CVPR 2017 论文列表

Deep Auto-Encoder (DAE) has shown its promising power in high-level representation learning. From the perspective of manifold learning, we propose a graph regularized deep neural network (GR-DNN) to endue traditional DAEs with the ability of retaining local geometric structure. A deep-structured regularizer is formulated upon multi-layer perceptions to capture this structure. The robust and discriminative embedding space is learned to simultaneously preserve the high-level semantics and the geometric structure within local manifold tangent space. Theoretical analysis presents the close relationship between the proposed graph regularizer and the graph Laplacian regularizer in terms of the optimization objective. We also alleviate the growth of the network complexity by introducing the anchor-based bipartite graph, which guarantees the good scalability for large scale data. The experiments on four datasets show the comparable results of the proposed GR-DNN with the state-of-the-art methods.

We introduce a novel technique for knowledge transfer, where knowledge from a pretrained deep neural network (DNN) is distilled and transferred to another DNN. As the DNN performs a mapping from the input space to the output space through many layers sequentially, we define the distilled knowledge to be transferred in terms of flow between layers, which is calculated by computing the inner product between features from two layers. When we compare the student DNN and the original network with the same size as the student DNN but trained without a teacher network, the proposed method of transferring the distilled knowledge as the flow between two layers exhibits three important phenomena : (1) the student DNN that learns the distilled knowledge is optimized much faster than the original model; (2) the student DNN outperforms the original DNN; and (3) the student DNN can learn the distilled knowledge from a teacher DNN that is trained at a different task, and the student DNN outperforms the original DNN that is trained from scratch.

We consider the problem of depth-based robust 3D facial pose tracking under unconstrained scenarios with heavy occlusions and arbitrary facial expression variations. Unlike the previous depth-based discriminative or data-driven methods that require sophisticated training or manual intervention, we propose a generative framework that unifies pose tracking and face model adaptation on-the-fly. Particularly, we propose a statistical 3D face model that owns the flexibility to generate and predict the distribution and uncertainty underlying the face model. Moreover, unlike prior arts employing the ICP-based facial pose estimation, we propose a ray visibility constraint that regularizes the pose based on the face model's visibility against the input point cloud, which augments the robustness against the occlusions. The experimental results on Biwi and ICT-3DHP datasets reveal that the proposed framework is effective and outperforms the state-of-the-art depth-based methods.

This paper introduces a general setting for the construction of data fidelity metrics between oriented or non-oriented geometric shapes like curves, curve sets or surfaces. These metrics are based on the representation of shapes as distributions of their local tangent or normal vectors and the definition of reproducing kernels on these spaces. The construction, that combines in one common setting and extends the previous frameworks of currents and varifolds, provides a very large class of kernel metrics which can be easily computed without requiring any kind of parametrization of shapes and which are smooth enough to give robustness to certain imperfections that could result e.g. from bad segmentation. We then give a sense, with synthetic examples, of the versatility and potentialities of such metrics when used in various problems like shape comparison, clustering and diffeomorphic registration.

We propose a general dual ascent (message passing) framework for Lagrangean (dual) decomposition of combinatorial problems. Although methods of this type have shown their efficiency for a number of problems, so far there was no general algorithm applicable to multiple problem types. In this work, we propose such a general algorithm. It depends on several parameters, which can be used to optimize its performance in each particular setting. We demonstrate efficiency of our method on the graph matching and the multicut problems, where it outperforms state-of-the-art solvers including those based on the subgradient optimization and off-the-shelf linear programming solvers.

Social relations are the foundation of human daily life. Developing techniques to analyze such relations from visual data bears great potential to build machines that better understand us and are capable of interacting with us at a social level. Previous investigations have remained partial due to the overwhelming diversity and complexity of the topic and consequently have only focused on a handful of social relations. In this paper, we argue that the domain-based theory from social psychology is a great starting point to systematically approach this problem. The theory provides coverage of all aspects of social relations and equally is concrete and predictive about the visual attributes and behaviors defining the relations included in each domain. We provide the first dataset built on this holistic conceptualization of social life that is composed of a hierarchical label space of social domains and social relations. We also contribute the first models to recognize such domains and relations and find superior performance for attribute based features. Beyond the encouraging performance of the attribute based approach, we also find interpretable features that are in accordance with the predictions from social psychology literature. Beyond our findings, we believe that our contributions more tightly interleave visual recognition and social psychology theory that has the potential to complement the theoretical work in the area with empirical and data-driven models of social life.

