CVPR 2015 Papers

The objective of this paper is to design an embedding method mapping local features describing image (e.g. SIFT) to a higher dimensional representation used for image retrieval problem. By investigating the relationship between the linear approximation of a nonlinear function in high dimensional space and state-of-the-art feature representation used in image retrieval, i.e., VLAD, we first introduce a new approach for the approximation. The embedded vectors resulted by the function approximation process are then aggregated to form a single representation used in the image retrieval framework. The evaluation shows that our embedding method gives a performance boost over the state of the art in image retrieval, as demonstrated by our experiments on the standard public image retrieval benchmarks.

Visual attention is a crucial indicator of the relative importance of objects in visual scenes to human viewers. In this paper, we propose an algorithm to extract objects which attract visual attention from videos. As human attention is naturally biased towards high level semantic objects in visual scenes, this information can be valuable to extract salient objects. The proposed algorithm extracts dominant visual tracks using eye tracking data from multiple subjects on a video sequence by a combination of mean-shift clustering and Hungarian algorithm. These visual tracks guide a generic object search algorithm to get candidate object locations and extents in every frame. Further, we propose a novel multiple object extraction algorithm by constructing a spatio-temporal mixed graph over object candidates. Bounding box based object extraction inference is performed using binary linear integer programming on a cost function defined over the graph. Finally, the object boundaries are refined using grabcut segmentation. The proposed technique outperforms state-of-the-art video segmentation using eye tracking prior and obtains favorable object extraction over algorithms which do not utilize eye tracking data.

Recent advances in camera relocalization use predictions from a regression forest to guide the camera pose optimization procedure. In these methods, each tree associates one pixel with a point in the scene's 3D world coordinate frame. In previous work, these predictions were point estimates and the subsequent camera pose optimization implicitly assumed an isotropic distribution of these estimates. In this paper, we train a regression forest to predict mixtures of anisotropic 3D Gaussians and show how the predicted uncertainties can be taken into account for continuous pose optimization. Experiments show that our proposed method is able to relocalize up to 40% more frames than the state-of-the-art.

Over the last several years it has been shown that image-based object detectors are sensitive to the training data and often fail to generalize to examples that fall outside the original training sample domain (e.g., videos). A number of domain adaptation (DA) techniques have been proposed to address this problem. DA approaches are designed to adapt a fixed complexity model to the new (e.g., video) domain. We posit that unlabeled data should not only allow adaptation, but also improve (or at least maintain) performance on the original and other domains by dynamically adjusting model complexity and parameters. We call this notion domain expansion. To this end, we develop a new scalable and accurate incremental object detection algorithm, based on several extensions of large-margin embedding (LME). Our detection model consists of an embedding space and multiple class prototypes in that embedding space, that represent object classes; distance to those prototypes allows us to reason about multi-class detection. By incrementally detecting object instances in video and adding confident detections into the model, we are able to dynamically adjust the complexity of the detector over time by instantiating new prototypes to span all domains the model has seen. We test performance of our approach by expanding an object detector trained on ImageNet to detect objects in ego-centric videos of Activity Daily Living (ADL) dataset and challenging YouTube Objects (YTO) dataset.

In this work, we investigate the use of exemplar SVMs (linear SVMs trained with one positive example only and a vast collection of negative examples) as encoders that turn generic image features into new, task-tailored features. The proposed feature encoding leverages the ability of the exemplar-SVM (E-SVM) classifier to extract, from the original representation of the exemplar image, what is unique about it. While existing image description pipelines rely on the intuition of the designer to encode uniqueness into the feature encoding process, our proposed approach does it explicitly relative to a "universe" of features represented by the generic negatives. We show that such a post-processing enhances the performance of state-of-the art image retrieval methods based on aggregated image features, as well as the performance of nearest class mean and K-nearest neighbor image classification methods. We establish these advantages for several features, including "traditional" features as well as features derived from deep convolutional neural nets. As an additional contribution, we also propose a recursive extension of this E-SVM encoding scheme (RE-SVM) that provides further performance gains.

