CVPR 2015 论文列表

In this paper, we propose a novel algorithm that infers the 3D layout of building facades from a single 2D image of an urban scene. Different from existing methods that only yield coarse orientation labels or qualitative block approximations, our algorithm quantitatively reconstructs building facades in 3D space using a set of planes mutually related by 3D geometric constraints. Each plane is characterized by a continuous orientation vector and a depth distribution. An optimal solution is reached through inter-planar interactions. Due to the quantitative and plane-based nature of our geometric reasoning, our model is more expressive and informative than existing approaches. Experiments show that our method compares competitively with the state of the art on both 2D and 3D measures, while yielding a richer interpretation of the 3D scene behind the image.

We tackle the problem of single image depth estimation, which, without additional knowledge, suffers from many ambiguities. Unlike previous approaches that only reason locally, we propose to exploit the global structure of the scene to estimate its depth. To this end, we introduce a hierarchical representation of the scene, which models local depth jointly with mid-level and global scene structures. We formulate single image depth estimation as inference in a graphical model whose edges let us encode the interactions within and across the different layers of our hierarchy. Our method therefore still produces detailed depth estimates, but also leverages higher-level information about the scene. We demonstrate the benefits of our approach over local depth estimation methods on standard indoor datasets.

In this paper, we address the problem of model-free online object tracking based on color representations. According to the findings of recent benchmark evaluations, such trackers often tend to drift towards regions which exhibit a similar appearance compared to the object of interest. To overcome this limitation, we propose an efficient discriminative object model which allows us to identify potentially distracting regions in advance. Furthermore, we exploit this knowledge to adapt the object representation beforehand so that distractors are suppressed and the risk of drifting is significantly reduced. We evaluate our approach on recent online tracking benchmark datasets demonstrating state-of-the-art results. In particular, our approach performs favorably both in terms of accuracy and robustness compared to recent tracking algorithms. Moreover, the proposed approach allows for an efficient implementation to enable online object tracking in real-time.

Recent advances in object detection have exploited object proposals to speed up object searching. However, many of existing object proposal generators have strong localization bias or require computationally expensive diversification strategies. In this paper, we present an effective approach to address these issues. We first propose a simple and useful localization bias measure, called superpixel tightness. Based on the characteristics of superpixel tightness distribution, we propose an effective method, namely multi-thresholding straddling expansion (MTSE) to reduce localization bias via fast diversification. Our method is essentially a box refinement process, which is intuitive and beneficial, but seldom exploited before. The greatest benefit of our method is that it can be integrated into any existing model to achieve consistently high recall across various intersection over union thresholds. Experiments on PASCAL VOC dataset demonstrates that our approach improves numerous existing models significantly with little computational overhead.

Object detection systems based on the deep convolutional neural network (CNN) have recently made ground- breaking advances on several object detection benchmarks. While the features learned by these high-capacity neural networks are discriminative for categorization, inaccurate localization is still a major source of error for detection. Building upon high-capacity CNN architectures, we address the localization problem by 1) using a search algorithm based on Bayesian optimization that sequentially proposes candidate regions for an object bounding box, and 2) training the CNN with a structured loss that explicitly penalizes the localization inaccuracy. In experiments, we demonstrated that each of the proposed methods improves the detection performance over the baseline method on PASCAL VOC 2007 and 2012 datasets. Furthermore, two methods are complementary and significantly outperform the previous state-of-the-art when combined.

For some images, descriptions written by multiple people are consistent with each other. But for other images, descriptions across people vary considerably. In other words, some images are specific - they elicit consistent descriptions from different people - while other images are ambiguous. Applications involving images and text can benefit from an understanding of which images are specific and which ones are ambiguous. For instance, consider text-based image retrieval. If a query description is moderately similar to the caption (or reference description) of an ambiguous image, that query may be considered a decent match to the image. But if the image is very specific, a moderate similarity between the query and the reference description may not be sufficient to retrieve the image. In this paper, we introduce the notion of image specificity. We present two mechanisms to measure specificity given multiple descriptions of an image: an automated measure and a measure that relies on human judgement. We analyze image specificity with respect to image content and properties to better understand what makes an image specific. We then train models to automatically predict the specificity of an image from image features alone without requiring textual descriptions of the image. Finally, we show that modeling image specificity leads to improvements in a text-based image retrieval application.

