CVPR 2016 Papers

Capturing hyperspectral images requires expensive and specialized hardware that is not readily accessible to most users. Digital cameras, on the other hand, are significantly cheaper in comparison and can be easily purchased and used. In this paper, we present a framework for reconstructing hyperspectral images by using multiple consumer-level digital cameras. Our approach works by exploiting the different spectral sensitivities of different camera sensors. In particular, due to the differences in spectral sensitivities of the cameras, different cameras yield different RGB measurements for the same spectral signal. We introduce an algorithm that is able to combine and convert these different RGB measurements into a single hyperspectral image for both indoor and outdoor scenes. This camera-based approach allows hyperspectral imaging at a fraction of the cost of most existing hyperspectral hardware. We validate the accuracy of our reconstruction against ground truth hyperspectral images (using both synthetic and real cases) and show its usage on relighting applications.

Recent advances in neural networks have revolutionized computer vision, but these algorithms are still outperformed by humans. Could this performance gap be due to systematic differences between object representations in humans and machines? To answer this question we collected a large dataset of 26,675 perceived dissimilarity measurements from 2,801 visual objects across 269 human subjects, and used this dataset to train and test leading computational models. The best model (a combination of all models) accounted for 68% of the explainable variance. Importantly, all computational models showed systematic deviations from perception: (1) They underestimated perceptual distances between objects with symmetry or large area differences; (2) They overestimated perceptual distances between objects with shared features. Our results reveal critical elements missing in computer vision algorithms and point to explicit encoding of these properties in higher visual areas in the brain.

In real world applications, more and more data, for example, image/video data, are high dimensional and represented by multiple views which describe different perspectives of the data. Efficiently clustering such data is a challenge. To address this problem, this paper proposes a novel multi-view clustering method called Discriminatively Embedded K-Means (DEKM), which embeds the synchronous learning of multiple discriminative subspaces into multi-view K-Means clustering to construct a unified framework, and adaptively control the intercoordinations between these subspaces simultaneously. In this framework, we firstly design a weighted multi-view Linear Discriminant Analysis (LDA), and then develop an unsupervised optimization scheme to alternatively learn the common clustering indicator, multiple discriminative subspaces and weights for heterogeneous features with convergence. Comprehensive evaluations on three benchmark datasets and comparisons with several state-of-the-art multi-view clustering algorithms demonstrate the superiority of the proposed work.

We propose an explicitly discriminative and `simple' approach to generate invariance to nuisance transformations modeled as unitary. In practice, the approach works well to handle non-unitary transformations as well. Our theoretical results extend the reach of a recent theory of invariance to discriminative and kernelized features based on unitary kernels. As a special case, a single common framework can be used to generate subject-specific pose-invariant features for face recognition and vice-versa for pose estimation. We show that our main proposed method (DIKF) can perform well under very challenging large-scale semi-synthetic face matching and pose estimation protocols with unaligned faces using no land-marking whatsoever. We additionally benchmark on CMU MPIE and outperform previous work in almost all cases on off-angle face matching while we are on par with the previous state-of-the-art on the LFW unsupervised and image-restricted protocols, without any low-level image descriptors other than raw-pixels.

We propose a novel method for detecting horizontal vanishing points and the zenith vanishing point in man-made environments. The dominant trend in existing methods is to first find candidate vanishing points, then remove outliers by enforcing mutual orthogonality. Our method reverses this process: we propose a set of horizon line candidates and score each based on the vanishing points it contains. A key element of our approach is the use of global image context, extracted with a deep convolutional network, to constrain the set of candidates under consideration. Our method does not make a Manhattan-world assumption and can operate effectively on scenes with only a single horizontal vanishing point. We evaluate our approach on three benchmark datasets and achieve state-of-the-art performance on each. In addition, our approach is significantly faster than the previous best method.

