ICCV 2017 论文列表

In this paper, we propose a generative model, Temporal Generative Adversarial Nets (TGAN), which can learn a semantic representation of unlabeled videos, and is capable of generating videos. Unlike existing Generative Adversarial Nets (GAN)-based methods that generate videos with a single generator consisting of 3D deconvolutional layers, our model exploits two different types of generators: a temporal generator and an image generator. The temporal generator takes a single latent variable as input and outputs a set of latent variables, each of which corresponds to an image frame in a video. The image generator transforms a set of such latent variables into a video. To deal with instability in training of GAN with such advanced networks, we adopt a recently proposed model, Wasserstein GAN, and propose a novel method to train it stably in an end-to-end manner. The experimental results demonstrate the effectiveness of our methods.

In this paper, we investigate a weakly-supervised object detection framework. Most existing frameworks focus on using static images to learn object detectors. However, these detectors often fail to generalize to videos because of the existing domain shift. Therefore, we investigate learning these detectors directly from boring videos of daily activities. Instead of using bounding boxes, we explore the use of action descriptions as supervision since they are relatively easy to gather. A common issue, however, is that objects of interest that are not involved in human actions are often absent in global action descriptions known as "missing label". To tackle this problem, we propose a novel temporal dynamic graph Long Short-Term Memory network (TD- Graph LSTM). TD-Graph LSTM enables global temporal reasoning by constructing a dynamic graph that is based on temporal correlations of object proposals and spans the entire video. The missing label issue for each individual frame can thus be significantly alleviated by transferring knowledge across correlated objects proposals in the whole video. Extensive evaluations on a large-scale daily-life action dataset (i.e., Charades) demonstrates the superiority of our proposed method. We also release object bounding-box annotations for more than 5,000 frames in Charades. We believe this annotated data can also benefit other research on video-based object recognition in the future.

We present a Temporal Context Network (TCN) for precise temporal localization of human activities. Similar to the Faster-RCNN architecture, proposals are placed at equal intervals in a video which span multiple temporal scales. We propose a novel representation for ranking these proposals. Since pooling features only inside a segment is not sufficient to predict activity boundaries, we construct a representation which explicitly captures context around a proposal for ranking it. For each temporal segment inside a proposal, features are uniformly sampled at a pair of scales and are input to a temporal convolutional neural network for classification. After ranking proposals, non-maximum suppression is applied and classification is performed to obtain final detections. TCN outperforms state-of-the-art methods on the ActivityNet dataset and the THUMOS14 dataset.

Detecting actions in untrimmed videos is an important yet challenging task. In this paper, we present the structured segment network (SSN), a novel framework which models the temporal structure of each action instance via a structured temporal pyramid. On top of the pyramid, we further introduce a decomposed discriminative model comprising two classifiers, respectively for classifying actions and determining completeness. This allows the framework to effectively distinguish positive proposals from background or incomplete ones, thus leading to both accurate recognition and localization. These components are integrated into a unified network that can be efficiently trained in an end-to-end fashion. Additionally, a simple yet effective temporal action proposal scheme, dubbed temporal actionness grouping (TAG) is devised to generate high quality action proposals. On two challenging benchmarks, THUMOS'14 and ActivityNet, our method remarkably outperforms previous state-of-the-art methods, demonstrating superior accuracy and strong adaptivity in handling actions with various temporal structures.

Reconstruction of 3D environments is a problem that has been widely addressed in the literature. While many approaches exist to perform reconstruction, few of them take an active role in deciding where the next observations should come from. Furthermore, the problem of travelling from the camera's current position to the next, known as pathplanning, usually focuses on minimising path length. This approach is ill-suited for reconstruction applications, where learning about the environment is more valuable than speed of traversal. We present a novel Scenic Route Planner that selects paths which maximise information gain, both in terms of total map coverage and reconstruction accuracy. We also introduce a new type of collaborative behaviour into the planning stage called opportunistic collaboration, which allows sensors to switch between acting as independent Structure from Motion (SfM) agents or as a variable baseline stereo pair. We show that Scenic Planning enables similar performance to state-of-the-art batch approaches using less than 0.00027% of the possible stereo pairs (3% of the views). Comparison against length-based pathplanning approaches show that our approach produces more complete and more accurate maps with fewer frames. Finally, we demonstrate the Scenic Pathplanner's ability to generalise to live scenarios by mounting cameras on autonomous ground-based sensor platforms and exploring an environment.

