CVPR 2016 Papers

The conventional sparse model relies on data representation in the form of vectors. It represents the vector-valued or vectorized one dimensional (1D) version of an signal as a highly sparse linear combination of basis atoms from a large dictionary. The 1D modeling, though simple, ignores the inherent structure and breaks the local correlation inside multidimensional (MD) signals. It also dramatically increases the demand of memory as well as computational resources especially when dealing with high dimensional signals. In this paper, we propose a new sparse model TenSR based on tensor for MD data representation along with the corresponding MD sparse coding and MD dictionary learning algorithms. The proposed TenSR model is able to well approximate the structure in each mode inherent in MD signals with a series of adaptive separable structure dictionaries via dictionary learning. The proposed MD sparse coding algorithm by proximal method further reduces the computational cost significantly. Experimental results with real world MD signals, i.e. 3D Multi-spectral images, show the proposed TenSR greatly reduces both the computational and memory costs with competitive performance in comparison with the state-of-the-art sparse representation methods. We believe our proposed TenSR model is a promising way to empower the sparse representation especially for large scale high order signals.

In view of the advantages of deep networks in producing useful representation, the generated features of different modality data (such as image, audio) can be jointly learned using Multimodal Restricted Boltzmann Machines (MRBM). Recently, audiovisual speech recognition based the MRBM has attracted much attention, and the MRBM shows its effectiveness in learning the joint representation across audiovisual modalities. However, the built networks have weakness in modeling the multimodal sequence which is the natural property of speech signal. In this paper, we will introduce a novel temporal multimodal deep learning architecture, named as Recurrent Temporal Multimodal RBM (RTMRBM), that models multimodal sequences by transforming the sequence of connected MRBMs into a probabilistic series model. Compared with existing multimodal networks, it's simple and efficient in learning temporal joint representation. We evaluate our model on audiovisual speech datasets, two public (AVLetters and AVLetters2) and one self-build. The experimental results demonstrate that our approach can obviously improve the accuracy of recognition compared with standard MRBM and the temporal model based on conditional RBM. In addition, RTMRBM still outperforms non-temporal multimodal deep networks in the presence of the weakness of long-term dependencies.

Dynamic events are often photographed by a number of people from different viewpoints at different times, resulting in an unconstrained set of images. Finding the corresponding moving features in each of the images allows us to extract information about objects of interest in the scene. Computing correspondence of moving features in such a set of images is considerably more challenging than computing correspondence in video due to possible significant differences in viewpoints and inconsistent timing between image captures. The prediction methods used in video for improving robustness and efficiency are not applicable to a set of still images. In this paper we propose a novel method to predict locations of an approximately linear moving feature point, given a small subset of correspondences and the temporal order of image captures. Our method extends the use of epipolar geometry to divide images into valid and invalid regions, termed Temporal Epipolar Regions (TERs). We formally prove that the location of a feature in a new image is restricted to valid TERs. We demonstrate the effectiveness of our method in reducing the search space for correspondence on both synthetic and challenging real world data, and show the improved matching.

We address temporal action localization in untrimmed long videos. This is important because videos in real applications are usually unconstrained and contain multiple action instances plus video content of background scenes or other activities. To address this challenging issue, we exploit the effectiveness of deep networks in temporal action localization via three segment-based 3D ConvNets: (1) a proposal network identifies candidate segments in a long video that may contain actions; (2) a classification network learns one-vs-all action classification model to serve as initialization for the localization network; and (3) a localization network fine-tunes the learned classification network to localize each action instance. We propose a novel loss function for the localization network to explicitly consider temporal overlap and achieve high temporal localization accuracy. In the end, only the proposal network and the localization network are used during prediction. On two large-scale benchmarks, our approach achieves significantly superior performances compared with other state-of-the-art systems: mAP increases from 1.7% to 7.4% on MEXaction2 and increases from 15.0% to 19.0% on THUMOS 2014.

In this paper we present a deep neural network topology that incorporates a simple to implement transformation-invariant pooling operator (TI-pooling). This operator is able to efficiently handle prior knowledge on nuisance variations in the data, such as rotation or scale changes. Most current methods usually make use of dataset augmentation to address this issue, but this requires larger number of model parameters and more training data, and results in significantly increased training time and larger chance of under- or overfitting. The main reason for these drawbacks is that that the learned model needs to capture adequate features for all the possible transformations of the input. On the other hand, we formulate features in convolutional neural networks to be transformation-invariant. We achieve that using parallel siamese architectures for the considered transformation set and applying the TI-pooling operator on their outputs before the fully-connected layers. We show that this topology internally finds the most optimal "canonical" instance of the input image for training and therefore limits the redundancy in learned features. This more efficient use of training data results in better performance on popular benchmark datasets with smaller number of parameters when comparing to standard convolutional neural networks with dataset augmentation and to other baselines.

