ICCV 2015 Papers

In recent years, contextual models that exploit maps have been shown to be very effective for many recognition and localization tasks. In this paper we propose to exploit aerial images in order to enhance freely available world maps. Towards this goal, we make use of OpenStreetMap and formulate the problem as the one of inference in a Markov random field parameterized in terms of the location of the road-segment centerlines as well as their width. This parameterization enables very efficient inference and returns only topologically correct roads. In particular, we can segment all OSM roads in the whole world in a single day using a small cluster of 10 computers. Importantly, our approach generalizes very well; it can be trained using only 1.5 km2 aerial imagery and produce very accurate results in any location across the globe. We demonstrate the effectiveness of our approach outperforming the state-of-the-art in two new benchmarks that we collect. We then show how our enhanced maps are beneficial for semantic segmentation of ground images.

Video segmentation is the task of grouping similar pixels in the spatio-temporal domain, and has become an important preprocessing step for subsequent video analysis. Most video segmentation and supervoxel methods output a hierarchy of segmentations, but while this provides useful multiscale information, it also adds difficulty in selecting the appropriate level for a task. In this work, we propose an efficient and robust video segmentation framework based on parametric graph partitioning (PGP), a fast, almost parameter free graph partitioning method that identifies and removes between-cluster edges to form node clusters. Apart from its computational efficiency, PGP performs clustering of the spatio-temporal volume without requiring a pre-specified cluster number or bandwidth parameters, thus making video segmentation more practical to use in applications. The PGP framework also allows processing sub-volumes, which further improves performance, contrary to other streaming video segmentation methods where sub-volume processing reduces performance. We evaluate the PGP method using the SegTrack v2 and Chen Xiph.org datasets, and show that it outperforms related state-of-the-art algorithms in 3D segmentation metrics and running time.

The estimation of the epipolar geometry of two cameras from image matches is a fundamental problem of computer vision with many applications. While the closely related problem of estimating relative pose of two different uncalibrated cameras with radial distortion is of particular importance, none of the previously published methods is suitable for practical applications. These solutions are either numerically unstable, sensitive to noise, based on a large number of point correspondences, or simply too slow for real-time applications. In this paper, we present a new efficient solution to this problem that uses 10 image correspondences. By manipulating ten input polynomial equations, we derive a degree 10 polynomial equation in one variable. The solutions to this equation are efficiently found using the Sturm sequences method. In the experiments, we show that the proposed solution is stable, noise resistant, and fast, and as such efficiently usable in a practical Structure-from-Motion pipeline.

Person re-identification is becoming a hot research topic due to its value in both machine learning research and video surveillance applications. For this challenging problem, distance metric learning is shown to be effective in matching person images. However, existing approaches either require a heavy computation due to the positive semidefinite (PSD) constraint, or ignore the PSD constraint and learn a free distance function that makes the learned metric potentially noisy. We argue that the PSD constraint provides a useful regularization to smooth the solution of the metric, and hence the learned metric is more robust than without the PSD constraint. Another problem with metric learning algorithms is that the number of positive sample pairs is very limited, and the learning process is largely dominated by the large amount of negative sample pairs. To address the above issues, we derive a logistic metric learning approach with the PSD constraint and an asymmetric sample weighting strategy. Besides, we successfully apply the accelerated proximal gradient approach to find a global minimum solution of the proposed formulation, with a convergence rate of O(1/t^2) where t is the number of iterations. The proposed algorithm termed MLAPG is shown to be computationally efficient and able to perform low rank selection. We applied the proposed method for person re-identification, achieving state-of-the-art performance on four challenging databases (VIPeR, QMUL GRID, CUHK Campus, and CUHK03), compared to existing metric learning methods as well as published results.

Formulations of the Image Decomposition Problem as a Multicut Problem (MP) w.r.t. a superpixel graph have received considerable attention. In contrast, instances of the MP w.r.t. a pixel grid graph have received little attention, firstly, because the MP is NP-hard and instances w.r.t. a pixel grid graph are hard to solve in practice, and, secondly, due to the lack of long-range terms in the objective function of the MP. We propose a generalization of the MP with long-range terms (LMP). We design and implement two efficient algorithms (primal feasible heuristics) for the MP and LMP which allow us to study instances of both problems w.r.t. the pixel grid graphs of the images in the BSDS-500 benchmark. The decompositions we obtain do not differ significantly from the state of the art, suggesting that the LMP is a competitive formulation of the Image Decomposition Problem. To demonstrate the generality of the LMP, we apply it also to the Mesh Decomposition Problem posed by the Princeton benchmark, obtaining state-of-the-art decompositions.