We address the problem of Temporal Action Proposal (TAP) generation. This is an important problem, as fast extraction of semantically important (e.g. human actions) segments from untrimmed videos is an important step for large-scale video analysis. To tackle this problem, we propose a novel Temporal Unit Regression Network (TURN) model. There are two salient aspects of TURN: (1) TURN jointly predicts action proposals and refines the temporal boundaries by temporal coordinate regression with contextual information; (2) Fast computation is enabled by unit feature reuse: a long untrimmed video is decomposed into video units, which are reused as basic building blocks of temporal proposals. TURN outperforms the state-of-the-art methods under average recall (AR) by a large margin on THUMOS-14 and ActivityNet datasets, and runs over 900 frames per second (FPS) on a TITAN X GPU. We further apply TURN as a proposal generation stage for existing temporal action localization pipelines, and outperforms state-of-the-art performance on THUMOS-14 and ActivityNet.

Given a video clip, action proposal aims to quickly generate a number of spatio-temporal tubes that enclose candidate human activities. Recently, the regression-based object detectors and long-term recurrent convolutional network (LRCN) have demonstrated superior performance in human action detection and recognition. However, the regression-based detectors performs inference without considering the temporal context among neighboring frames, and the LRCN using global visual percepts lacks the capability to capture local temporal dynamics. In this paper, we present a novel framework called TORNADO for human action proposal detection in un-trimmed video clips. Specifically, we propose a spatial-temporal convolutional network that combines the advantages of regression-based detector and LRCN by empowering Convolutional LSTM with regression capability. Our approach consists of a temporal convolutional regression network (T-CRN) and a spatial regression network (S-CRN) which are trained end-to-end on both RGB and OpticalFlow streams. They fuse appearance, motion and temporal contexts to regress the bounding boxes of candidate human actions simultaneously in 28 FPS. The action proposals are constructed by solving dynamic programming with peak trimming of the generated action boxes. Extensive experiments on the challenging UCF-101 and UCF-Sports datasets show that our method achieves superior performance as compared with the state-of-the-arts.

This paper focuses on temporal localization of actions from untrimmed videos. Existing methods typically involve training classifiers for a pre-defined list of actions and applying the classifiers in a sliding window fashion. However, activities in the wild consist of a wide combination of actors, actions and objects; it is difficult to design a proper activity list that meets users' needs. We propose to localize activities by natural language queries. Temporal Activity Localization via Language (TALL) is challenging as it requires: (1) suitable design of text and video representations to allow cross-modal matching of actions and language queries; (2) ability to locate actions accurately given features from sliding windows of limited granularity. We propose a novel Cross-modal Temporal Regression Localizer (CTRL) to jointly model text query and video clips, output alignment scores and location regression results for candidate clips. For evaluation, we adopt TaCoS dataset, and build a new dataset for this task on top of Charades by adding sentence temporal annotations, called Charades-STA. Experimental results show that CTRL outperforms previous methods significantly on both datasets.

An open question in facial landmark localization in video is whether one should perform tracking or tracking-by-detection (i.e. face alignment). Tracking produces fittings of high accuracy but is prone to drifting. Tracking-by-detection is drift-free but results in low accuracy fittings. To provide a solution to this problem, we describe the very first, to the best of our knowledge, synergistic approach between detection (face alignment) and tracking which completely eliminates drifting from face tracking, and does not merely perform tracking-by-detection. Our first main contribution is to show that one can achieve this synergy between detection and tracking using a principled optimization framework based on the theory of Global Variable Consensus Optimization using ADMM; Our second contribution is to show how the proposed analytic framework can be integrated within state-of-the-art discriminative methods for face alignment and tracking based on cascaded regression and deeply learned features. Overall, we call our method Discriminative Global Consensus Model (DGCM). Our third contribution is to show that DGCM achieves large performance improvement over the currently best performing face tracking methods on the most challenging category of the 300-VW dataset.