Regression based facial landmark detection methods usually learns a series of regression functions to update the landmark positions from an initial estimation. Most of existing approaches focus on learning effective mapping functions with robust image features to improve performance. The approach to dealing with the initialization issue, however, receives relatively fewer attentions. In this paper, we present a deep regression architecture with two-stage re-initialization to explicitly deal with the initialization problem. At the global stage, given an image with a rough face detection result, the full face region is firstly re-initialized by a supervised spatial transformer network to a canonical shape state and then trained to regress a coarse landmark estimation. At the local stage, different face parts are further separately re-initialized to their own canonical shape states, followed by another regression subnetwork to get the final estimation. Our proposed deep architecture is trained from end to end and obtains promising results using different kinds of unstable initialization. It also achieves superior performances over many competing algorithms.

A novel dataset for benchmarking image-based localization is presented. With increasing research interests in visual place recognition and localization, several datasets have been published in the past few years. One of the evident limitations of existing datasets is that precise ground truth camera poses of query images are not available in a meaningful 3D metric system. This is in part due to the underlying 3D models of these datasets are reconstructed from Structure from Motion methods. So far little attention has been paid to metric evaluations of localization accuracy. In this paper we address the problem of whether state-of-the-art visual localization techniques can be applied to tasks with demanding accuracy requirements. We acquired training data for a large indoor environment with cameras and a LiDAR scanner. In addition, we collected over 2000 query images with cell phone cameras. Using LiDAR point clouds as a reference, we employed a semi-automatic approach to estimate the 6 degrees of freedom camera poses precisely in the world coordinate system. The proposed dataset enables us to quantitatively assess the performance of various algorithms using a fair and intuitive metric.

While deep convolutional neural networks frequently approach or exceed human-level performance in benchmark tasks involving static images, extending this success to moving images is not straightforward. Video understanding is of interest for many applications, including content recommendation, prediction, summarization, event/object detection, and understanding human visual perception. However, many domains lack sufficient data to explore and perfect video models. In order to address the need for a simple, quantitative benchmark for developing and understanding video, we present MovieFIB, a fill-in-the-blank question-answering dataset with over 300,000 examples, based on descriptive video annotations for the visually impaired. In addition to presenting statistics and a description of the dataset, we perform a detailed analysis of 5 different models' predictions, and compare these with human performance. We investigate the relative importance of language, static (2D) visual features, and moving (3D) visual features; the effects of increasing dataset size, the number of frames sampled; and of vocabulary size. We illustrate that: this task is not solvable by a language model alone; our model combining 2D and 3D visual information indeed provides the best result; all models perform significantly worse than human-level. We provide human evaluation for responses given by different models and find that accuracy on the MovieFIB evaluation corresponds well with human judgment. We suggest avenues for improving video models, and hope that the MovieFIB challenge can be useful for measuring and encouraging progress in this very interesting field.

Recently, DNN model compression based on network architecture design, e.g., SqueezeNet, attracted a lot attention. No accuracy drop on image classification is observed on these extremely compact networks, compared to well-known models. An emerging question, however, is whether these model compression techniques hurt DNN's learning ability other than classifying images on a single dataset. Our preliminary experiment shows that these compression methods could degrade domain adaptation (DA) ability, though the classification performance is preserved. Therefore, we propose a new compact network architecture and unsupervised DA method in this paper. The DNN is built on a new basic module Conv-M which provides more diverse feature extractors without significantly increasing parameters. The unified framework of our DA method will simultaneously learn invariance across domains, reduce divergence of feature representations, and adapt label prediction. Our DNN has 4.1M parameters, which is only 6.7% of AlexNet or 59% of GoogLeNet. Experiments show that our DNN obtains GoogLeNet-level accuracy both on classification and DA, and our DA method slightly outperforms previous competitive ones. Put all together, our DA strategy based on our DNN achieves state-of-the-art on sixteen of total eighteen DA tasks on popular Office-31 and Office-Caltech datasets.