The prospect of neural reconstruction from Electron Microscopy (EM) images has been elucidated by the automatic segmentation algorithms. Although segmentation algorithms eliminate the necessity of tracing the neurons by hand, significant manual effort is still essential for correcting the mistakes they make. A considerable amount of human labor is also required for annotating groundtruth volumes for training the classifiers of a segmentation framework. It is critically important to diminish the dependence on human interaction in the overall reconstruction system. This study proposes a novel classifier training algorithm for EM segmentation aimed to reduce the amount of manual effort demanded by the groundtruth annotation and error refinement tasks. Instead of using an exhaustive pixel level groundtruth, an active learning algorithm is proposed for sparse labeling of pixel and boundaries of superpixels. Because over-segmentation errors are in general more tolerable and easier to correct than the under-segmentation errors, our algorithm is designed to prioritize minimization of false-merges over false-split mistakes. Our experiments on both 2D and 3D data suggest that the proposed method yields segmentation outputs that are more amenable to neural reconstruction than those of existing methods.

Dynamic textures (DTs) are video sequences with stationary properties, which exhibit repetitive patterns over space and time. This paper aims at investigating the sparse coding based approach to characterizing local DT patterns for recognition. Owing to the high dimensionality of DT sequences, existing dictionary learning algorithms are not suitable for our purpose due to their high computational costs as well as poor scalability. To overcome these obstacles, we proposed a structured tensor dictionary learning method for sparse coding, which learns a dictionary structured with orthogonality and separability. The proposed method is very fast and more scalable to high-dimensional data than the existing ones. In addition, based on the proposed dictionary learning method, a DT descriptor is developed, which has better adaptivity, discriminability and scalability than the existing approaches. These advantages are demonstrated by the experiments on multiple datasets.

To break down the geometry assumptions of traditional motion models (e.g., homography, affine), warping-based motion model recently becomes very popular and is adopted in many latest applications (e.g., image stitching, video stabilization). With high degrees of freedom, the accuracy of model heavily relies on data-terms (keypoint correspondences). In some low-texture environments (e.g., indoor) where keypoint feature is insufficient or unreliable, the warping model is often erroneously estimated. In this paper we propose a simple and effective approach by considering both keypoint and line segment correspondences as data-term. Line segment is a prominent feature in artificial environments and it can supply sufficient geometrical and structural information of scenes, which not only helps guild to a correct warp in low-texture condition, but also prevents the undesired distortion induced by warping. The combination aims to complement each other and benefit for a wider range of scenes. Our method is general and can be ported to many existing applications. Experiments demonstrate that using dual-feature yields more robust and accurate result especially for those low-texture images.

The problem of domain generalization is to take knowledge acquired from a number of related domains, where training data is available, and to then successfully apply it to previously unseen domains. We propose a new feature learning algorithm, Multi-Task Autoencoder (MTAE), that provides good generalization performance for cross-domain object recognition. The algorithm extends the standard denoising autoencoder framework by substituting artificially induced corruption with naturally occurring inter-domain variability in the appearance of objects. Instead of reconstructing images from noisy versions, MTAE learns to transform the original image into analogs in multiple related domains. It thereby learns features that are robust to variations across domains. The learnt features are then used as inputs to a classifier. We evaluated the performance of the algorithm on benchmark image recognition datasets, where the task is to learn features from multiple datasets and to then predict the image label from unseen datasets. We found that (denoising) MTAE outperforms alternative autoencoder-based models as well as the current state-of-the-art algorithms for domain generalization.

