ICCV 2015 Papers

We propose a new approach for general object tracking with fully convolutional neural network. Instead of treating convolutional neural network (CNN) as a black-box feature extractor, we conduct in-depth study on the properties of CNN features offline pre-trained on massive image data and classification task on ImageNet. The discoveries motivate the design of our tracking system. It is found that convolutional layers in different levels characterize the target from different perspectives. A top layer encodes more semantic features and serves as a category detector, while a lower layer carries more discriminative information and can better separate the target from distracters with similar appearance. Both layers are jointly used with a switch mechanism during tracking. It is also found that for a tracking target, only a subset of neurons are relevant. A feature map selection method is developed to remove noisy and irrelevant feature maps, which can reduce computation redundancy and improve tracking accuracy. Extensive evaluation on the widely used tracking benchmark shows that the proposed tacker outperforms the state-of-the-art significantly.

Recently, exemplar based approaches have been successfully applied for face detection in the wild. Contrary to traditional approaches that model face variations from a large and diverse set of training examples, exemplar-based approaches use a collection of discriminatively trained exemplars for detection. In this paradigm, each exemplar casts a vote using retrieval framework and generalized Hough voting, to locate the faces in the target image. The advantage of this approach is that by having a large database that covers all possible variations, faces in challenging conditions can be detected without having to learn explicit models for different variations. Current schemes, however, make an assumption of independence between the visual words, ignoring their relations in the process. They also ignore the spatial consistency of the visual words. Consequently, every exemplar word contributes equally during voting regardless of its location. In this paper, we propose a novel approach that incorporates higher order information in the voting process. We discover visual phrases that contain semantically related visual words and exploit them for detection along with the visual words. For spatial consistency, we estimate the spatial distribution of visual words and phrases from the entire database and then weigh their occurrence in exemplars. This ensures that a visual word or a phrase in an exemplar makes a major contribution only if it occurs at its semantic location, thereby suppressing the noise significantly. We perform extensive experiments on standard FDDB, AFW and G-album datasets and show significant improvement over previous exemplar approaches.

In this paper, we introduce a new dataset consisting of 360,001 focused natural language descriptions for 10,738 images. This dataset, the Visual Madlibs dataset, is collected using automatically produced fill-in-the-blank templates designed to gather targeted descriptions about: people and objects, their appearances, activities, and interactions, as well as inferences about the general scene or its broader context. We provide several analyses of the Visual Madlibs dataset and demonstrate its applicability to two new description generation tasks: focused description generation, and multiple-choice question-answering for images. Experiments using joint-embedding and deep learning methods show promising results on these tasks.

We propose a new direction for fast video super-resolution (VideoSR) via a SR draft ensemble, which is defined as the set of high-resolution patch candidates before final image deconvolution. Our method contains two main components -- i.e., SR draft ensemble generation and its optimal reconstruction. The first component is to renovate traditional feedforward reconstruction pipeline and greatly enhance its ability to compute different super resolution results considering large motion variation and possible errors arising in this process. Then we combine SR drafts through the nonlinear process in a deep convolutional neural network (CNN). We analyze why this framework is proposed and explain its unique advantages compared to previous iterative methods to update different modules in passes. Promising experimental results are shown on natural video sequences.

As the use of videos is becoming more popular in computer vision, the need for annotated video datasets increases. Such datasets are required either as training data or simply as ground truth for benchmark datasets. A particular challenge in video segmentation is due to disocclusions, which hamper frame-to-frame propagation, in conjunction with non-moving objects. We show that a combination of motion from point trajectories, as known from motion segmentation, along with minimal supervision can largely help solve this problem. Moreover, we integrate a new constraint that enforces consistency of the color distribution in successive frames. We quantify user interaction effort with respect to segmentation quality on challenging ego motion videos. We compare our approach to a diverse set of algorithms in terms of user effort and in terms of performance on common video segmentation benchmarks.

A common video degradation problem, which is largely untreated in literature, is what we call Yin-Yang Phasing (YYP). YYP is characterized by involuntary, dramatic flip-flop in the intensity and possibly chromaticity of an object as the video plays. Such temporal artifacts occur under ill illumination conditions and are triggered by object or/and camera motions, which mislead the settings of camera's auto-exposure and white point. In this paper, we investigate the problem and propose a video restoration technique to suppress YYP artifacts and retain temporal consistency of objects appearance via inter-frame, spatially-adaptive, optimal tone mapping. The video quality can be further improved by a novel image enhancer designed in Weber's perception principle and by exploiting the second-order statistics of the scene. Experimental results are encouraging, pointing to an effective, practical solution for a common but surprisingly understudied problem.

