CVPR 2015 Papers

Computer-aided diagnosis of medical images requires thorough analysis of image details. For example, examining all cells enables fine-grained categorization of histopathological images. Traditional computational methods may have efficiency issues when performing such detailed analysis. In this paper, we propose a robust and scalable solution to achieve this. Specifically, a robust segmentation method is developed to delineate region-of-interests (e.g., cells) accurately, using hierarchical voting and repulsive active contour. A hashing-based large-scale retrieval approach is also designed to examine and classify them by comparing with a massive training database. We evaluate this proposed framework on a challenging and important clinical use case, i.e., differentiation of two types of lung cancers (the adenocarcinoma and the squamous carcinoma), using thousands of histopathological images extracted from hundreds of patients. Our method has achieved promising performance, i.e., 87.3% accuracy and 1.68 seconds by searching among half-million cells.

A video dataset that is designed to study fine-grained categorisation of pedestrians is introduced. Pedestrians were recorded ``in-the-wild'' from a moving vehicle. Annotations include bounding boxes, tracks, 14 keypoints with occlusion information and the fine-grained categories of age (5 classes), sex (2 classes), weight (3 classes) and clothing style (4 classes). There are a total of 27,454 bounding box and pose labels across 4222 tracks. This dataset is designed to train and test algorithms for fine-grained categorisation of people; it is also useful for benchmarking tracking, detection and pose estimation of pedestrians. State-of-the-art algorithms for fine-grained classification and pose estimation were tested using the dataset and the results are reported as a useful performance baseline.

We propose a new computer vision task we call "distractor prediction." Distractors are the regions of an image that draw attention away from the main subjects and reduce the overall image quality. Removing distractors --- for example, using in-painting --- can improve the composition of an image. In this work we created two datasets of images with user annotations to identify the characteristics of distractors. We use these datasets to train an algorithm to predict distractor maps. Finally, we use our predictor to automatically enhance images.

We address the problem of action detection in videos. Driven by the latest progress in object detection from 2D images, we build action models using rich feature hierarchies derived from shape and kinematic cues. We incorporate appearance and motion in two ways. First, starting from image region proposals we select those that are motion salient and thus are more likely to contain the action. This leads to a significant reduction in the number of regions being processed and allows for faster computations. Second, we extract spatio-temporal feature representations to build strong classifiers using Convolutional Neural Networks. We link our predictions to produce detections consistent in time, which we call action tubes. We show that our approach outperforms other techniques in the task of action detection.

This paper starts from the observation that multiple top performing pedestrian detectors can be modelled by using an intermediate layer filtering low-level features in combination with a boosted decision forest. Based on this observation we propose a unifying framework and experimentally explore different filter families. We report extensive results enabling a systematic analysis. Using filtered channel features we obtain top performance on the challenging Caltech and KITTI datasets, while using only HOG+LUV as low-level features. When adding optical flow features we further improve detection quality and report the best known result on the Caltech dataset, reaching 93% recall at 1 FPPI.

We introduce a purely feed-forward architecture for semantic segmentation. We map small image elements (superpixels) to rich feature representations extracted from a sequence of nested regions of increasing extent. These regions are obtained by "zooming out" from the superpixel all the way to scene-level resolution. This approach exploits statistical structure in the image and in the label space without setting up explicit structured prediction mechanisms, and thus avoids complex and expensive inference. Instead superpixels are classified by a feedforward multilayer network. Our architecture achieves 69.6% average accuracy on the PASCAL VOC 2012 test set, and 86.1 pixel accuracy on Stanford Background Dataset.

Automatic image annotation is a highly valuable tool for image search, retrieval and archival systems. In the absence of an annotation tool, such systems have to rely on either users' input or large amount of text on the webpage of the image, to acquire its textual description. Users may provide insufficient/noisy tags and all the text on the webpage may not be a description or an explanation of the accompanying image. Therefore, it is of extreme importance to develop efficient tools for automatic annotation of images with correct and sufficient tags. The context of the image plays a significant role in this process, along with the content of the image. A suitable quantification of the context of the image may reduce the semantic gap between visual features and appropriate textual description of the image. In this paper, we present an unsupervised feature-independent quantification of the context of the image through tensor decomposition. We incorporate the estimated context as prior knowledge in the process of automatic image annotation. Evaluation of the predicted annotations provides evidence of the effectiveness of our feature-independent context estimation method.

Markerless tracking of hands and fingers is a promising enabler for human-computer interaction. However, adoption has been limited because of tracking inaccuracies, incomplete coverage of motions, low framerate, complex camera setups, and high computational requirements. In this paper, we present a fast method for accurately tracking rapid and complex articulations of the hand using a single depth camera. Our algorithm uses a novel detection-guided optimization strategy that increases the robustness and speed of pose estimation. In the detection step, a randomized decision forest classifies pixels into parts of the hand. In the optimization step, a novel objective function combines the detected part labels and a Gaussian mixture representation of the depth to estimate a pose that best fits the depth. Our approach needs comparably less computational resources which makes it extremely fast (50 fps without GPU support). The approach also supports varying static, or moving, camera-to-scene arrangements. We show the benefits of our method by evaluating on public datasets and comparing against previous work.