We propose a combinatorial solution for the problem of non-rigidly matching a 3D shape to 3D image data. To this end, we model the shape as a triangular mesh and allow each triangle of this mesh to be rigidly transformed to achieve a suitable matching to the image. By penalising the distance and the relative rotation between neighbouring triangles our matching compromises between the image and the shape information. In this paper, we resolve two major challenges: Firstly, we address the resulting large and NP-hard combinatorial problem with a suitable graph-theoretic approach. Secondly, we propose an efficient discretisation of the unbounded 6-dimensional Lie group SE(3). To our knowledge this is the first combinatorial formulation for non-rigid 3D shape-to-image matching. In contrast to existing local (gradient descent) optimisation methods, we obtain solutions that do not require a good initialisation and that are within a bound of the optimal solution. We evaluate the proposed combinatorial method on the two problems of non-rigid 3D shape-to-shape and non-rigid 3D shape-to-image registration and demonstrate that it provides promising results.

We present a new insight into the systematic generation of minimal solvers in computer vision, which leads to smaller and faster solvers. Many minimal problem formulations are coupled sets of linear and polynomial equations where image measurements enter the linear equations only. We show that it is useful to solve such systems by first eliminating all the unknowns that do not appear in the linear equations and then extending solutions to the rest of unknowns. This can be generalized to fully non-linear systems by linearization via lifting. We demonstrate that this approach leads to more efficient solvers in three problems of partially calibrated relative camera pose computation with unknown focal length and/or radial distortion. Our approach also generates new interesting constraints on the fundamental matrices of partially calibrated cameras, which were not known before.

Superpixel segmentation of 2D image has been widely used in many computer vision tasks. However, limited to the Gaussian imaging principle, there is not a thorough segmentation solution to the ambiguity in defocus and occlusion boundary areas. In this paper, we consider the essential element of image pixel, i.e., rays in the light space and propose light field superpixel (LFSP) segmentation to eliminate the ambiguity. The LFSP is first defined mathematically and then a refocus-invariant metric named LFSP self-similarity is proposed to evaluate the segmentation performance. By building a clique system containing 80 neighbors in light field, a robust refocus-invariant LFSP segmentation algorithm is developed. Experimental results on both synthetic and real light field datasets demonstrate the advantages over the state-of-the-arts in terms of traditional evaluation metrics. Additionally the LFSP self-similarity evaluation under different light field refocus levels shows the refocus-invariance of the proposed algorithm.

Matching local geometric features on real-world depth images is a challenging task due to the noisy, low-resolution, and incomplete nature of 3D scan data. These difficulties limit the performance of current state-of-art methods, which are typically based on histograms over geometric properties. In this paper, we present 3DMatch, a data-driven model that learns a local volumetric patch descriptor for establishing correspondences between partial 3D data. To amass training data for our model, we propose a self-supervised feature learning method that leverages the millions of correspondence labels found in existing RGB-D reconstructions. Experiments show that our descriptor is not only able to match local geometry in new scenes for reconstruction, but also generalize to different tasks and spatial scales (e.g. instance-level object model alignment for the Amazon Picking Challenge, and mesh surface correspondence). Results show that 3DMatch consistently outperforms other state-of-the-art approaches by a significant margin. Code, data, benchmarks, and pre-trained models are available online at http://3dmatch.cs.princeton.edu

This paper introduces a deep architecture for segmenting 3D objects into their labeled semantic parts. Our architecture combines image-based Fully Convolutional Networks (FCNs) and surface-based Conditional Random Fields (CRFs) to yield coherent segmentations of 3D shapes. The image-based FCNs are used for efficient view-based reasoning about 3D object parts. Through a special projection layer, FCN outputs are effectively aggregated across multiple views and scales, then are projected onto the 3D object surfaces. Finally, a surface-based CRF combines the projected outputs with geometric consistency cues to yield coherent segmentations. The whole architecture (multi-view FCNs and CRF) is trained end-to-end. Our approach significantly outperforms the existing state-of-the-art methods in the currently largest segmentation benchmark (ShapeNet). Finally, we demonstrate promising segmentation results on noisy 3D shapes acquired from consumer-grade depth cameras.

