CVPR 2016 Papers

In this work we propose a technique that transfers supervision between images from different modalities. We use learned representations from a large labeled modality as supervisory signal for training representations for a new unlabeled paired modality. Our method enables learning of rich representations for unlabeled modalities and can be used as a pre-training procedure for new modalities with limited labeled data. We transfer supervision from labeled RGB images to unlabeled depth and optical flow images and demonstrate large improvements for both these cross modal supervision transfers.

There is a great deal of interest in using large scale brain imaging studies to understand how brain connectivity evolves over time for an individual and how it varies over different levels/quantiles of cognitive function. To do so, one typically performs so-called tractography procedures on diffusion MR brain images and derives measures of brain connectivity expressed as graphs. The nodes correspond to distinct brain regions and the edges encode the strength of the connection. The scientific interest is in characterizing the evolution of these graphs over time or from healthy individuals to diseased. We pose this important question in terms of the Laplacian of the connectivity graphs derived from various longitudinal or disease time points - quantifying its progression is then expressed in terms of coupling the harmonic bases of a full set of Laplacians. We derive a coupled system of generalized eigenvalue problems (and corresponding numerical optimization schemes) whose solution helps characterize the full life cycle of brain connectivity evolution in a given dataset. Finally, we show a set of results on a diffusion MR imaging dataset of middle aged people at risk for Alzheimer's disease (AD), who are cognitively healthy. In such asymptomatic adults, we find that a framework for characterizing brain connectivity evolution provides the ability to predict cognitive scores for individual subjects, and for estimating the progression of participant's brain connectivity into the future.

Joint modeling of the intensity of facial action units (AUs) from face images is challenging due to the large number of AUs (30+) and their intensity levels (6). This is in part due to the lack of suitable models that can efficiently handle such a large number of outputs/classes simultaneously, but also due to the lack of target data. For this reason, majority of the methods proposed resort to independent classifiers for the AU intensity. This is suboptimal for at least two reasons: the facial appearance of some AUs changes depending on the intensity of other AUs, and some AUs co-occur more often than others. Encoding this is expected to improve the estimation of target AU intensities, especially in the case of noisy image features, head-pose variations and imbalanced training data. To this end, we introduce a novel modeling framework, Copula Ordinal Regression (COR), that leverages the power of copula functions and CRFs, to detangle the probabilistic modeling of AU dependencies from the marginal modeling of the AU intensity. Consequently, the COR model achieves the joint learning and inference of intensities of multiple AUs, while being computationally tractable. We show on two challenging datasets of naturalistic facial expressions that the proposed approach consistently outperforms (i) independent modeling of AU intensities, and (ii) the state-of-the-art approach for the target task.

While great progress has been made in stereo computation over the last decades, large textureless regions remain challenging. Segment-based methods can tackle this problem properly, but their performances are sensitive to the segmentation results. In this paper, we alleviate the sensitivity by generating multiple proposals on absolute and relative disparities from multi-segmentations. These proposals supply rich descriptions of surface structures. Especially, the relative disparity between distant pixels can encode the large structure, which is critical to handle the large texture-less regions. The proposals are coordinated by point-wise competition and pairwise collaboration within a MRF model. During inference, a dynamic programming is performed in different directions with various step sizes, so the long-range connections are better preserved. In the experiments, we carefully analyzed the effectiveness of the major components. Results on the 2014 Middlebury and KITTI 2015 stereo benchmark show that our method is comparable to state-of-the-art.

Recent applications of Convolutional Neural Networks (ConvNets) for human action recognition in videos have proposed different solutions for incorporating the appearance and motion information. We study a number of ways of fusing ConvNet towers both spatially and temporally in order to best take advantage of this spatio-temporal information. We make the following findings: (i) that rather than fusing at the softmax layer, a spatial and temporal network can be fused at a convolution layer without loss of performance, but with a substantial saving in parameters; (ii) that it is better to fuse such networks spatially at the last convolutional layer than earlier, and that additionally fusing at the class prediction layer can boost accuracy; finally (iii) that pooling of abstract convolutional features over spatiotemporal neighbourhoods further boosts performance. Based on these studies we propose a new ConvNet architecture for spatiotemporal fusion of video snippets, and evaluate its performance on standard benchmarks where this architecture achieves state-of-the-art results.