We introduce a novel method of video matting via sparse and low-rank representation. Previous matting methods [10, 9] introduced a nonlocal prior to estimate the alpha matte and have achieved impressive results on some data. However, on one hand, searching inadequate or excessive samples may miss good samples or introduce noise; on the other hand, it is difficult to construct consistent nonlocal structures for pixels with similar features, yielding spatially and temporally inconsistent video mattes. In this paper, we proposed a novel video matting method to achieve spatially and temporally consistent matting result. Toward this end, a sparse and low-rank representation model is introduced to pursue consistent nonlocal structures for pixels with similar features. The sparse representation is used to adaptively select best samples and accurately construct the nonlocal structures for all pixels, while the low-rank representation is used to globally ensure consistent nonlocal structures for pixels with similar features. The two representations are combined to generate consistent video mattes. Experimental results show that our method has achieved high quality results in a variety of challenging examples featuring illumination changes, feature ambiguity, topology changes, transparency variation, dis-occlusion, fast motion and motion blur.

In this work we propose a novel approach to the problem of multi-view stereo reconstruction. Building upon the previously proposed PatchMatch stereo and PM-Huber algorithm we introduce an extension to the multi-view scenario that employs an iterative refinement scheme. Our proposed approach uses an extended and robustified volumetric truncated signed distance function representation, which is advantageous for the fusion of refined depth maps and also for raycasting the current reconstruction estimation together with estimated depth normals into arbitrary camera views. We formulate the combined multi-view stereo reconstruction and refinement as a variational optimization problem. The newly introduced plane based smoothing term in the energy formulation is guided by the current reconstruction confidence and the image contents. Further we propose an extension of the PatchMatch scheme with an additional KLT step to avoid unnecessary sampling iterations. Improper camera poses are corrected by a direct image aligment step that performs robust outlier compensation by means of a recently proposed kernel lifting framework. To speed up the optimization of the variational formulation an adapted scheme is used for faster convergence.

In this paper we propose a novel method for depth image superresolution which combines recent advances in example based upsampling with variational superresolution based on a known blur kernel. Most traditional depth superresolution approaches try to use additional high resolution intensity images as guidance for superresolution. In our method we learn a dictionary of edge priors from an external database of high and low resolution examples. In a novel variational sparse coding approach this dictionary is used to infer strong edge priors. Additionally to the traditional sparse coding constraints the difference in the overlap of neighboring edge patches is minimized in our optimization. These edge priors are used in a novel variational superresolution as anisotropic guidance of the higher order regularization. Both the sparse coding and the variational superresolution of the depth are solved based on a primal-dual formulation. In an exhaustive numerical and visual evaluation we show that our method clearly outperforms existing approaches on multiple real and synthetic datasets.

We propose the task of free-form and open-ended Visual Question Answering (VQA). Given an image and a natural language question about the image, the task is to provide an accurate natural language answer. Mirroring real-world scenarios, such as helping the visually impaired, both the questions and answers are open-ended. Visual questions selectively target different areas of an image, including background details and underlying context. As a result, a system that succeeds at VQA typically needs a more detailed understanding of the image and complex reasoning than a system producing generic image captions. Moreover, VQA is amenable to automatic evaluation, since many open-ended answers contain only a few words or a closed set of answers that can be provided in a multiple-choice format. We provide a dataset containing 0.25M images, 0.76M questions, and 10M answers (www.visualqa.org), and discuss the information it provides. Numerous baselines for VQA are provided and compared with human performance.

We address the problem of automatic extraction of foreground objects from videos. The goal is to provide a method for unsupervised collection of samples which can be further used for object detection training without any human intervention. We use the well known Selective Search approach to produce an initial still-image based segmentation of the video frames. This initial set of proposals is pruned and temporally extended using optical flow and transductive learning. Specifically, we propose to use Dense Trajectories in order to robustly match and track candidate boxes over different frames. The obtained box tracks are used to collect samples for unsupervised training of track-specific detectors. Finally, the detectors are run on the videos to extract the final tubes. The combination of appearance-based static ''objectness'' (Selective Search), motion information (Dense Trajectories) and transductive learning (detectors are forced to "overfit" on the unsupervised data used for training) makes the proposed approach extremely robust. We outperform state-of-the-art systems by a large margin on common benchmarks used for tube proposal evaluation.

This paper proposes a shrinkage-based framework for unsupervised trajectory clustering. Facing to the challenges of trajectory clustering, e.g., large variations within a cluster and ambiguities across clusters, we first introduce an adaptive multi-kernel-based estimation process to estimate the `shrunk' positions and speeds of trajectories' points. This kernel-based estimation effectively leverages both multiple structural information within a trajectory and the local motion patterns across multiple trajectories, such that the discrimination of the shrunk point can be properly increased. We further introduce a speed-regularized optimization process, which utilizes the estimated speeds to regularize the optimal shrunk points, so as to guarantee the smoothness and the discriminative pattern of the final shrunk trajectory. Using our approach, the variations among similar trajectories can be reduced while the boundaries between different clusters are enlarged. Experimental results demonstrate that our approach is superior to the state-of-art approaches on both clustering accuracy and robustness. Besides, additional experiments further reveal the effectiveness of our approach when applied to trajectory analysis applications such as anomaly detection.

