ICCV 2015 论文列表

We propose the product quantization table (PQTable), a product quantization-based hash table that is fast and requires neither parameter tuning nor training steps. The PQTable produces exactly the same results as a linear PQ search, and is 10^2 to 10^5 times faster when tested on the SIFT1B data. In addition, although state-of-the-art performance can be achieved by previous inverted-indexing-based approaches, such methods do require manually designed parameter setting and much training, whereas our method is free from them. Therefore, PQTable offers a practical and useful solution for real-world problems.

There are many applications where multiple images are fused to form a single summary greyscale or colour output, including computational photography (e.g. RGB-NIR), diffusion tensor imaging (medical), and remote sensing. Often, and intuitively, image fusion is carried out in the derivative domain. Here, a new composite fused derivative is found that best accounts for the detail across all images and then the resulting gradient field is reintegrated. However, the reintegration step generally hallucinates new detail (not appearing in any of the input image bands) including halo and bending artifacts. In this paper we avoid these hallucinated details by avoiding the reintegration step. Our work builds directly on the work of Socolinsky and Wolff who derive their equivalent gradient field from the per-pixel Di Zenzo structure tensor which is defined as the inner product of the image Jacobian. We show that the x- and y- derivatives of the projection of the original image onto the Principal characteristic vector of the Outer Product (POP) of the Jacobian generates the same equivalent gradient field. In so doing, we have derived a fused image that has the derivative structure we seek. Of course, this projection will be meaningful only where the Jacobian has non-zero derivatives, so we diffuse the projection directions using a bilateral filter before we calculate the fused image. The resulting POP fused image has maximal fused detail but avoids hallucinated artifacts. Experiments demonstrate our method delivers state of the art image fusion performance.

Face alignment, especially on real-time or large-scale sequential images, is a challenging task with broad applications. Both generic and joint alignment approaches have been proposed with varying degrees of success. However, many generic methods are heavily sensitive to initializations and usually rely on offline-trained static models, which limit their performance on sequential images with extensive variations. On the other hand, joint methods are restricted to offline applications, since they require all frames to conduct batch alignment. To address these limitations, we propose to exploit incremental learning for personalized ensemble alignment. We sample multiple initial shapes to achieve image congealing within one frame, which enables us to incrementally conduct ensemble alignment by group-sparse regularized rank minimization. At the same time, personalized modeling is obtained by subspace adaptation under the same incremental framework, while correction strategy is used to alleviate model drifting. Experimental results on multiple controlled and in-the-wild databases demonstrate the superior performance of our approach compared with state-of-the-arts in terms of fitting accuracy and efficiency.

This work targets human action recognition in video. While recent methods typically represent actions by statistics of local video features, here we argue for the importance of a representation derived from human pose. To this end we propose a new Pose-based Convolutional Neural Network descriptor (P-CNN) for action recognition. The descriptor aggregates motion and appearance information along tracks of human body parts. We investigate different schemes of temporal aggregation and experiment with P-CNN features obtained both for automatically estimated and manually annotated human poses. We evaluate our method on the recent and challenging JHMDB and MPII Cooking datasets. For both datasets our method shows consistent improvement over the state of the art.

In this paper, we propose a new approach to generate oriented object proposals (OOPs) to reduce the detection error caused by various orientations of the object. To this end, we propose to efficiently locate object regions according to pixelwise object probability, rather than measuring the objectness from a set of sampled windows. We formulate the proposal generation problem as a generative probabilistic model such that object proposals of different shapes (i.e., sizes and orientations) can be produced by locating the local maximum likelihoods. The new approach has three main advantages. First, it helps the object detector handle objects of different orientations. Second, as the shapes of the proposals may vary to fit the objects, the resulting proposals are tighter than the sampling windows with fixed sizes. Third, it avoids massive window sampling, and thereby reducing the number of proposals while maintaining a high recall. Experiments on the PASCAL VOC 2007 dataset show that the proposed OOP outperforms the state-of-the-art fast methods. Further experiments show that the rotation invariant property helps a class-specific object detector achieve better performance than the state-of-the-art proposal generation methods in either object rotation scenarios or general scenarios. Generating OOPs is very fast and takes only 0.5s per image.