We present an approach for correcting the bias in 3D reconstruction of points imaged by a calibrated stereo rig. Our analysis is based on the observation that, due to quantization error, a 3D point reconstructed by triangulation essentially represents an entire region in space. The true location of the world point that generated the triangulated point could be anywhere in this region. We argue that the reconstructed point, if it is to represent this region in space without bias, should be located at the centroid of this region, which is not what has been done in the literature. We derive the exact geometry of these regions in space, which we call 3D cells, and we show how they can be viewed as uniform distributions of possible pre-images of the pair of corresponding pixels. By assuming a uniform distribution of points in 3D, as opposed to a uniform distribution of the projections of these 3D points on the images, we arrive at a fast and exact computation of the triangulation bias in each cell. In addition, we derive the exact covariance matrices of the 3D cells. We validate our approach in a variety of simulations ranging from 3D reconstruction to camera localization and relative motion estimation. In all cases, we are able to demonstrate a marked improvement compared to conventional techniques for small disparity values, for which bias is significant and the required corrections are large.

This paper presents a stereo matching approach for a novel multi-perspective panoramic stereo vision system, making use of asynchronous and non-simultaneous stereo imaging towards real-time 3D 360deg vision. The method is designed for events representing the scenes visual contrast as a sparse visual code allowing the stereo reconstruction of high resolution panoramic views. We propose a novel cost measure for the stereo matching, which makes use of a similarity measure based on event distributions. Thus, the robustness to variations in event occurrences was increased. An evaluation of the proposed stereo method is presented using distance estimation of panoramic stereo views and ground truth data. Furthermore, our approach is compared to standard stereo methods applied on event-data. Results show that we obtain 3D reconstructions of 1024 x 3600 round views and outperform depth reconstruction accuracy of state-of-the-art methods on event data.

Image classification has advanced significantly in recent years with the availability of large-scale image sets. However, fine-grained classification remains a major challenge due to the annotation cost of large numbers of fine-grained categories. This project shows that compelling classification performance can be achieved on such categories even without labeled training data. Given image and class embeddings, we learn a compatibility function such that matching embeddings are assigned a higher score than mismatching ones; zero-shot classification of an image proceeds by finding the label yielding the highest joint compatibility score. We use state-of-the-art image features and focus on different supervised attributes and unsupervised output embeddings either derived from hierarchies or learned from unlabeled text corpora. We establish a substantially improved state-of-the-art on the Animals with Attributes and Caltech-UCSD Birds datasets. Most encouragingly, we demonstrate that purely unsupervised output embeddings (learned from Wikipedia and improved with fine-grained text) achieve compelling results, even outperforming the previous supervised state-of-the-art. By combining different output embeddings, we further improve results.

We propose a novel approach for optical flow estimation, targeted at large displacements with significant occlusions. It consists of two steps: i) dense matching by edge-preserving interpolation from a sparse set of matches; ii) variational energy minimization initialized with the dense matches. The sparse-to-dense interpolation relies on an appropriate choice of the distance, namely an edge-aware geodesic distance. This distance is tailored to handle occlusions and motion boundaries - two common and difficult issues for optical flow computation. We also propose an approximation scheme for the geodesic distance to allow fast computation without loss of performance. Subsequent to the dense interpolation step, standard one-level variational energy minimization is carried out on the dense matches to obtain the final flow estimation. The proposed approach, called Edge-Preserving Interpolation of Correspondences (EpicFlow) is fast and robust to large displacements. It significantly outperforms the state of the art on MPI-Sintel and performs on par on Kitti and Middlebury.

A large body of recent work on object detection has focused on exploiting 3D CAD model databases to improve detection performance. Many of these approaches work by aligning exact 3D models to images using templates generated from renderings of the 3D models at a set of discrete viewpoints. However, the training procedures for these approaches are computationally expensive and require gigabytes of memory and storage, while the viewpoint discretization hampers pose estimation performance. We propose an efficient method for synthesizing templates from 3D models that runs on the fly -- that is, it quickly produces detectors for an arbitrary viewpoint of a 3D model without expensive dataset-dependent training or template storage. Given a 3D model and an arbitrary continuous detection viewpoint, our method synthesizes a discriminative template by extracting features from a rendered view of the object and decorrelating spatial dependences among the features. Our decorrelation procedure relies on a gradient-based algorithm that is more numerically stable than standard decomposition-based procedures, and we efficiently search for candidate detections by computing FFT-based template convolutions. Due to the speed of our template synthesis procedure, we are able to perform joint optimization of scale, translation, continuous rotation, and focal length using Metropolis-Hastings algorithm. We provide an efficient GPU implementation of our algorithm, and we validate its performance on 3D Object Classes and PASCAL3D+ datasets.