Convolutional sparse coding (CSC) has become an increasingly important tool in machine learning and computer vision. Image features can be learned and subsequently used for classification and reconstruction tasks. As opposed to patch-based methods, convolutional sparse coding operates on whole images, thereby seamlessly capturing the correlation between local neighborhoods. In this paper, we propose a new approach to solving CSC problems and show that our method converges significantly faster and also finds better solutions than the state of the art. In addition, the proposed method is the first efficient approach to allow for proper boundary conditions to be imposed and it also supports feature learning from incomplete data as well as general reconstruction problems.

The aim of single image super-resolution is to reconstruct a high-resolution image from a single low-resolution input. Although the task is ill-posed it can be seen as finding a non-linear mapping from a low to high-dimensional space. Recent methods that rely on both neighborhood embedding and sparse-coding have led to tremendous quality improvements. Yet, many of the previous approaches are hard to apply in practice because they are either too slow or demand tedious parameter tweaks. In this paper, we propose to directly map from low to high-resolution patches using random forests. We show the close relation of previous work on single image super-resolution to locally linear regression and demonstrate how random forests nicely fit into this framework. During training the trees, we optimize a novel and effective regularized objective that not only operates on the output space but also on the input space, which especially suits the regression task. During inference, our method comprises the same well-known computational efficiency that has made random forests popular for many computer vision problems. In the experimental part, we demonstrate on standard benchmarks for single image super-resolution that our approach yields highly accurate state-of-the-art results, while being fast in both training and evaluation.

Rank minimization problem can be boiled down to either Nuclear Norm Minimization (NNM) or Weighted NNM (WNNM) problem. The problems related to NNM (or WNNM) can be solved iteratively by applying a closed-form proximal operator, called Singular Value Thresholding (SVT) (or Weighted SVT), but they suffer from high computational cost to compute a Singular Value Decomposition (SVD) at each iteration. In this paper, we propose an accurate and fast approximation method for SVT, called fast randomized SVT (FRSVT), where we avoid direct computation of SVD. The key idea is to extract an approximate basis for the range of a matrix from its compressed matrix. Given the basis, we compute the partial singular values of the original matrix from a small factored matrix. While the basis approximation is the bottleneck, our method is already severalfold faster than thin SVD. By adopting a range propagation technique, we can further avoid one of the bottleneck at each iteration. Our theoretical analysis provides a stepping stone between the approximation bound of SVD and its effect to NNM via SVT. Along with the analysis, our empirical results on both quantitative and qualitative studies show our approximation rarely harms the convergence behavior of the host algorithms. We apply it and validate the efficiency of our method on various vision problems, e.g. subspace clustering, weather artifact removal, simultaneous multi-image alignment and rectification.

Given a stereo pair it is possible to recover a depth map and use that depth to render a synthetically defocused image. Though stereo algorithms are well-studied, rarely are those algorithms considered solely in the context of producing these defocused renderings. In this paper we present a technique for efficiently producing disparity maps using a novel optimization framework in which inference is performed in "bilateral-space". Our approach produces higher-quality "defocus" results than other stereo algorithms while also being 10-100 times faster than comparable techniques.

In this paper we target at generating generic action proposals in unconstrained videos. Each action proposal corresponds to a temporal series of spatial bounding boxes, i.e., a spatio-temporal video tube, which has a good potential to locate one human action. Assuming each action is performed by a human with meaningful motion, both appearance and motion cues are utilized to measure the actionness of the video tubes. After picking those spatiotemporal paths of high actionness scores, our action proposal generation is formulated as a maximum set coverage problem, where greedy search is performed to select a set of action proposals that can maximize the overall actionness score. Compared with existing action proposal approaches, our action proposals do not rely on video segmentation and can be generated in nearly real-time. Experimental results on two challenging datasets, MSRII and UCF 101, validate the superior performance of our action proposals as well as competitive results on action detection and search.

A method for efficient border ownership assignment in 2D images is proposed. Leveraging on recent advances using Structured Random Forests (SRF) for boundary detection, we impose a novel border ownership structure that detects both boundaries and border ownership at the same time. Key to this work are features that predict ownership cues from 2D images. To this end, we use several different local cues: shape, spectral properties of boundary patches, and semi-global grouping cues that are indicative of perceived depth. For shape, we use HoG-like descriptors that encode local curvature (convexity and concavity). For spectral properties, such as extremal edges, we first learn an orthonormal basis spanned by the top K eigenvectors via PCA over common types of contour tokens. For grouping, we introduce a novel mid-level descriptor that captures patterns near edges and indicates ownership information of the boundary. Experimental results over a subset of the Berkeley Segmentation Dataset (BSDS) and the NYU Depth V2 dataset show that our method's performance exceeds current state-of-the-art multi-stage approaches that use more complex features.