2D spatial image windows are used for comparing pixel values in computer vision applications such as correspondence for optical flow and 3D reconstruction, bilateral filtering, and image segmentation. However, pixel window comparisons can suffer from varying defocus blur and perspective at different depths, and can also lead to a loss of precision. In this paper, we leverage the recent use of light-field cameras to propose alternative - oriented light-field windows that enable more robust and accurate pixel comparisons. For Lambertian surfaces focused to the correct depth, the 2D distribution of angular rays from a pixel remains consistent. We build on this idea to develop an oriented 4D light-field window that accounts for shearing (depth), translation (matching), and windowing. Our main application is to scene flow, a generalization of optical flow to the 3D vector field describing the motion of each point in the scene. We show significant benefits of oriented light-field windows over standard 2D spatial windows. We also demonstrate additional applications of oriented light-field windows for bilateral filtering and image segmentation.

The viewing graph represents a set of views that are related by pairwise relative geometries. In the context of Structure-from-Motion (SfM), the viewing graph is the input to the incremental or global estimation pipeline. Much effort has been put towards developing robust algorithms to overcome potentially inaccurate relative geometries in the viewing graph during SfM. In this paper, we take a fundamentally different approach to SfM and instead focus on improving the quality of the viewing graph before applying SfM. Our main contribution is a novel optimization that improves the quality of the relative geometries in the viewing graph by enforcing loop consistency constraints with the epipolar point transfer. We show that this optimization greatly improves the accuracy of relative poses in the viewing graph and removes the need for filtering steps or robust algorithms typically used in global SfM methods. In addition, the optimized viewing graph can be used to efficiently calibrate cameras at scale. We combine our viewing graph optimization and focal length calibration into a global SfM pipeline that is more efficient than existing approaches. To our knowledge, ours is the first global SfM pipeline capable of handling uncalibrated image sets.

We study the question of how to make loss-aware predictions in image segmentation settings where the evaluation function is the Intersection-over-Union (IoU) measure that is used widely in evaluating image segmentation systems. Currently, there are two dominant approaches: the first approximates the Expected-IoU (EIoU) score as Expected-Intersection-over-Expected-Union (EIoEU); and the second approach is to compute exact EIoU but only over a small set of high-quality candidate solutions. We begin by asking which approach we should favor for two typical image segmentation tasks. Studying this question leads to two new methods that draw ideas from both existing approaches. Our new methods use the EIoEU approximation paired with high quality candidate solutions. Experimentally we show that our new approaches lead to improved performance on both image segmentation tasks.

We address the problem of hand pose estimation, formulated as an inverse problem. Typical approaches optimize an energy function over pose parameters using a `black box' image generation procedure. This procedure knows little about either the relationships between the parameters or the form of the energy function. In this paper, we show that we can significantly improving upon black box optimization by exploiting high-level knowledge of the structure of the parameters and using a local surrogate energy function. Our new framework, hierarchical sampling optimization, consists of a sequence of predictors organized into a kinematic hierarchy. Each predictor is conditioned on its ancestors, and generates a set of samples over a subset of the pose parameters. The highly-efficient surrogate energy is used to select among samples. Having evaluated the full hierarchy, the partial pose samples are concatenated to generate a full-pose hypothesis. Several hypotheses are generated using the same procedure, and finally the original full energy function selects the best result. Experimental evaluation on three publically available datasets show that our method is particularly impressive in low-compute scenarios where it significantly outperforms all other state-of-the-art methods.

Tracking-by-detection approaches are some of the most successful object trackers in recent years. Their success is largely determined by the detector model they learn initially and then update over time. However, under challenging conditions where an object can undergo transformations, e.g., severe rotation, these methods are found to be lacking. In this paper, we address this problem by formulating it as a proposal selection task and making two contributions. The first one is introducing novel proposals estimated from the geometric transformations undergone by the object, and building a rich candidate set for predicting the object location. The second one is devising a novel selection strategy using multiple cues, i.e., detection score and edgeness score computed from state-of-the-art object edges and motion boundaries. We extensively evaluate our approach on the visual object tracking 2014 challenge and online tracking benchmark datasets, and show the best performance.