Consider the following scenario between a human user and the computer. Given an image, the user thinks of an object to be segmented within this picture, but is only allowed to provide binary inputs to the computer (yes or no). In these conditions, can the computer guess this hidden segmentation by asking well-chosen questions to the user? We introduce a strategy for the computer to increase the accuracy of its guess in a minimal number of questions. At each turn, the current belief about the answer is encoded in a Bayesian fashion via a probability distribution over the set of all possible segmentations. To efficiently handle this huge space, the distribution is approximated by sampling representative segmentations using an adapted version of the Metropolis-Hastings algorithm, whose proposal moves build on a geodesic distance transform segmentation method. Following a dichotomic search, the question halving the weighted set of samples is finally picked, and the provided answer is used to update the belief for the upcoming rounds. The performance of this strategy is assessed on three publicly available datasets with diverse visual properties. Our approach shows to be a tractable and very adaptive solution to this problem.

This paper develops a novel framework for semantic image retrieval based on the notion of a scene graph. Our scene graphs represent objects ("man", "boat"), attributes of objects ("boat is white") and relationships between objects ("man standing on boat"). We use these scene graphs as queries to retrieve semantically related images. To this end, we design a conditional random field model that reasons about possible groundings of scene graphs to test images. The likelihoods of these groundings are used as ranking scores for retrieval. We introduce a novel dataset of 5,000 human-generated scene graphs grounded to images and use this dataset to evaluate our method for image retrieval. In particular, we evaluate retrieval using full scene graphs and small scene subgraphs, and show that our method outperforms retrieval methods that use only objects or low-level image features. In addition, we show that our full model can be used to improve object localization compared to baseline methods.

The over-segmentation of images into atomic regions has become a standard and powerful tool in Vision. Traditional superpixel methods, that operate at the pixel level, cannot directly capture the geometric information disseminated into the images. We propose an alternative to these methods by operating at the level of geometric shapes. Our algorithm partitions images into convex polygons. It presents several interesting properties in terms of geometric guarantees, region compactness and scalability. The overall strategy consists in building a Voronoi diagram that conforms to preliminarily detected line-segments, before homogenizing the partition by spatial point process distributed over the image gradient. Our method is particularly adapted to images with strong geometric signatures, typically man-made objects and environments. We show the potential of our approach with experiments on large-scale images and comparisons with state-of-the-art superpixel methods.

This paper presents a nonparametric scene parsing approach that improves the overall accuracy, as well as the coverage of foreground classes in scene images. We first improve the label likelihood estimates at superpixels by merging likelihood scores from different probabilistic classifiers. This boosts the classification performance and enriches the representation of less-represented classes. Our second contribution consists of incorporating semantic context in the parsing process through global label costs. Our method does not rely on image retrieval sets but rather assigns a global likelihood estimate to each label, which is plugged into the overall energy function. We evaluate our system on two large-scale datasets, SIFTflow and LMSun. We achieve state-of-the-art performance on the SIFTflow dataset and near-record results on LMSun.

This paper proposes a new image denoising approach using adaptive signal modeling and adaptive soft-thresholding. It improves the image quality by regularizing all the patches in image based on distribution modeling in transform domain. Instead of using a global model for all patches, it employs content adaptive models to address the non-stationarity of image signals. The distribution model of each patch is estimated individually and can vary for different transform bands and for different patch locations. In particular, we allow the distribution model for each individual patch to have non-zero expectation. To estimate the expectation and variance parameters for the transform bands of a particular patch, we exploit the non-local correlation of image and collect a set of similar patches as data samples to form the distribution. Irrelevant patches are excluded so that this non-local based modeling is more accurate than global modeling. Adaptive soft-thresholding is employed since we observed that the distribution of non-local samples can be approximated by Laplacian distribution. Experimental results show that the proposed scheme outperforms the state-of-the-art denoising methods such as BM3D and CSR in both the PSNR and the perceptual quality.

This paper addresses the illumination and reflectance spectra separation (IRSS) problem of a hyperspectral image captured under general spectral illumination. The huge amount of pixels in a hypersepctral image poses tremendous challenges on computational efficiency, yet in turn offers greater color variety that might be utilized to improve separation accuracy and relax the restrictive subspace illumination assumption in existing works. We show that this IRSS problem can be modeled into a low-rank matrix factorization problem, and prove that the separation is unique up to an unknown scale under the standard low-dimensionality assumption of reflectance. We also develop a scalable algorithm for this separation task that works in the presence of model error and image noise. Experiments on both synthetic data and real images have demonstrated that our separation results are sufficiently accurate, and can benefit some important applications, such as spectra relighting and illumination swapping.

Recognition algorithms based on convolutional networks (CNNs) typically use the output of the last layer as feature representation. However, the information in this layer may be too coarse to allow precise localization. On the contrary, earlier layers may be precise in localization but will not capture semantics. To get the best of both worlds, we define the hypercolumn at a pixel as the vector of activations of all CNN units above that pixel. Using hypercolumns as pixel descriptors, we show results on three fine-grained localization tasks: simultaneous detection and segmentation[20], where we improve state-of-the-art from 49.7 mean AP^r[20] to 59.0, keypoint localization, where we get a 3.3 point boost over [19] and part labeling, where we show a 6.6 point gain over a strong baseline.