This paper proposes an end-to-end learning framework for multiview stereopsis. We term the network SurfaceNet. It takes a set of images and their corresponding camera parameters as input and directly infers the 3D model. The key advantage of the framework is that both photo-consistency as well geometric relations of the surface structure can be directly learned for the purpose of multiview stereopsis in an end-to-end fashion. SurfaceNet is a fully 3D convolutional network which is achieved by encoding the camera parameters together with the images in a 3D voxel representation. We evaluate SurfaceNet on the large-scale DTU benchmark.

This work introduces a novel surface registration method based on foliation. A foliation decomposes the surface into a family of closed loops, such that the decomposition has local tensor product structure. By projecting each loop to a point, the surface is collapsed into a graph. Two homeomorphic surfaces with consistent foliations can be registered by first matching their foliation graphs, then matching the corresponding leaves. This foliation based method is capable of handling surfaces with complicated topologies and large non-isometric deformations, rigorous with solid theoretic foundation, easy to implement, robust to compute. The result mapping is diffeomorphic. Our experimental results show the efficiency and efficacy of the proposed method.

We study the problem of single-image depth estimation for images in the wild. We collect human annotated surface normals and use them to help train a neural network that directly predicts pixel-wise depth. We propose two novel loss functions for training with surface normal annotations. Experiments on NYU Depth, KITTI, and our own dataset demonstrate that our approach can significantly improve the quality of depth estimation in the wild.

Data diversity in terms of types, styles, as well as radiometric, exposure and texture conditions widely exists in training and test data of vision applications. However, learning in traditional neural networks (NNs) only tries to find a model with fixed parameters that optimize the average behavior over all inputs, without using data-specific properties. In this paper, we develop a meta-level NN (MLNN) model that learns meta-knowledge on data-specific properties of images during learning and that dynamically adapts its weights during application according to the properties of the images input. MLNN consists of two parts: the dynamic supplementary NN (SNN) that learns meta-information on each type of inputs, and the fixed base-level NN (BLNN) that incorporates the meta-information from SNN into its weights at run time to realize the generalization for each type of inputs. We verify our approach using over ten network architectures under various application scenarios and loss functions. In low-level vision applications on image super-resolution and denoising, MLNN has 0.1 0.3 dB improvements on PSNR, whereas for high-level image classification, MLNN has accuracy improvement of 0.4 0.6% for Cifar10 and 1.2 2.1% for ImageNet when compared to convolutional NNs (CNNs). Improvements are more pronounced as the scale or diversity of data is increased.

In light of the powerful learning capability of deep neural networks (DNNs), deep (convolutional) models have been built in recent years to address the task of salient object detection. Although training such deep saliency models can significantly improve the detection performance, it requires large-scale manual supervision in the form of pixel-level human annotation, which is highly labor-intensive and time-consuming. To address this problem, this paper makes the earliest effort to train a deep salient object detector without using any human annotation. The key insight is "supervision by fusion", i.e., generating useful supervisory signals from the fusion process of weak but fast unsupervised saliency models. Based on this insight, we combine an intra-image fusion stream and a inter-image fusion stream in the proposed framework to generate the learning curriculum and pseudo ground-truth for supervising the training of the deep salient object detector. Comprehensive experiments on four benchmark datasets demonstrate that our method can approach the same network trained with full supervision (within 2-5% performance gap) and, more encouragingly, even outperform a number of fully supervised state-of-the-art approaches.