Deep networks have recently enjoyed enormous success when applied to recognition and classification problems in computer vision [22, 32], but their use in graphics problems has been limited ([23, 7] are notable recent exceptions). In this work, we present a novel deep architecture that per- forms new view synthesis directly from pixels, trained from a large number of posed image sets. In contrast to tradi- tional approaches which consist of multiple complex stages of processing, each of which require careful tuning and can fail in unexpected ways, our system is trained end-to-end. The pixels from neighboring views of a scene are presented to the network which then directly produces the pixels of the unseen view. The benefits of our approach include gen- erality (we only require posed image sets and can easily apply our method to different domains), and high quality results on traditionally difficult scenes. We believe this is due to the end-to-end nature of our system which is able to plausibly generate pixels according to color, depth, and tex- ture priors learnt automatically from the training data. We show view interpolation results on imagery from the KITTI dataset [12], from data from [1] as well as on StreetView images. To our knowledge, our work is the first to apply deep learning to the problem of new view synthesis from sets of real-world, natural imagery.

This paper considers the task of articulated human pose estimation of multiple people in real world images. We propose an approach that jointly solves the tasks of detection and pose estimation: it infers the number of persons in a scene, identifies occluded body parts, and disambiguates body parts between people in close proximity of each other. This joint formulation is in contrast to previous strategies, that address the problem by first detecting people and subsequently estimating their body pose. We propose a partitioning and labeling formulation of a set of body-part hypotheses generated with CNN-based part detectors. Our formulation, an instance of an integer linear program, implicitly performs non-maximum suppression on the set of part candidates and groups them to form configurations of body parts respecting geometric and appearance constraints. Experiments on four different datasets demonstrate state-of-the-art results for both single person and multi person pose estimation.

We focus on the task of amodal 3D object detection in RGB-D images, which aims to produce a 3D bounding box of an object in metric form at its full extent. We introduce Deep Sliding Shapes, a 3D ConvNet formulation that takes a 3D volumetric scene from a RGB-D image as input and outputs 3D object bounding boxes. In our approach, we propose the first 3D Region Proposal Network (RPN) to learn objectness from geometric shapes and the first joint Object Recognition Network (ORN) to extract geometric features in 3D and color features in 2D. In particular, we handle objects of various sizes by training an amodal RPN at two different scales and an ORN to regress 3D bounding boxes. Experiments show that our algorithm outperforms the state-of-the-art by 13.8 in mAP and is 200x faster than the original Sliding Shapes.

Learning the distance metric between pairs of examples is of great importance for learning and visual recognition. With the remarkable success from the state of the art convolutional neural networks, recent works have shown promising results on discriminatively training the networks to learn semantic feature embeddings where similar examples are mapped close to each other and dissimilar examples are mapped farther apart. In this paper, we describe an algorithm for taking full advantage of the training batches in the neural network training by lifting the vector of pairwise distances within the batch to the matrix of pairwise distances. This step enables the algorithm to learn the state of the art feature embedding by optimizing a novel structured prediction objective for active hard negative mining on the lifted problem. Additionally, we collected Online Products dataset: 120k images of 23k classes of online products for metric learning. Our experiments on the CUB-200-2011, CARS196, and Online Products datasets demonstrate significant improvement over existing deep feature embedding methods on all experimented embedding sizes with the GoogLeNet network. The source code and the dataset are available at: https://github.com/rksltnl/Deep-Metric-Learning-CVPR16

Multi-task learning in Convolutional Networks has displayed remarkable success in the field of recognition. This success can be largely attributed to learning shared representations from multiple supervisory tasks. However, existing multi-task approaches rely on enumerating multiple network architectures specific to the tasks at hand, that do not generalize. In this paper, we propose a principled approach to learn shared representations in ConvNets using multi-task learning. Specifically, we propose a new sharing unit: "cross-stitch" unit. These units combine the activations from multiple networks and can be trained end-to-end. A network with cross-stitch units can learn an optimal combination of shared and task-specific representations. Our proposed method generalizes across multiple tasks and shows dramatically improved performance over baseline methods for categories with few training examples.