Deep convolutional neural networks (CNN) have seen tremendous success in large-scale generic object recognition. In comparison with generic object recognition, fine-grained image classification (FGIC) is much more challenging because (i) fine-grained labeled data is much more expensive to acquire (usually requiring domain expertise); (ii) there exists large intra-class and small inter-class variance. Most recent work exploiting deep CNN for image recognition with small training data adopts a simple strategy: pre-train a deep CNN on a large-scale external dataset (e.g., ImageNet) and fine-tune on the small-scale target data to fit the specific classification task. In this paper, beyond the fine-tuning strategy, we propose a systematic framework of learning a deep CNN that addresses the challenges from two new perspectives: (i) identifying easily annotated hyper-classes inherent in the fine-grained data and acquiring a large number of hyper-class-labeled images from readily available external sources (e.g., image search engines), and formulating the problem into multi-task learning; (ii) a novel learning model by exploiting a regularization between the fine-grained recognition model and the hyper-class recognition model. We demonstrate the success of the proposed framework on two small-scale fine-grained datasets (Stanford Dogs and Stanford Cars) and on a large-scale car dataset that we collected.

Detailed analysis of human action, such as action classification, detection and localization has received increasing attention from the community; datasets like JHMDB have made it plausible to conduct studies analyzing the impact that such deeper information has on the greater action understanding problem. However, detailed automatic segmentation of human action has comparatively been unexplored. In this paper, we take a step in that direction and propose a hierarchical MRF model to bridge low-level video fragments with high-level human motion and appearance; novel higher-order potentials connect different levels of the supervoxel hierarchy to enforce the consistency of the human segmentation by pulling from different segment-scales. Our single layer model significantly outperforms the current state-of-the-art on actionness, and our full model improves upon the single layer baselines in action segmentation.

Visual saliency has been shown to depend on the unpredictability of the visual stimulus given its surround. Various previous works have advocated the equivalence between stimulus saliency and uncompressibility. We propose a direct measure of this quantity, namely the number of bits required by an optimal video compressor to encode a given video patch, and show that features derived from this measure are highly predictive of eye fixations. To account for global saliency effects, these are embedded in a Markov random field model. The resulting saliency measure is shown to achieve state-of-the-art accuracy for the prediction of fixations, at a very low computational cost. Since most modern cameras incorporate video encoders, this paves the way for in-camera saliency estimation, which could be useful in a variety of computer vision applications.

Our aim is to show how state-of-the-art computer vision techniques can be used to advance prehensile analysis (i.e., understanding the functionality of human hands). Prehensile analysis is a broad field of multi-disciplinary interest, where researchers painstakingly manually analyze hours of hand-object interaction videos to understand the mechanics of hand manipulation. In this work, we present promising empirical results indicating that wearable cameras and unsupervised clustering techniques can be used to automatically discover common modes of human hand use. In particular, we use a first-person point-of-view camera to record common manipulation tasks and leverage its strengths for reliably observing human hand use. To learn a diverse set of hand-object interactions, we propose a fast online clustering algorithm based on the Determinantal Point Process (DPP). Furthermore, we develop a hierarchical extension to the DPP clustering algorithm and show that it can be used to discover appearance-based grasp taxonomies. Using a purely data-driven approach, our proposed algorithm is able to obtain hand grasp taxonomies that roughly correspond to the classic Cutkosky grasp taxonomy. We validate our approach on over 10 hours of first-person point-of-view videos in both choreographed and real-life scenarios.

In this paper we are interested in exploiting geographic priors to help outdoor scene understanding. Towards this goal we propose a holistic approach that reasons jointly about 3D object detection, pose estimation, semantic segmentation as well as depth reconstruction from a single image. Our approach takes advantage of large-scale crowd-sourced maps to generate dense geographic, geometric and semantic priors by rendering the 3D world. We demonstrate the effectiveness of our holistic model on the challenging KITTI dataset, and show significant improvements over the baselines in all metrics and tasks.

We propose a novel dual-camera design to acquire 4D high-speed hyperspectral (HSHS) videos with high spatial and spectral resolution. Our work has two key technical contributions. First, we build a dual-camera system that simultaneously captures a panchromatic video at a high frame rate and a hyperspectral video at a low frame rate, which jointly provide reliable projections for the underlying HSHS video. Second, we exploit the panchromatic video to learn an over-complete 3D dictionary to represent each band-wise video sparsely, and a robust computational reconstruction is then employed to recover the HSHS video based on the joint videos and the self-learned dictionary. Experimental results demonstrate that, for the first time to our knowledge, the hyperspectral video frame rate reaches up to 100fps with decent quality, even when the incident light is not strong.