We introduce a novel semi-supervised video segmentation approach based on an efficient video representation, called as "super-trajectory". Each super-trajectory corresponds to a group of compact trajectories that exhibit consistent motion patterns, similar appearance and close spatiotemporal relationships. We generate trajectories using a probabilistic model, which handles occlusions and drifts in a robust and natural way. To reliably group trajectories, we adopt a modified version of the density peaks based clustering algorithm that allows capturing rich spatiotemporal relations among trajectories in the clustering process. The presented video representation is discriminative enough to accurately propagate the initial annotations in the first frame onto the remaining video frames. Extensive experimental analysis on challenging benchmarks demonstrate our method is capable of distinguishing the target objects from complex backgrounds and even reidentifying them after occlusions.

A popular approach to training classifiers of new image classes is to use lower levels of a pre-trained feed-forward neural network and retrain only the top. Thus, most layers simply serve as highly nonlinear feature extractors. While these features were found useful for classifying a variety of scenes and objects, previous work also demonstrated unusual levels of sensitivity to the input especially for images which are veering too far away from the training distribution. This can lead to surprising results as an imperceptible change in an image can be enough to completely change the predicted class. This occurs in particular in applications involving personaldata, typically acquired with wearable cameras (e.g., visual lifelogs), where the problem is also made more complex by the dearth of new labeled training data that make supervised learning with deep models difficult. To alleviate these problems, in this paper we propose a new generative model that captures the feature distribution in new data. Its latent space then becomes more representative of the new data, while still retaining the generalization properties. In particular, we use constrained Markov walks over a counting grid for modeling image sequences, which not only yield good latent representations, but allow for excellent classification with only a handful of labeled training examples of the new scenes or objects, a scenario typical in lifelogging applications.

Drones equipped with cameras are emerging as a powerful tool for large-scale aerial 3D scanning, but existing automatic flight planners do not exploit all available information about the scene, and can therefore produce inaccurate and incomplete 3D models. We present an automatic method to generate drone trajectories, such that the imagery acquired during the flight will later produce a high-fidelity 3D model. Our method uses a coarse estimate of the scene geometry to plan camera trajectories that: (1) cover the scene as thoroughly as possible; (2) encourage observations of scene geometry from a diverse set of viewing angles; (3) avoid obstacles; and (4) respect a user-specified flight time budget. Our method relies on a mathematical model of scene coverage that exhibits an intuitive diminishing returns property known as submodularity. We leverage this property extensively to design a trajectory planning algorithm that reasons globally about the non-additive coverage reward obtained across a trajectory, jointly with the cost of traveling between views. We evaluate our method by using it to scan three large outdoor scenes, and we perform a quantitative evaluation using a photorealistic video game simulator.

In this work we show how sublabel-accurate multilabeling approaches can be derived by approximating a classical label-continuous convex relaxation of nonconvex free-discontinuity problems. This insight allows to extend these sublabel-accurate approaches from total variation to general convex and nonconvex regularizations. Furthermore, it leads to a systematic approach to the discretization of continuous convex relaxations. We study the relationship to existing discretizations and to discrete-continuous MRFs. Finally, we apply the proposed approach to obtain a sublabel-accurate and convex solution to the vectorial Mumford-Shah functional and show in several experiments that it leads to more precise solutions using fewer labels.

We propose a novel deep learning approach to solve simultaneous alignment and recognition problems (referred to as "Sequence-to-sequence" learning). We decompose the problem into a series of specialised expert systems referred to as SubUNets. The spatio-temporal relationships between these SubUNets are then modelled to solve the task, while remaining trainable end-to-end. The approach mimics human learning and educational techniques, and has a number of significant advantages. SubUNets allow us to inject domain-specific expert knowledge into the system regarding suitable intermediate representations. They also allow us to implicitly perform transfer learning between different interrelated tasks, which also allows us to exploit a wider range of more varied data sources. In our experiments we demonstrate that each of these properties serves to significantly improve the performance of the overarching recognition system, by better constraining the learning problem. The proposed techniques are demonstrated in the challenging domain of sign language recognition. We demonstrate state-of-the-art performance on hand-shape recognition outperforming previous techniques by more than 30%). Furthermore, we are able to obtain comparable sign recognition rates to previous research, without the need for an alignment step to segment out the signs for recognition.