This is a reaction to the poor performance of symmetry detection algorithms on real-world images, benchmarked since CVPR 2011. Our systematic study reveals significant difference between human labeled (reflection and rotation) symmetries on photos and the output of computer vision algorithms on the same photo set. We exploit this human-machine symmetry perception gap by proposing a novel symmetry-based Turing test. By leveraging a comprehensive user interface, we collected more than 78,000 symmetry labels from 400 Amazon Mechanical Turk raters on nearly 1,000 photos from the Microsoft COCO dataset. Using a set of ground-truth symmetries automatically generated from noisy human labels, the effectiveness of our work is evidenced by a separate test where over 96% success rate is achieved. We demonstrate statistically significant outcomes for using symmetry perception as a powerful, alternative, image-based reCAPTCHA.

In this paper, we address the problem of searching for semantically similar images from a large database. We present a compact coding approach, supervised quantization. Our approach simultaneously learns feature selection that linearly transforms the database points into a low-dimensional discriminative subspace, and quantizes the data points in the transformed space. The optimization criterion is that the quantized points not only approximate the transformed points accurately, but also are semantically separable: the points belonging to a class lie in a cluster that is not overlapped with other clusters corresponding to other classes, which is formulated as a classification problem. The experiments on several standard datasets show the superiority of our approach over the state-of-the art supervised hashing and unsupervised quantization algorithms.

Video summarization has unprecedented importance to help us digest, browse, and search today's ever-growing video collections. We propose a novel subset selection technique that leverages supervision in the form of human-created summaries to perform automatic keyframe-based video summarization. The main idea is to nonparametrically transfer summary structures from annotated videos to unseen test videos. We show how to extend our method to exploit semantic side information about the video's category/genre to guide the transfer process by those training videos semantically consistent with the test input. We also show how to generalize our method to subshot-based summarization, which not only reduces computational costs but also provides more flexible ways of defining visual similarity across subshots spanning several frames. We conduct extensive evaluation on several benchmarks and demonstrate promising results, outperforming existing methods in several settings.

This paper considers the problem of recovering a subspace arrangement from noisy samples, potentially corrupted with outliers. Our main result shows that this problem can be formulated as a convex semi-definite optimization problem subject to an additional rank constrain that involves only a very small number of variables. This is established by first reducing the problem to a (generically non-convex) quadratically constrained quadratic problem and then using its special sparse structure to find conditions guaranteeing that a suitably built convex relaxation is indeed exact. When combined with the commonly used nuclear norm relaxation for rank, the results above lead to computationally efficient algorithms with optimality guarantees. A salient feature of the proposed approach is its ability to incorporate existing a-priori information about the noise, co-ocurrences, and percentage of outliers. These results are illustrated with several examples where the proposed algorithm is shown to outperform existing approaches.

Visual recognition research often assumes a sufficient resolution of the region of interest (ROI). That is usually violated in practice, inspiring us to explore the Very Low Resolution Recognition (VLRR) problem. Typically, the ROI in a VLRR problem can be smaller than 16 x16 pixels, and is challenging to be recognized even by human experts. We attempt to solve the VLRR problem using deep learning methods. Taking advantage of techniques primarily in super resolution, domain adaptation and robust regression, we formulate a dedicated deep learning method and demonstrate how these techniques are incorporated step by step. Any extra complexity, when introduced, is fully justified by both analysis and simulation results. The resulting Robust Partially Coupled Networks achieves feature enhancement and recognition simultaneously. It allows for both the flexibility to combat the LR-HR domain mismatch, and the robustness to outliers. Finally, the effectiveness of the proposed models is evaluated on three different VLRR tasks, including face identification, digit recognition and font recognition, all of which obtain very impressive performances.

We propose a new way to train a structured output prediction model. More specifically, we train nonlinear data terms in a Gaussian Conditional Random Field (GCRF) by a generalized version of gradient boosting. The approach is evaluated on three challenging regression benchmarks: vessel detection, single image depth estimation and image inpainting. These experiments suggest that the proposed boosting framework matches or exceeds the state-of-the-art.

Learning powerful feature representations with CNNs is hard when training data are limited. Pre-training is one way to overcome this, but it requires large datasets sufficiently similar to the target domain. Another option is to design priors into the model, which can range from tuned hyperparameters to fully engineered representations like Scattering Networks. We combine these ideas into structured receptive field networks, a model which has a fixed filter basis and yet retains the flexibility of CNNs. This flexibility is achieved by expressing receptive fields in CNNs as a weighted sum over a fixed basis which is similar in spirit to Scattering Networks. The key difference is that we learn arbitrary effective filter sets from the basis rather than modeling the filters. This approach explicitly connects classical multiscale image analysis with general CNNs. With structured receptive field networks, we improve considerably over unstructured CNNs for small and medium dataset scenarios as well as over Scattering for large datasets. We validate our findings on ILSVRC2012, Cifar-10, Cifar-100 and MNIST. As a realistic small dataset example, we show state-of-the-art classification results on popular 3D MRI brain-disease datasets where pre-training is difficult due to a lack of large public datasets in a similar domain.