Convolutional Neural Networks (CNNs) have recently been successfully applied to various Computer Vision (CV) applications. In this paper we utilize CNNs to predict depth information for given Light Field (LF) data. The proposed method learns an end-to-end mapping between the 4D light field and a representation of the corresponding 4D depth field in terms of 2D hyperplane orientations. The obtained prediction is then further refined in a post processing step by applying a higher-order regularization. Existing LF datasets are not sufficient for the purpose of the training scheme tackled in this paper. This is mainly due to the fact that the ground truth depth of existing datasets is inaccurate and/or the datasets are limited to a small number of LFs. This made it necessary to generate a new synthetic LF dataset, which is based on the raytracing software POV-Ray. This new dataset provides floating point accurate ground truth depth fields, and due to a random scene generator the dataset can be scaled as required.

Segmenting an image into multiple components is a central task in computer vision. In many practical scenarios, prior knowledge about plausible components is available. Incorporating such prior knowledge into models and algorithms for image segmentation is highly desirable, yet can be non-trivial. In this work, we introduce a new approach that allows, for the first time, to constrain some or all components of a segmentation to have convex shapes. Specifically, we extend the Minimum Cost Multicut Problem by a class of constraints that enforce convexity. To solve instances of this NP-hard integer linear program to optimality, we separate the proposed constraints in the branch-and-cut loop of a state-of-the-art ILP solver. Results on photographs and micrographs demonstrate the effectiveness of the approach as well as its advantages over the state-of-the-art heuristic.

Point set registration (PSR) is a fundamental problem in computer vision and pattern recognition, and it has been successfully applied to many applications. Although widely used, existing PSR methods cannot align point sets robustly under degradations, such as deformation, noise, occlusion, outlier, rotation, and multi-view changes. This paper proposes context-aware Gaussian fields (CA-LapGF) for non-rigid PSR subject to global rigid and local non-rigid geometric constraints, where a laplacian regularized term is added to preserve the intrinsic geometry of the transformed set. CA-LapGF uses a robust objective function and the quasi-Newton algorithm to estimate the likely correspondences, and the non-rigid transformation parameters between two point sets iteratively. The CA-LapGF can estimate non-rigid transformations, which are mapped to reproducing kernel Hilbert spaces, accurately and robustly in the presence of degradations. Experimental results on synthetic and real images reveal that how CA-LapGF outperforms state-of-the-art algorithms for non-rigid PSR.

We present an unsupervised visual feature learning algorithm driven by context-based pixel prediction. By analogy with auto-encoders, we propose Context Encoders -- a convolutional neural network trained to generate the contents of an arbitrary image region conditioned on its surroundings. In order to succeed at this task, context encoders need to both understand the content of the entire image, as well as produce a plausible hypothesis for the missing part(s). When training context encoders, we have experimented with both a standard pixel-wise reconstruction loss, as well as a reconstruction plus an adversarial loss. The latter produces much sharper results because it can better handle multiple modes in the output. We found that a context encoder learns a representation that captures not just appearance but also the semantics of visual structures. We quantitatively demonstrate the effectiveness of our learned features for CNN pre-training on classification, detection, and segmentation tasks. Furthermore, context encoders can be used for semantic inpainting tasks, either stand-alone or as initialization for non-parametric methods.

In view selection, little work has been done for optimizing the search process; views must be densely distributed and checked individually. Thus, evaluating poor views wastes much time, and a poor view may even be misidentified as a best one. In this paper, we propose a search strategy by identifying the regions that are very likely to contain best views, referred to as canonical regions. It is by decomposing the model under investigation into meaningful parts, and using the canonical views of these parts to generate canonical regions. Applying existing view selection methods in the canonical regions can not only accelerate the search process but also guarantee the quality of obtained views. As a result, when our canonical regions are used for searching N-best views during comprehensive model analysis, we can attain greater search speed and reduce the number of views required. Experimental results show the effectiveness of our method.