We present an algorithm for registration between a large-scale point cloud and a close-proximity scanned point cloud, providing a localization solution that is fully independent of prior information about the initial positions of the two point cloud coordinate systems. The algorithm, denoted LORAX, selects super-points--local subsets of points--and describes the geometric structure of each with a low-dimensional descriptor. These descriptors are then used to infer potential matching regions for an efficient coarse registration process, followed by a fine-tuning stage. The set of super-points is selected by covering the point clouds with overlapping spheres, and then filtering out those of low-quality or nonsalient regions. The descriptors are computed using state-of-the-art unsupervised machine learning, utilizing the technology of deep neural network based auto-encoders. Abstract This novel framework provides a strong alternative to the common practice of using manually designed key-point descriptors for coarse point cloud registration. Utilizing super-points instead of key-points allows the available geometrical data to be better exploited to find the correct transformation. Encoding local 3D geometric structures using a deep neural network auto-encoder instead of traditional descriptors continues the trend seen in other computer vision applications and indeed leads to superior results. The algorithm is tested on challenging point cloud registration datasets, and its advantages over previous approaches as well as its robustness to density changes, noise, and missing data are shown.

There has been significant work on learning realistic, articulated, 3D models of the human body. In contrast, there are few such models of animals, despite many applications. The main challenge is that animals are much less cooperative than humans. The best human body models are learned from thousands of 3D scans of people in specific poses, which is infeasible with live animals. Consequently, we learn our model from a small set of 3D scans of toy figurines in arbitrary poses. We employ a novel part-based shape model to compute an initial registration to the scans. We then normalize their pose, learn a statistical shape model, and refine the registrations and the model together. In this way, we accurately align animal scans from different quadruped families with very different shapes and poses. With the registration to a common template we learn a shape space representing animals including lions, cats, dogs, horses, cows and hippos. Animal shapes can be sampled from the model, posed, animated, and fit to data. We demonstrate generalization by fitting it to images of real animals including species not seen in training.

This paper addresses the problem of 3D human pose estimation from a single image. We follow a standard two-step pipeline by first detecting the 2D position of the N body joints, and then using these observations to infer 3D pose. For the first step, we use a recent CNN-based detector. For the second step, most existing approaches perform 2N-to-3N regression of the Cartesian joint coordinates. We show that more precise pose estimates can be obtained by representing both the 2D and 3D human poses using NxN distance matrices, and formulating the problem as a 2D-to-3D distance matrix regression. For learning such a regressor we leverage on simple Neural Network architectures, which by construction, enforce positivity and symmetry of the predicted matrices. The approach has also the advantage to naturally handle missing observations and allowing to hypothesize the position of non-observed joints. Quantitative results on Humaneva and Human3.6M datasets demonstrate consistent performance gains over state-of-the-art. Qualitative evaluation on the images in-the-wild of the LSP dataset, using the regressor learned on Human3.6M, reveals very promising generalization results.

We explore 3D human pose estimation from a single RGB image. While many approaches try to directly predict 3D pose from image measurements, we explore a simple architecture that reasons through intermediate 2D pose predictions. Our approach is based on two key observations (1) Deep neural nets have revolutionized 2D pose estimation, producing accurate 2D predictions even for poses with self-occlusions (2) "Big-data"sets of 3D mocap data are now readily available, making it tempting to "lift" predicted 2D poses to 3D through simple memorization (e.g., nearest neighbors). The resulting architecture is straightforward to implement with off-the-shelf 2D pose estimation systems and 3D mocap libraries. Importantly, we demonstratethatsuchmethodsoutperformalmostallstate-of-theart 3D pose estimation systems, most of which directly try to regress 3D pose from 2D measurements.