Despite significant recent advances in the field of face recognition [DeepFace, DeepId2], implementing face verification and recognition efficiently at scale presents serious challenges to current approaches. In this paper we present a system, called FaceNet, that directly learns a mapping from face images to a compact Euclidean space where distances directly correspond to a measure of face similarity. Once this space has been produced, tasks such as face recognition, verification and clustering can be easily implemented using standard techniques with FaceNet embeddings as feature vectors. Our method uses a deep convolutional network trained to directly optimize the embedding itself, rather than an intermediate bottleneck layer as in previous approaches. To train, we use triplets of roughly aligned matching / non-matching face patches generated using a novel online triplet mining method. The benefit of our approach is much greater representational efficiency: we achieve state-of-the-art face recognition performance using only 128 bytes per face. On the widely used Labeled Faces in the Wild (LFW) dataset, our system achieves a new record accuracy of 99.63%. On YouTube Faces DB it achieves 95.12%. Our system cuts the error rate in comparison to the best published result [DeepId2+] by 30% on both datasets.

Retrieving videos of a specific person given his/her face image as query becomes more and more appealing for applications like smart movie fast-forwards and suspect searching. It also forms an interesting but challenging computer vision task, as the visual data to match, i.e., still image and video clip are usually represented quite differently. Typically, face image is represented as point (i.e., vector) in Euclidean space, while video clip is seemingly modeled as a point (e.g., covariance matrix) on some particular Riemannian manifold in the light of its recent promising success. It thus incurs a new hashing-based retrieval problem of matching two heterogeneous representations, respectively in Euclidean space and Riemannian manifold. This work makes the first attempt to embed the two heterogeneous spaces into a common discriminant Hamming space. Specifically, we propose Hashing across Euclidean space and Riemannian manifold (HER) by deriving a unified framework to firstly embed the two spaces into corresponding reproducing kernel Hilbert spaces, and then iteratively optimize the intra- and inter-space Hamming distances in a maxmargin framework to learn the hash functions for the two spaces. Extensive experiments demonstrate the impressive superiority of our method over the state-of-the-art competitive hash learning methods.

We present a novel face alignment framework based on coarse-to-fine shape searching. Unlike the conventional cascaded regression approaches that start with an initial shape and refine the shape in a cascaded manner, our approach begins with a coarse search over a shape space that contains diverse shapes, and employs the coarse solution to constrain subsequent finer search of shapes. The unique stage-by-stage progressive and adaptive search i) prevents the final solution from being trapped in local optima due to poor initialisation, a common problem encountered by cascaded regression approaches; and ii) improves the robustness in coping with large pose variations. The framework demonstrates real-time performance and state-of-theart results on various benchmarks including the challenging 300-W dataset.

In this paper, we propose a face alignment method that uses cascade Gaussian process regression trees (cGPRT) constructed by combining Gaussian process regression trees (GPRT) in a cascade stage-wise manner. Here, GPRT is a Gaussian process with a kernel defined by a set of trees. The kernel measures the similarity between two inputs as the number of trees where the two inputs fall in the same leaves. Without increasing prediction time, the prediction of cGPRT can be performed in the same framework as the cascade regression trees (CRT) but with better generalization. Features for GPRT are designed using shape-indexed difference of Gaussian (DoG) filter responses sampled from local retinal patterns to increase stability and to attain robustness against geometric variances. Compared with the previous CRT-based face alignment methods that have shown state-of-the-art performances, cGPRT using shape-indexed DoG features performed best on the HELEN and 300-W datasets which are the most challenging dataset today.

Low rank representation (LRR) has shown promising performance for various computer vision applications such as face clustering. Existing algorithms for solving LRR usually depend on its two-variable formulation which contains the original data matrix. In this paper, we develop a fast LRR solver called FaLRR, by reformulating LRR as a new optimization problem with regard to factorized data (which is obtained by skinny SVD of the original data matrix). The new formulation benefits the corresponding optimization and theoretical analysis. Specifically, to solve the resultant optimization problem, we propose a new algorithm which is not only efficient but also theoretically guaranteed to obtain a globally optimal solution. Regarding the theoretical analysis, the new formulation is helpful for deriving some interesting properties of LRR. Last but not least, the proposed algorithm can be readily incorporated into an existing distributed framework of LRR for further acceleration. Extensive experiments on synthetic and real-world datasets demonstrate that our FaLRR achieves order-of-magnitude speedup over existing LRR solvers, and the efficiency can be further improved by incorporating our algorithm into the distributed framework of LRR.