Deformable Parts Models and Convolutional Networks each have achieved notable performance in object detection. Yet these two approaches find their strengths in complementary areas: DPMs are well-versed in object composition, modeling fine-grained spatial relationships between parts; likewise, ConvNets are adept at producing powerful image features, having been discriminatively trained directly on the pixels. In this paper, we propose a new model that combines these two approaches, obtaining the advantages of each. We train this model using a new structured loss function that considers all bounding boxes within an image, rather than isolated object instances. This enables the non-maximal suppression (NMS) operation, previously treated as a separate post-processing stage, to be integrated into the model. This allows for discriminative training of our combined Convnet + DPM + NMS model in end-to-end fashion. We evaluate our system on PASCAL VOC 2007 and 2011 datasets, achieving competitive results on both bench- marks.

We present a novel video saliency detection method to support human activity recognition and weakly supervised training of activity detection algorithms. Recent research has emphasized the need for analyzing salient information in videos to minimize dataset bias or to supervise weakly labeled training of activity detectors. In contrast to previous methods we do not rely on training information given by either eye-gaze or annotation data, but propose a fully unsupervised algorithm to find salient regions within videos. In general, we enforce the Gestalt principle of figure-ground segregation for both appearance and motion cues. We introduce an encoding approach that allows for efficient computation of saliency by approximating joint feature distributions. We evaluate our approach on several datasets, including challenging scenarios with cluttered background and camera motion, as well as salient object detection in images. Overall, we demonstrate favorable performance compared to state-of-the-art methods in estimating both ground-truth eye-gaze and activity annotations.

We introduce Embedded PS, a new robust and accurate phase shifting algorithm for 3D scanning. The method projects only high frequency sinusoidal patterns in order to reduce errors due to global illumination effects, such as subsurface scattering and interreflections. The frequency set for the projected patterns is specially designed so that our algorithm can extract a set of embedded low frequency sinusoidals with simple math. All the signals, patterns high and embedded low frequencies, are used with temporal phase unwrapping to compute absolute phase values in closed-form, without quantization or approximation via LUT, resulting in fast computation. The absolute phases provide correspondences from projector to camera pixels which enable to recover 3D points using optical triangulation. The algorithm estimates multiple absolute phase values per pixel which are combined to reduce measurement noise while preserving fine details. We prove that embedded periodic signals can be recovered from any periodic signal, not just sinusoidal signals, which may result in further improvements for other 3D imaging methods. Several experiments are presented showing that our algorithm produces more robust and accurate 3D scanning results than state-of-the-art methods for challenging surface materials, with an equal or smaller number of projected patterns and at lower computational cost.

Learning a low-dimensional structure plays an important role in computer vision. Recently, a new family of methods, such as l1 minimization and robust principal component analysis, has been proposed for low-rank matrix approximation problems and shown to be robust against outliers and missing data. But these methods often require heavy computational load and can fail to find a solution when highly corrupted data are presented. In this paper, an elastic-net regularization based low-rank matrix factorization method for subspace learning is proposed. The proposed method finds a robust solution efficiently by enforcing a strong convex constraint to improve the algorithm's stability while maintaining the low-rank property of the solution. It is shown that any stationary point of the proposed algorithm satisfies the Karush-Kuhn-Tucker optimality conditions. The proposed method is applied to a number of low-rank matrix approximation problems to demonstrate its efficiency in the presence of heavy corruptions and to show its effectiveness and robustness compared to the existing methods.