Pose invariant matching is a very important and challenging problem with various applications like recognizing faces in uncontrolled scenarios, matching objects taken from different view points, etc. In this paper, we propose a discriminative pose-free descriptor (DPFD) which can be used to match faces/objects across pose variations. Training examples at very few representative poses are used to generate virtual intermediate pose subspaces. An image or image region is then represented by a feature set obtained by projecting it on all these subspaces and a discriminative transform is applied on this feature set to make it suitable for classification tasks. Finally, this discriminative feature set is represented by a single feature vector, termed as DPFD. The DPFD of images taken from different viewpoints can be directly compared for matching. Extensive experiments on recognizing faces across pose, pose and resolution on the Multi-PIE and Surveillance Cameras Face datasets and comparisons with state-of-the-art approaches show the effectiveness of the proposed approach. Experiments on matching general objects across viewpoints show the generalizability of the proposed approach beyond faces.

Good tracking performance is in general attributed to accurate representation over previously obtained targets or reliable discrimination between the target and the surrounding background. In this work, we exploit the advantages of the both approaches to achieve a robust tracker. We construct a subspace to represent the target and the neighboring background, and simultaneously propagate their class labels via the learned subspace. Moreover, we propose a novel criterion to identify the target from numerous target candidates on each frame, which takes into account both discrimination reliability and representation accuracy. In addition, with the proposed criterion, the ambiguity in the class labels of the neighboring background samples, which often influences the reliability of discriminative tracking model, is effectively alleviated, while the training set is still kept small. Extensive experiments demonstrate that our tracker performs favourably against many other state-of-the-art trackers.

Deep learning has revolutionalized image-level tasks such as classification, but patch-level tasks, such as correspondence, still rely on hand-crafted features, e.g. SIFT. In this paper we use Convolutional Neural Networks (CNNs) to learn discriminant patch representations and in particular train a Siamese network with pairs of (non-)corresponding patches. We deal with the large number of potential pairs with the combination of a stochastic sampling of the training set and an aggressive mining strategy biased towards patches that are hard to classify. By using the L2 distance during both training and testing we develop 128-D descriptors whose euclidean distances reflect patch similarity, and which can be used as a drop-in replacement for any task involving SIFT. We demonstrate consistent performance gains over the state of the art, and generalize well against scaling and rotation, perspective transformation, non-rigid deformation, and illumination changes. Our descriptors are efficient to compute and amenable to modern GPUs, and are publicly available.

Graph matching is a fundamental problem in computer vision. In this paper, we propose a novel graph matching algorithm based on tabu search. The proposed method solves graph matching problem by casting it into an equivalent weighted maximum clique problem of the corresponding association graph, which we further penalize through introducing negative weights. Subsequent tabu search optimization allows for overcoming the convention of using positive weights. The method's distinct feature is that it utilizes the history of search to make more strategic decisions while looking for the optimal solution, thus effectively escaping local optima and in practice achieving superior results. The proposed method, unlike the existing algorithms, enables direct optimization in the original discrete space while encouraging rather than artificially enforcing hard one-to-one constraint, thus resulting in better solution. The experiments demonstrate the robustness of the algorithm in a variety of settings, presenting the state-of-the-art results. The code is available at http://cv.snu.ac.kr/research/ DTSGM/

In this paper we tackle the problem of capturing the dense, detailed 3D geometry of generic, complex non-rigid meshes using a single RGB-only commodity video camera and a direct approach. While robust and even real-time solutions exist to this problem if the observed scene is static, for non-rigid dense shape capture current systems are typically restricted to the use of complex multi-camera rigs, take advantage of the additional depth channel available in RGB-D cameras, or deal with specific shapes such as faces or planar surfaces. In contrast, our method makes use of a single RGB video as input; it can capture the deformations of generic shapes; and the depth estimation is dense, per-pixel and direct. We first compute a dense 3D template of the shape of the object, using a short rigid sequence, and subsequently perform online reconstruction of the non-rigid mesh as it evolves over time. Our energy optimization approach minimizes a robust photometric cost that simultaneously estimates the temporal correspondences and 3D deformations with respect to the template mesh. In our experimental evaluation we show a range of qualitative results on novel datasets; we compare against an existing method that requires multi-frame optical flow; and perform a quantitative evaluation against other template-based approaches on a ground truth dataset.