The task of discriminating one object from another is almost trivial for a human being. However, this task is computationally taxing for most modern machine learning methods; whereas, we perform this task at ease given very few examples for learning. It has been proposed that the quick grasp of concept may come from the shared knowledge between the new example and examples previously learned. We believe that the key to one-shot learning is the sharing of common parts as each part holds immense amounts of information on how a visual concept is constructed. We propose an unsupervised method for learning a compact dictionary of image patches representing meaningful components of an objects. Using those patches as features, we build a compositional model that outperforms a number of popular algorithms on a one-shot learning task. We demonstrate the effectiveness of this approach on hand-written digits and show that this model generalizes to multiple datasets.

Is it possible to determine the visible subset of points directly from a given point cloud? Interestingly, in [7] it was shown that this is indeed the case - despite the fact that points cannot occlude each other, this task can be performed without surface reconstruction or normal estimation. The operator is very simple - it first transforms the points to a new domain and then constructs the convex hull in that domain. Points that lie on the convex hull of the transformed set of points are the images of the visible points. This operator found numerous applications in computer vision, including face reconstruction, keypoint detection, finding the best viewpoints, reduction of points, and many more. The current paper addresses a fundamental question: What properties should a transformation function satisfy, in order to be utilized in this operator? We show that three such properties are sufficient: the sign of the function, monotonicity, and a condition regarding the function's parameter. The correctness of an algorithm that satisfies these three properties is proved. Finally, we show an interesting application of the operator - assignment of visibility-confidence score. This feature is missing from previous approaches, where a binary yes/no visibility is determined. This score can be utilized in various applications; we illustrate its use in view-dependent curvature estimation.

Radial distortion for ordinary (non-fisheye) camera lenses has traditionally been modeled as an infinite series function of radial location of an image pixel from the image center. While there has been enough empirical evidence to show that such a model is accurate and sufficient for radial distortion calibration, there has not been much analysis on the geometric/physical understanding of radial distortion from a camera calibration perspective. In this paper, we show using a thick-lens imaging model, that the variation of entrance pupil location as a function of incident image ray angle is directly responsible for radial distortion in captured images. Thus, unlike as proposed in the current state-of-the-art in camera calibration, radial distortion and entrance pupil movement are equivalent and need not be modeled together. By modeling only entrance pupil motion instead of radial distortion, we achieve two main benefits; first, we obtain comparable if not better pixel re-projection error than traditional methods; second, and more importantly, we directly back-project a radially distorted image pixel along the true image ray which formed it. Using a thick-lens setting, we show that such a back-projection is more accurate than the two-step method of undistorting an image pixel and then back-projecting it. We have applied this calibration method to the problem of generative depth-from-focus using focal stack to get accurate depth estimates.

A variety of studies in neuroscience/neuroimaging seek to perform statistical inference on the acquired brain image scans for diagnosis as well as understanding the pathological manifestation of diseases. To do so, an important first step is to register (or co-register) all of the image data into a common coordinate system. This permits meaningful comparison of the intensities at each voxel across groups (e.g., diseased versus healthy) to evaluate the effects of the disease and/or use machine learning algorithms in a subsequent step. But errors in the underlying registration make this problematic, they either decrease the statistical power or make the follow-up inference tasks less effective/accurate. In this paper, we derive a novel algorithm which offers immunity to local errors in the underlying deformation field obtained from registration procedures. By deriving a deformation invariant representation of the image, the downstream analysis can be made more robust as if one had access to a (hypothetical) far superior registration procedure. Our algorithm is based on recent work on Scattering coefficients. Using this as a starting point, we show how results from harmonic analysis (especially, non-Euclidean wavelets) yields strategies for designing deformation and additive noise invariant representations of large 3-D brain image volumes. We present a set of results on synthetic and real brain images where we achieve robust statistical analysis even in the presence of substantial deformation errors; here, standard analysis procedures significantly under-perform and fail to identify the true signal.