Cascade regression framework has been shown to be effective for facial landmark detection. It starts from an initial face shape and gradually predicts the face shape update from the local appearance features to generate the facial landmark locations in the next iteration until convergence. In this paper, we improve upon the cascade regression framework and propose the Constrained Joint Cascade Regression Framework (CJCRF) for simultaneous facial action unit recognition and facial landmark detection, which are two related face analysis tasks, but are seldomly exploited together. In particular, we first learn the relationships among facial action units and face shapes as a constraint. Then, in the proposed constrained joint cascade regression framework, with the help from the constraint, we iteratively update the facial landmark locations and the action unit activation probabilities until convergence. Experimental results demonstrate that the intertwined relationships of facial action units and face shapes boost the performances of both facial action unit recognition and facial landmark detection. The experimental results also demonstrate the effectiveness of the proposed method comparing to the state-of-the-art works.

Feature learning with deep models has achieved impressive results for both data representation and classification for various vision tasks. Deep feature learning, however,typically requires a large amount of training data, which may not be feasible for some application domains. Transfer learning can be one of the approaches to alleviate this problem by transferring data from data-rich source domain to data-scarce target domain. Existing transfer learning methods typically perform one-shot transfer learning and often ignore the specific properties that the transferred data must satisfy. To address these issues, we introduce a constrained deep transfer feature learning method to perform simultaneous transfer learning and feature learning by performing transfer learning in a progressively improving feature space iteratively in order to better narrow the gap between the target domain and the source domain for effective transfer of the data from source domain to target domain. Furthermore, we propose to exploit the target domain knowledge and incorporate such prior knowledge as constraint during transfer learning to ensure that the transferred data satisfies certain properties of the target domain. To demonstrate the effectiveness of the proposed constrained deep transfer feature learning method, we apply it to thermal feature learning for eye detection by transferring from the visible domain. We also applied the proposed method for cross-view facial expression recognition as a second application. The experimental results demonstrate the effectiveness of the proposed method for both applications.

Silhouettes provide rich information on three-dimensional shape, since the intersection of the associated visual cones generates the "visual hull", which encloses and approximates the original shape. However, not all silhouettes can actually be projections of the same object in space: this simple observation has implications in object recognition and multi-view segmentation, and has been (often implicitly) used as a basis for camera calibration. In this paper, we investigate the conditions for multiple silhouettes, or more generally arbitrary closed image sets, to be geometrically "consistent". We present this notion as a natural generalization of traditional multi-view geometry, which deals with consistency for points. After discussing some general results, we present a "dual" formulation for consistency, that gives conditions for a family of planar sets to be sections of the same object. Finally, we introduce a more general notion of silhouette "compatibility" under partial knowledge of the camera projections, and point out some possible directions for future research.

Deformations of surfaces with the same intrinsic shape can often be described accurately by a conformal model. A major focus of computational conformal geometry is the estimation of the conformal mapping that aligns a given pair of object surfaces. The uniformization theorem en- ables this task to be acccomplished in a canonical 2D do- main, wherein the surfaces can be aligned using a M obius transformation. Current algorithms for estimating M obius transformations, however, often cannot provide satisfactory alignment or are computationally too costly. This paper in- troduces a novel globally optimal algorithm for estimating Mobius transformations to align surfaces that are topologi- cal discs. Unlike previous methods, the proposed algorithm deterministically calculates the best transformation, with- out requiring good initializations. Further, our algorithm is also much faster than previous techniques in practice. We demonstrate the efficacy of our algorithm on data commonly used in computational conformal geometry.