Pose Machines provide a sequential prediction framework for learning rich implicit spatial models. In this work we show a systematic design for how convolutional networks can be incorporated into the pose machine framework for learning image features and image-dependent spatial models for the task of pose estimation. The contribution of this paper is to implicitly model long-range dependencies between variables in structured prediction tasks such as articulated pose estimation. We achieve this by designing a sequential architecture composed of convolutional networks that directly operate on belief maps from previous stages, producing increasingly refined estimates for part locations, without the need for explicit graphical model-style inference. Our approach addresses the characteristic difficulty of vanishing gradients during training by providing a natural learning objective function that enforces intermediate supervision, thereby replenishing back-propagated gradients and conditioning the learning procedure. We demonstrate state-of-the-art performance and outperform competing methods on standard benchmarks including the MPII, LSP, and FLIC datasets.

This paper studies a combination of generative Markov random field (MRF) models and discriminatively trained deep convolutional neural networks (dCNNs) for synthesizing 2D images. The generative MRF acts on higher-levels of a dCNN feature pyramid, controling the image layout at an abstract level. We apply the method to both photographic and non-photo-realistic (artwork) synthesis tasks. The MRF regularizer prevents over-excitation artifacts and reduces implausible feature mixtures common to previous dCNN inversion approaches, permitting synthezing photographic content with increased visual plausibility. Unlike standard MRF-based texture synthesis, the combined system can both match and adapt local features with considerable variability, yielding results far out of reach of classic generative MRF methods.

Convolutional neural nets (CNNs) have demonstrated remarkable performance in recent history. Such approaches tend to work in a "unidirectional" bottom-up feed-forward fashion. However, practical experience and biological evidence tells us that feedback plays a crucial role, particularly for detailed spatial understanding tasks. This work explores "bidirectional" architectures that also reason with top-down feedback: neural units are influenced by both lower and higher-level units. We do so by treating units as rectified latent variables in a quadratic energy function, which can be seen as a hierarchical Rectified Gaussian model (RGs). We show that RGs can be optimized with a quadratic program (QP), that can in turn be optimized with a recurrent neural network (with rectified linear units). This allows RGs to be trained with GPU-optimized gradient descent. From a theoretical perspective, RGs help establish a connection between CNNs and hierarchical probabilistic models. From a practical perspective, RGs are well suited for detailed spatial tasks that can benefit from top-down reasoning. We illustrate them on the challenging task of keypoint localization under occlusions, where local bottom-up evidence may be misleading. We demonstrate state-of-the-art results on challenging benchmarks.

Most tracking-by-detection methods employ a local search window around the predicted object location in the current frame assuming the previous location is accurate, the trajectory is smooth, and the computational capacity permits a search radius that can accommodate the maximum speed yet small enough to reduce mismatches. These, however, may not be valid always, in particular for fast and irregularly moving objects. Here, we present an object tracker that is not limited to a local search window and has ability to probe efficiently the entire frame. Our method generates a small number of "high-quality" proposals by a novel instance-specific objectness measure and evaluates them against the object model that can be adopted from an existing tracking-by-detection approach as a core tracker. During the tracking process, we update the object model concentrating on hard false-positives supplied by the proposals, which help suppressing distractors caused by difficult background clutters, and learn how to re-rank proposals according to the object model. Since we reduce significantly the number of hypotheses the core tracker evaluates, we can use richer object descriptors and stronger detector. Our method outperforms most recent state-of-the-art trackers on popular tracking benchmarks, and provides improved robustness for fast moving objects as well as for ultra low-frame-rate videos.