We present a novel approach to relative pose estimation which is tailored to 4D light field cameras. From the relationships between scene geometry and light field structure and an analysis of the light field projection in terms of Pluecker ray coordinates, we deduce a set of linear constraints on ray space correspondences between a light field camera pair. These can be applied to infer relative pose of the light field cameras and thus obtain a point cloud reconstruction of the scene. While the proposed method has interesting relationships to pose estimation for generalized cameras based on ray-to-ray correspondence, our experiments demonstrate that our approach is both more accurate and computationally more efficient. It also compares favourably to direct linear pose estimation based on aligning the 3D point clouds obtained by reconstructing depth for each individual light field. To further validate the method, we employ the pose estimates to merge light fields captured with hand-held consumer light field cameras into refocusable panoramas.

Consumer-level and high-end light-field cameras are now widely available. Recent work has demonstrated practical methods for passive depth estimation from light-field images. However, most previous approaches do not explicitly model occlusions, and therefore cannot capture sharp transitions around object boundaries. A common assumption is that a pixel exhibits photo-consistency when focused to its correct depth, i.e., all viewpoints converge to a single (Lambertian) point in the scene. This assumption does not hold in the presence of occlusions, making most current approaches unreliable precisely where accurate depth information is most important - at depth discontinuities. In this paper, we develop a depth estimation algorithm that treats occlusion explicitly; the method also enables identification of occlusion edges, which may be useful in other applications. We show that, although pixels at occlusions do not preserve photo-consistency in general, they are still consistent in approximately half the viewpoints. Moreover, the line separating the two view regions (correct depth vs. occluder) has the same orientation as the occlusion edge has in the spatial domain. By treating these two regions separately, depth estimation can be improved. Occlusion predictions can also be computed and used for regularization. Experimental results show that our method outperforms current state-of-the-art light-field depth estimation algorithms, especially near occlusion boundaries.

The goal of this paper is to recognize actions in video without the need for examples. Different from traditional zero-shot approaches we do not demand the design and specification of attribute classifiers and class-to-attribute mappings to allow for transfer from seen classes to unseen classes. Our key contribution is objects2action, a semantic word embedding that is spanned by a skip-gram model of thousands of object categories. Action labels are assigned to an object encoding of unseen video based on a convex combination of action and object affinities. Our semantic embedding has three main characteristics to accommodate for the specifics of actions. First, we propose a mechanism to exploit multiple-word descriptions of actions and objects. Second, we incorporate the automated selection of the most responsive objects per action. And finally, we demonstrate how to extend our zero-shot approach to the spatio-temporal localization of actions in video. Experiments on four action datasets demonstrate the potential of our approach.

We propose an object detection system that relies on a multi-region deep convolutional neural network (CNN) that also encodes semantic segmentation-aware features. The resulting CNN-based representation aims at capturing a diverse set of discriminative appearance factors and exhibits localization sensitivity that is essential for accurate object localization. We exploit the above properties of our recognition module by integrating it on an iterative localization mechanism that alternates between scoring a box proposal and refining its location with a deep CNN regression model. Thanks to the efficient use of our modules, we detect objects with very high localization accuracy. On the detection challenges of PASCAL VOC2007 and PASCAL VOC2012 we achieve mAP of 78.2% and 73.9% correspondingly, surpassing any other published work by a significant margin.

In this paper, we propose a high-accuracy object detector based on co-occurrence features. Firstly, we introduce three kinds of local co-occurrence features constructed by the traditional Haar, LBP, and HOG respectively. Then the boosted detectors are learned, where each weak classifier corresponds to a local image region with a co-occurrence feature. In addition, we propose a Generalization and Efficiency Balanced (GEB) framework for boosting training. In the feature selection procedure, the discrimination ability, the generalization power, and the computation cost of the candidate features are all evaluated for decision. As a result, the boosted detector achieves both high accuracy and good efficiency. It also shows performance competitive with the state-of-the-art methods for pedestrian detection and general object detection tasks.