose estimation in the wild. This task is challenging due to lack of training data, as existing datasets are either in the wild images with 2D pose or in the lab images with 3D pose. We propose a weakly-supervised transfer learning method that uses mixed 2D and 3D labels in a unified deep neutral network that presents two-stage cascaded structure. Our network augments a state-of-the-art 2D pose estimation sub-network with a 3D depth regression sub-network. Unlike previous two stage approaches that train the two sub-networks sequentially and separately, our training is end-to-end and fully exploits the correlation between the 2D pose and depth estimation sub-tasks. The deep features are better learnt through shared representations. In doing so, the 3D pose labels in controlled lab environments are transferred to in the wild images. In addition, we introduce a 3D geometric constraint to regularize the 3D pose prediction, which is effective in the absence of ground truth depth labels. Our method achieves competitive results on both 2D and 3D benchmarks.

Two types of information exist in a stereo pair: correlation (matching) and decorrelation (half-occlusion). Vision science has shown that both types of information are used in the visual cortex, and that people can perceive depth even when correlation cues are absent or very weak, a capability that remains absent from most computational stereo systems. As a step toward stereo algorithms that are more consistent with these perceptual phenomena, we re-examine the topic of scanline stereo as energy minimization. We represent a disparity profile as a piecewise smooth function with explicit breakpoints between its smooth pieces, and we show this allows correlation and decorrelation to be integrated into an objective that requires only two types of local information: the correlation and its spatial gradient. Experimentally, we show the global optimum of this objective matches human perception on a broad collection of wellknown perceptual stimuli, and that it also provides reasonable piecewise-smooth interpretations of depth in natural images, even without exploiting monocular boundary cues.

We propose an efficient and unified framework, namely ThiNet, to simultaneously accelerate and compress CNN models in both training and inference stages. We focus on the filter level pruning, i.e., the whole filter would be discarded if it is less important. Our method does not change the original network structure, thus it can be perfectly supported by any off-the-shelf deep learning libraries. We formally establish filter pruning as an optimization problem, and reveal that we need to prune filters based on statistics information computed from its next layer, not the current layer, which differentiates ThiNet from existing methods. Experimental results demonstrate the effectiveness of this strategy, which has advanced the state-of-the-art. We also show the performance of ThiNet on ILSVRC-12 benchmark. ThiNet achieves 3.31x FLOPs reduction and 16.63x compression on VGG-16, with only 0.52% top-5 accuracy drop. Similar experiments with ResNet-50 reveal that even for a compact network, ThiNet can also reduce more than half of the parameters and FLOPs, at the cost of roughly 1% top-5 accuracy drop. Moreover, the original VGG-16 model can be further pruned into a very small model with only 5.05MB model size, preserving AlexNet level accuracy but showing much stronger generalization ability.

Current approaches to video forecasting attempt to generate videos directly in pixel space using Generative Adversarial Networks (GANs) or Variational Autoencoders (VAEs). However, since these approaches try to model all the structure and scene dynamics at once, in unconstrained settings they often generate uninterpretable results. Our insight is that forecasting needs to be done first at a higher level of abstraction. Specifically, we exploit human pose detectors as a free source of supervision and break the video forecasting problem into two discrete steps. First we explicitly model the high level structure of active objects in the scene (humans) and use a VAE to model the possible future movements of humans in the pose space. We then use the future poses generated as conditional information to a GAN to predict the future frames of the video in pixel space. By using the structured space of pose as an intermediate representation, we sidestep the problems that GANs have in generating video pixels directly. We show through quantitative and qualitative evaluation that our method outperforms state-of-the-art methods for video prediction.

The Mapillary Vistas Dataset is a novel, large-scale street-level image dataset containing 25,000 high-resolution images annotated into 66 object categories with additional, instance-specific labels for 37 classes. Annotation is performed in a dense and fine-grained style by using polygons for delineating individual objects. Our dataset is 5x larger than the total amount of fine annotations for Cityscapes and contains images from all around the world, captured at various conditions regarding weather, season and daytime. Images come from different imaging devices (mobile phones, tablets, action cameras, professional capturing rigs) and differently experienced photographers. In such a way, our dataset has been designed and compiled to cover diversity, richness of detail and geographic extent. As default benchmark tasks, we define semantic image segmentation and instance-specific image segmentation, aiming to significantly further the development of state-of-the-art methods for visual road-scene understanding.

Neural networks trained on datasets such as ImageNet have led to major advances in visual object classification. One obstacle that prevents networks from reasoning more deeply about complex scenes and situations, and from integrating visual knowledge with natural language, like humans do, is their lack of common sense knowledge about the physical world. Videos, unlike still images, contain a wealth of detailed information about the physical world. However, most labelled video datasets represent high-level concepts rather than detailed physical aspects about actions and scenes. In this work, we describe our ongoing collection of the "something-something" database of video prediction tasks whose solutions require a common sense understanding of the depicted situation. The database currently contains more than 100,000 videos across 174 classes, which are defined as caption-templates. We also describe the challenges in crowd-sourcing this data at scale.