We introduce a new technique that automatically generates diverse, visually compelling stylizations for a photograph in an unsupervised manner. We achieve this by learning style ranking for a given input using a large photo collection and selecting a diverse subset of matching styles for final style transfer. We also propose an improved technique that transfers the global color and tone of the chosen exemplars to the input photograph while avoiding the common visual artifacts produced by the existing style transfer methods. Together, our style selection and transfer techniques produce compelling, artifact-free results on a wide range of input photographs, and a user study shows that our results are preferred over other techniques.

We propose a method for visual question answering which combines an internal representation of the content of an image with information extracted from a general knowledge base to answer a broad range of image-based questions. This allows more complex questions to be answered using the predominant neural network-based approach than has previously been possible. It particularly allows questions to be asked about the contents of an image, even when the image itself does not contain the whole answer. The method constructs a textual representation of the semantic content of an image, and merges it with textual information sourced from a knowledge base, to develop a deeper understanding of the scene viewed. Priming a recurrent neural network with this combined information, and the submitted question, leads to a very flexible visual question answering approach. We are specifically able to answer questions posed in natural language, that refer to information not contained in the image. We demonstrate the effectiveness of our model on two publicly available datasets, Toronto COCO-QA and VQA, and show that it produces the best reported results in both cases.

Anticipating actions and objects before they start or appear is a difficult problem in computer vision with several real-world applications. This task is challenging partly because it requires leveraging extensive knowledge of the world that is difficult to write down. We believe that a promising resource for efficiently learning this knowledge is through readily available unlabeled video. We present a framework that capitalizes on temporal structure in unlabeled video to learn to anticipate human actions and objects. The key idea behind our approach is that we can train deep networks to predict the visual representation of images in the future. Visual representations are a promising prediction target because they encode images at a higher semantic level than pixels yet are automatic to compute. We then apply recognition algorithms on our predicted representation to anticipate objects and actions. We experimentally validate this idea on two datasets, anticipating actions one second in the future and objects five seconds in the future.

Manifold-valued datasets are widely encountered in many computer vision tasks. A non-linear analog of the PCA algorithm, called the Principal Geodesic Analysis (PGA) algorithm suited for data lying on Riemannian manifolds was reported in literature a decade ago. Since the objective function in the PGA algorithm is highly non-linear and hard to solve efficiently in general, researchers have proposed a linear approximation. Though this linear approximation is easy to compute, it lacks accuracy especially when the data exhibits a large variance. Recently, an alternative called the exact PGA was proposed which tries to solve the optimization without any linearization. For general Riemannian manifolds, though it yields a better accuracy than the original (linearized) PGA, for data that exhibit large variance, the optimization is not computationally efficient. In this paper, we propose an efficient exact PGA algorithm for constant curvature Riemannian manifolds (CCM-EPGA). The CCM-EPGA algorithm differs significantly from existing PGA algorithms in two aspects, (i) the distance between a given manifold-valued data point and the principal submanifold is computed analytically and thus no optimization is required as in the existing methods. (ii) Unlike the existing PGA algorithms, the descent into codimension-1 submanifolds does not require any optimization but is accomplished through the use of the Rimeannian inverse Exponential map and the parallel transport operations. We present theoretical and experimental results for constant curvature Riemannian manifolds depicting favorable performance of the CCM-EPGA algorithm compared to existing PGA algorithms. We also present data reconstruction from the principal components which has not been reported in literature in this setting.

Spectral embedding provides a framework for solving perceptual organization problems, including image segmentation and figure/ground organization. From an affinity matrix describing pairwise relationships between pixels, it clusters pixels into regions, and, using a complex-valued extension, orders pixels according to layer. We train a convolutional neural network (CNN) to directly predict the pairwise relationships that define this affinity matrix. Spectral embedding then resolves these predictions into a globally-consistent segmentation and figure/ground organization of the scene. Experiments demonstrate significant benefit to this direct coupling compared to prior works which use explicit intermediate stages, such as edge detection, on the pathway from image to affinities. Our results suggest spectral embedding as a powerful alternative to the conditional random field (CRF)-based globalization schemes typically coupled to deep neural networks.