Multi-label image classification aims to predict multiple labels for a single image which contains diverse content. By utilizing label correlations, various techniques have been developed to improve classification performance. However, current existing methods either neglect image features when exploiting label correlations or lack the ability to learn image-dependent conditional label structures. In this paper, we develop conditional graphical Lasso (CGL) to handle these challenges. CGL provides a unified Bayesian framework for structure and parameter learning conditioned on image features. We formulate the multi-label prediction as CGL inference problem, which is solved by a mean field variational approach. Meanwhile, CGL learning is efficient due to a tailored proximal gradient procedure by applying the maximum a posterior (MAP) methodology. CGL performs competitively for multi-label image classification on benchmark datasets MULAN scene, PASCAL VOC 2007 and PASCAL VOC 2012, compared with the state-of-the-art multi-label classification algorithms.

Photo aesthetics assessment is challenging. Deep convolutional neural network (ConvNet) methods have recently shown promising results for aesthetics assessment. The performance of these deep ConvNet methods, however, is often compromised by the constraint that the neural network only takes the fixed-size input. To accommodate this requirement, input images need to be transformed via cropping, scaling, or padding, which often damages image composition, reduces image resolution, or causes image distortion, thus compromising the aesthetics of the original images. In this paper, we present a composition-preserving deep ConvNet method that directly learns aesthetics features from the original input images without any image transformations. Specifically, our method adds an adaptive spatial pooling layer upon the regular convolution and pooling layers to directly handle input images with original sizes and aspect ratios. To allow for multi-scale feature extraction, we develop the Multi-Net Adaptive Spatial Pooling ConvNet architecture which consists of multiple sub-networks with different adaptive spatial pooling sizes and leverage a scene-based aggregation layer to effectively combine the predictions from multiple sub-networks. Our experiments on the large-scale aesthetics assessment benchmark (AVA) demonstrate that our method can significantly improve the state-of-the-art results in photo aesthetics assessment.

In many image-related tasks, learning expressive and discriminative representations of images is essential, and deep learning has been studied for automating the learning of such representations. Some user-centric tasks, such as image recommendations, call for effective representations of not only images but also preferences and intents of users over images. Such representations are termed hybrid and addressed via a deep learning approach in this paper. We design a dual-net deep network, in which the two sub-networks map input images and preferences of users into a same latent semantic space, and then the distances between images and users in the latent space are calculated to make decisions. We further propose a comparative deep learning (CDL) method to train the deep network, using a pair of images compared against one user to learn the pattern of their relative distances. The CDL embraces much more training data than naive deep learning, and thus achieves superior performance than the latter, with no cost of increasing network complexity. Experimental results with real-world data sets for image recommendations have shown the proposed dual-net network and CDL greatly outperform other state-of-the-art image recommendation solutions.

Bilinear models has been shown to achieve impressive performance on a wide range of visual tasks, such as semantic segmentation, fine grained recognition and face recognition. However, bilinear features are high dimensional, typically on the order of hundreds of thousands to a few million, which makes them impractical for subsequent analysis. We propose two compact bilinear representations with the same discriminative power as the full bilinear representation but with only a few thousand dimensions. Our compact representations allow back-propagation of classification errors enabling an end-to-end optimization of the visual recognition system. The compact bilinear representations are derived through a novel kernelized analysis of bilinear pooling which provide insights into the discriminative power of bilinear pooling, and a platform for further research in compact pooling methods. Experimentation illustrate the utility of the proposed representations for image classification and few-shot learning across several datasets.

Cross-modal similarity search is a problem about designing a search system supporting querying across content modalities, e.g., using an image to search for texts or using a text to search for images. This paper presents a compact coding solution for efficient search, with a focus on the quantization approach which has already shown the superior performance over the hashing solutions in the single-modal similarity search. We propose a cross modal quantization approach, which is among the early attempts to introduce quantization into cross-modal search. The major contribution lies in jointly learning the quantizers for both modalities through aligning the quantized representations for each pair of image and text belonging to a document. In addition, our approach simultaneously learns the common space for both modalities in which quantization is conducted to enable efficient and effective search using the Euclidean distance computed in the common space with fast distance table lookup. Experimental results compared with several competitive algorithms over three benchmark datasets demonstrate that the proposed approach achieves the state-of-the-art performance.