Human activities observed from visual sensors often give rise to a sequence of smoothly varying features. In many cases, the space of features can be formally defined as a manifold, where the action becomes a trajectory on the manifold. Such trajectories are high dimensional in addition to being non-linear, which can severely limit computations on them. We also argue that by their nature, human actions themselves lie on a much lower dimensional manifold compared to the high dimensional feature space. Learning an accurate low dimensional embedding for actions could have a huge impact in the areas of efficient search and retrieval, visualization, learning, and recognition. Traditional manifold learning addresses this problem for static points in R^n, but its extension to trajectories on Riemannian manifolds is non-trivial and has remained unexplored. The challenge arises due to the inherent non-linearity, and temporal variability that can significantly distort the distance metric between trajectories. To address these issues we use the transport square-root velocity function (TSRVF) space, a recently proposed representation that provides a metric which has favorable theoretical properties such as invariance to group action. We propose to learn the low dimensional embedding with a manifold functional variant of principal component analysis (mfPCA). We show that mfPCA effectively models the manifold trajectories in several applications such as action recognition, clustering and diverse sequence sampling while reducing the dimensionality by a factor of ~250x. The mfPCA features can also be reconstructed back to the original manifold to allow for easy visualization of the latent variable space.

While egocentric cameras like GoPro are gaining popularity, the videos they capture are long, boring, and difficult to watch from start to end. Fast forwarding (i.e. frame sampling) is a natural choice for faster video browsing. However, this accentuates the shake caused by natural head motion, making the fast forwarded video useless. We propose EgoSampling, an adaptive frame sampling that gives more stable fast forwarded videos. Adaptive frame sampling is formulated as energy minimization, whose optimal solution can be found in polynomial time. In addition, egocentric video taken while walking suffers from the left-right movement of the head as the body weight shifts from one leg to another. We turn this drawback into a feature: Stereo video can be created by sampling the frames from the left most and right most head positions of each step, forming approximate stereo-pairs.

We envision a future time when wearable cameras (e.g., small cameras in glasses or pinned on a shirt collar) are worn by the masses and record first-person point-of-view (POV) videos of everyday life. While these cameras can enable new assistive technologies and novel research challenges, they also raise serious privacy concerns. For example, first-person videos passively recorded by wearable cameras will necessarily include anyone who comes into the view of a camera -- with or without consent. Motivated by these benefits and risks, we develop a self-search technique tailored to first-person POV videos. The key observation of our work is that the egocentric head motions of a target person (i.e., the self) are observed both in the POV video of the target and observer. The motion correlation between the target person's video and the observer's video can then be used to uniquely identify instances of the self. We incorporate this feature into our proposed approach that computes the motion correlation over supervoxel hierarchies to localize target instances in observer videos. Our proposed approach significantly improves self-search performance over several well-known face detectors and recognizers. Furthermore, we show how our approach can enable several practical applications such as privacy filtering, automated video collection and social group discovery.

This paper aims to accelerate the test-time computation of deep convolutional neural networks (CNNs). Unlike existing methods that are designed for approximating linear filters or linear responses, our method takes the nonlinear units into account. We minimize the reconstruction error of the nonlinear responses, subject to a low-rank constraint which helps to reduce the complexity of filters. We develop an effective solution to this constrained nonlinear optimization problem. An algorithm is also presented for reducing the accumulated error when multiple layers are approximated. A whole-model speedup ratio of 4x is demonstrated on a large network trained for ImageNet, while the top-5 error rate is only increased by 0.9%. Our accelerated model has a comparably fast speed as the "AlexNet", but is 4.7% more accurate.

We address the elusive goal of estimating optical flow both accurately and efficiently by adopting a sparse-to-dense approach. Given a set of sparse matches, we regress to dense optical flow using a learned set of full-frame basis flow fields. We learn the principal components of natural flow fields using flow computed from four Hollywood movies. Optical flow fields are then compactly approximated as a weighted sum of the basis flow fields. Our new PCA-Flow algorithm robustly estimates these weights from sparse feature matches. The method runs in under 200ms/frame on the MPI-Sintel dataset using a single CPU and is more accurate and significantly faster than popular methods such as LDOF and Classic+NL. For some applications, however, the results are too smooth. Consequently, we develop a novel sparse layered flow method in which each layer is represented by PCA-Flow. Unlike existing layered methods, estimation is fast because it uses only sparse matches. We combine information from different layers into a dense flow field using an image-aware MRF. The resulting PCA-Layers method runs in 3.2s/frame, is significantly more accurate than PCA-Flow, and achieves state-of-the-art performance in occluded regions on MPI-Sintel.