People are inherently social. Social interaction plays an important and natural role in human behavior. Most computational methods focus on individuals alone rather than in social context. They also require labelled training data. We present an unsupervised approach to discover interpersonal synchrony, referred as to two or more persons preforming common actions in overlapping video frames or segments. For computational efficiency, we develop a branch-and-bound (B&B) approach that affords exhaustive search while guaranteeing a globally optimal solution. The proposed method is entirely general. It takes from two or more videos any multi-dimensional signal that can be represented as a histogram. We derive three novel bounding functions and provide efficient extensions, including multi-synchrony detection and accelerated search, using a warm-start strategy and parallelism. We evaluate the effectiveness of our approach in multiple databases, including human actions using the CMU Mocap dataset, spontaneous facial behaviors using group-formation task dataset and parent-infant interaction dataset.

Human communication typically has an underlying structure. This is reflected in the fact that in many user generated videos, a starting point, ending, and certain objective steps between these two can be identified. In this paper, we propose a method for parsing a video into such semantic steps in an unsupervised way. The proposed method is capable of providing a semantic ``storyline'' of the video composed of its objective steps. We accomplish this utilizing both visual and language cues in a joint generative model. The proposed method can also provide a textual description for each of identified semantic steps and video segments. We evaluate this method on a large number of complex YouTube videos and show results of unprecedented quality for this new and impactful problem.

This paper addresses the problem of automatically localizing dominant objects as spatio-temporal tubes in a noisy collection of videos with minimal or even no supervision. We formulate the problem as a combination of two complementary processes: discovery and tracking. The first one establishes correspondences between prominent regions across videos, and the second one associates similar object regions within the same video. Interestingly, our algorithm also discovers the implicit topology of frames associated with instances of the same object class across different videos, a role normally left to supervisory information in the form of class labels in conventional image and video understanding methods. Indeed, as demonstrated by our experiments, our method can handle video collections featuring multiple object classes, and substantially outperforms the state of the art in colocalization, even though it tackles a broader problem with much less supervision.

Is strong supervision necessary for learning a good visual representation? Do we really need millions of semantically-labeled images to train a Convolutional Neural Network (CNN)? In this paper, we present a simple yet surprisingly powerful approach for unsupervised learning of CNN. Specifically, we use hundreds of thousands of unlabeled videos from the web to learn visual representations. Our key idea is that visual tracking provides the supervision. That is, two patches connected by a track should have similar visual representation in deep feature space since they probably belong to same object or object part. We design a Siamese-triplet network with a ranking loss function to train this CNN representation. Without using a single image from ImageNet, just using 100K unlabeled videos and the VOC 2012 dataset, we train an ensemble of unsupervised networks that achieves 52% mAP (no bounding box regression). This performance comes tantalizingly close to its ImageNet-supervised counterpart, an ensemble which achieves a mAP of 54.4%. We also show that our unsupervised network can perform competitively in other tasks such as surface-normal estimation.

Current state-of-the-art classification and detection algorithms train deep convolutional networks using labeled data. In this work we study unsupervised feature learning with convolutional networks in the context of temporally coherent unlabeled data. We focus on feature learning from unlabeled video data, using the assumption that adjacent video frames contain semantically similar information. This assumption is exploited to train a convolutional pooling auto-encoder regularized by slowness and sparsity. We establish a connection between slow feature learning and metric learning. Using this connection we define "temporal coherence"--a criterion which can be used to select hyper-parameters automatically. In a transfer learning experiment, we show that the resulting encoder can be used to define a more semantically coherent metric without the use of labeled data.

Object recognition approaches have recently been extended to yield, aside of the object class output, also viewpoint or pose. Training such approaches typically requires additional viewpoint or keypoint annotation in the training data or, alternatively, synthetic CAD models. In this paper,we present an approach that creates a dataset of images annotated with bounding boxes and viewpoint labels in a fully automated manner from videos. We assume that the scene is static in order to reconstruct 3D surfaces via structure from motion. We automatically detect when the reconstruction fails and normalize for the viewpoint of the 3D models by aligning the reconstructed point clouds. Exemplarily for cars we show that we can expand a large dataset of annotated single images and obtain improved performance when training a viewpoint regressor on this joined dataset.

With the growing popularity of short-form video sharing platforms such as Instagram and Vine, there has been an increasing need for techniques that automatically extract highlights from video. Whereas prior works have approached this problem with heuristic rules or supervised learning, we present an unsupervised learning approach that takes advantage of the abundance of user-edited videos on social media websites such as YouTube. Based on the idea that the most significant sub-events within a video class are commonly present among edited videos while less interesting ones appear less frequently, we identify the significant sub-events via a robust recurrent auto-encoder trained on a collection of user-edited videos queried for each particular class of interest. The auto-encoder is trained using a proposed shrinking exponential loss function that makes it robust to noise in the web-crawled training data, and is configured with bidirectional long short term memory (LSTM) cells to better model the temporal structure of highlight segments. Different from supervised techniques, our method can infer highlights using only a set of downloaded edited videos, without also needing their pre-edited counterparts which are rarely available online. Extensive experiments indicate the promise of our proposed solution in this challenging unsupervised setting.