We introduce a new approach to intrinsic image decomposition, the task of decomposing a single image into albedo and shading components. Our strategy, which we term direct intrinsics, is to learn a convolutional neural network (CNN) that directly predicts output albedo and shading channels from an input RGB image patch. Direct intrinsics is a departure from classical techniques for intrinsic image decomposition, which typically rely on physically-motivated priors and graph-based inference algorithms. The large-scale synthetic ground-truth of the MPI Sintel dataset plays the key role in training direct intrinsics. We demonstrate results on both the synthetic images of Sintel and the real images of the classic MIT intrinsic image dataset. On Sintel, direct intrinsics, using only RGB input, outperforms all prior work, including methods that rely on RGB+Depth input. Direct intrinsics also generalizes across modalities; our Sintel-trained CNN produces quite reasonable decompositions on the real images of the MIT dataset. Our results indicate that the marriage of CNNs with synthetic training data may be a powerful new technique for tackling classic problems in computer vision.

The long short-term memory (LSTM) neural network is capable of processing complex sequential information since it utilizes special gating schemes for learning representations from long input sequences. It has the potential to model any time-series or sequential data, where the current hidden state has to be considered in the context of the past hidden states. This property makes LSTM an ideal choice to learn the complex dynamics of various actions. Unfortunately, the conventional LSTMs do not consider the impact of spatio-temporal dynamics corresponding to the given salient motion patterns, when they gate the information that ought to be memorized through time. To address this problem, we propose a differential gating scheme for the LSTM neural network, which emphasizes on the change in information gain caused by the salient motions between the successive frames. This change in information gain is quantified by Derivative of States (DoS), and thus the proposed LSTM model is termed as differential Recurrent Neural Network (dRNN). We demonstrate the effectiveness of the proposed model by automatically recognizing actions from the real-world 2D and 3D human action datasets. Our study is one of the first works towards demonstrating the potential of learning complex time-series representations via high-order derivatives of states.

Symmetry, as one of the key components of Gestalt theory, provides an important mid-level cue that serves as input to higher visual processes such as segmentation. In this work, we propose a complete approach that links the detection of curved reflection symmetries to produce symmetry-constrained segments of structures/regions in real images with clutter. For curved reflection symmetry detection, we leverage on patch-based symmetric features to train a Structured Random Forest classifier that detects multiscaled curved symmetries in 2D images. Next, using these curved symmetries, we modulate a novel symmetry-constrained foreground-background segmentation by their symmetry scores so that we enforce global symmetrical consistency in the final segmentation. This is achieved by imposing a pairwise symmetry prior that encourages symmetric pixels to have the same labels over a MRF-based representation of the input image edges, and the final segmentation is obtained via graph-cuts. Experimental results over four publicly available datasets containing annotated symmetric structures: 1) SYMMAX-300, 2) BSD-Parts, 3) Weizmann Horse and 4) NY-roads demonstrate the approach's applicability to different environments with state-of-the-art performance.

We accurately estimate the 3D geometry and appearance of the human body from a monocular RGB-D sequence of a user moving freely in front of the sensor. Range data in each frame is first brought into alignment with a multi-resolution 3D body model in a coarse-to-fine process. The method then uses geometry and image texture over time to obtain accurate shape, pose, and appearance information despite unconstrained motion, partial views, varying resolution, occlusion, and soft tissue deformation. Our novel body model has variable shape detail, allowing it to capture faces with a high-resolution deformable head model and body shape with lower-resolution. Finally we combine range data from an entire sequence to estimate a high-resolution displacement map that captures fine shape details. We compare our recovered models with high-resolution scans from a professional system and with avatars created by a commercial product. We extract accurate 3D avatars from challenging motion sequences and even capture soft tissue dynamics.

Recent progress in using recurrent neural networks (RNNs) for image description has motivated the exploration of their application for video description. However, while images are static, working with videos requires modeling their dynamic temporal structure and then properly integrating that information into a natural language description model. In this context, we propose an approach that successfully takes into account both the local and global temporal structure of videos to produce descriptions. First, our approach incorporates a spatial temporal 3-D convolutional neural network (3-D CNN) representation of the short temporal dynamics. The 3-D CNN representation is trained on video action recognition tasks, so as to produce a representation that is tuned to human motion and behavior. Second we propose a temporal attention mechanism that allows to go beyond local temporal modeling and learns to automatically select the most relevant temporal segments given the text-generating RNN. Our approach exceeds the current state-of-art for both BLEU and METEOR metrics on the Youtube2Text dataset. We also present results on a new, larger and more challenging dataset of paired video and natural language descriptions.