Feature matching is a fundamental process in a variety of computer vision tasks. Beyond the standard L2 metric, various methods to measure similarity between features have been proposed mainly on the assumption that the features are defined in a histogram form. On the other hand, in a field of image quality assessment, SSIM produces effective similarity between images, taking the place of L2 metric. In this paper, we propose a feature similarity measurement method based on the SSIM. Unlike the previous methods, the proposed method is built on not a histogram form but a tensor structure of a feature array extracted such as on spatial grids, in order to construct effective SSIM-based similarity measure of high robustness which is a key requirement in feature matching. In addition, we provide the explicit feature map such that the proposed similarity metric is embedded as a dot product. It contributes to significant speedup in similarity measurement as well as to feature transformation toward an effective vector form to which linear classifiers are directly applicable. In the experiments on various tasks, the proposed method exhibits favorable performance in both feature matching and classification.

Incremental Structure-from-Motion is a prevalent strategy for 3D reconstruction from unordered image collections. While incremental reconstruction systems have tremendously advanced in all regards, robustness, accuracy, completeness, and scalability remain the key problems towards building a truly general-purpose pipeline. We propose a new SfM technique that improves upon the state of the art to make a further step towards this ultimate goal. The full reconstruction pipeline is released to the public as an open-source implementation.

Rich semantic relations are important in a variety of visual recognition problems. As a concrete example, group activity recognition involves the interactions and relative spatial relations of a set of people in a scene. State of the art recognition methods center on deep learning approaches for training highly effective, complex classifiers for interpreting images. However, bridging the relatively low-level concepts output by these methods to interpret higher-level compositional scenes remains a challenge. Graphical models are a standard tool for this task. In this paper, we propose a method to integrate graphical models and deep neural networks into a joint framework. Instead of using a traditional inference method, we use a sequential inference modeled by a recurrent neural network. Beyond this, the appropriate structure for inference can be learned by imposing gates on edges between nodes. Empirical results on group activity recognition demonstrate the potential of this model to handle highly structured learning tasks.

This paper shows for the first time that is possible to reconstruct the position of rigid objects and to jointly recover affine camera calibration solely from a set of object detections in a video sequence. In practice, this work can be considered as the extension of Tomasi and Kanade factorization method using objects. Instead of using points to form a rank constrained measurement matrix, we can form a matrix with similar rank properties using 2D object detection proposals. In detail, we first fit an ellipse onto the image plane at each bounding box as given by the object detector. The collection of all the ellipses in the dual space is used to create a measurement matrix that gives a specific rank constraint. This matrix can be factorised and metrically upgraded in order to provide the affine camera matrices and the 3D position of the objects as an ellipsoid. Moreover, we recover the full 3D quadric thus giving additional information about object occupancy and 3D pose. Finally, we also show that 2D points measurements can be seamlessly included in the framework to reduce the number of objects required. This last aspect unifies the classical point-based Tomasi and Kanade approach with objects in a unique framework. Experiments with synthetic and real data show the feasibility of our approach for the affine camera case.

In this paper, we propose a novel structural correlation filter (SCF) model for robust visual tracking. The proposed SCF model takes part-based tracking strategies into account in a correlation filter tracker, and exploits circular shifts of all parts for their motion modeling to preserve target object structure. Compared with existing correlation filter trackers, our proposed tracker has several advantages: (1) Due to the part strategy, the learned structural correlation filters are less sensitive to partial occlusion, and have computational efficiency and robustness. (2) The learned filters are able to not only distinguish the parts from the background as the traditional correlation filters, but also exploit the intrinsic relationship among local parts via spatial constraints to preserve object structure. (3) The learned correlation filters not only make most parts share similar motion, but also tolerate outlier parts that have different motion. Both qualitative and quantitative evaluations on challenging benchmark image sequences demonstrate that the proposed SCF tracking algorithm performs favorably against several state-of-the-art methods.

Consumer devices with stereo cameras have become popular because of their low-cost depth sensing capability. However, those systems usually suffer from low imaging quality and inaccurate depth acquisition under low-light conditions. To address the problem, we present a new stereo matching method with a color and monochrome camera pair. We focus on the fundamental trade-off that monochrome cameras have much better light-efficiency than color-filtered cameras. Our key ideas involve compensating for the radiometric difference between two cross-spectral images and taking full advantage of complementary data. Consequently, our method produces both an accurate depth map and high-quality images, which are applicable for various depth-aware image processing. Our method is evaluated using various datasets and the performance of our depth estimation consistently outperforms state-of-the-art methods.

Correlation Filter-based trackers have recently achieved excellent performance, showing great robustness to challenging situations exhibiting motion blur and illumination changes. However, since the model that they learn depends strongly on the spatial layout of the tracked object, they are notoriously sensitive to deformation. Models based on colour statistics have complementary traits: they cope well with variation in shape, but suffer when illumination is not consistent throughout a sequence. Moreover, colour distributions alone can be insufficiently discriminative. In this paper, we show that a simple tracker combining complementary cues in a ridge regression framework can operate faster than 80 FPS and outperform not only all entries in the popular VOT14 competition, but also recent and far more sophisticated trackers according to multiple benchmarks.