CVPR 2015 Papers

This paper presents effective combination models with certain combination features for human detection. In the past several years, many existing features/models have achieved impressive progress, but their performances are still limited by the biases rooted in their self-structures, that is, a particular kind of feature/model may work well for some types of human bodies, but not for all the types. To tackle this difficult problem, we combine certain complementary features/models together with effective organization/fusion methods. Specifically, the HOG features, color features and bar-shape features are combined together with a cell-based histogram structure to form the so-called HOG-III features. Moreover, the detections from different models are fused together with the new proposed weighted-NMS algorithm, which enhances the probable "true" activations as well as suppresses the overlapped detections. The experiments on PASCAL VOC datasets demonstrate that, both the HOG-III features and the weighted-NMS fusion algorithm are effective (obvious improvement for detection performance) and efficient (relatively less computation cost): When applied to human detection task with the Grammar model and Poselet model, they can boost the detection performance significantly; Also, when extended to detection of the whole VOC 20 object categories with the deformable part-based model and deepCNN-based model, they still show competitive improvements.

Image feature representation plays an essential role in image recognition and related tasks. The current state-of-the-art feature learning paradigm is supervised learning from labeled data. However, this paradigm requires large-scale category labels, which limits its applicability to domains where labels are hard to obtain. In this paper, we propose a new data-driven feature learning paradigm which does not rely on category labels. Instead, we learn from user behavior data collected on social media. Concretely, we use the image relationship discovered in the latent space from the user behavior data to guide the image feature learning. We collect a large-scale image and user behavior dataset from Behance.net. The dataset consists of 1.9 million images and over 300 million view records from 1.9 million users. We validate our feature learning paradigm on this dataset and find that the learned feature significantly outperforms the state-of-the-art image features in learning better image similarities. We also show that the learned feature performs competitively on various recognition benchmarks.

We present a new approach to wide baseline matching. We propose to use a hierarchical decomposition of the image domain and coarse-to-fine selection of regions to match. In contrast to interest point matching methods, which sample salient regions to reduce the cost of comparing all regions in two images, our method eliminates regions systematically to achieve efficiency. One advantage of our approach is that it is not restricted to covariant salient regions, which is too restrictive under large viewpoint and leads to few corresponding regions. Affine invariant matching of regions in the hierarchy is achieved efficiently by a coarse-to-fine search of the affine space. Experiments on two benchmark datasets shows that our method finds more correct correspondence of the image (with fewer false alarms) than other wide baseline methods on large viewpoint change.

With the goal of effectively identifying common and salient objects in a group of relevant images, co-saliency detection has become essential for many applications such as video foreground extraction, surveillance, image retrieval, and image annotation. In this paper, we propose a unified co-saliency detection framework by introducing two novel insights: 1) looking deep to transfer higher-level representations by using the convolutional neural network with additional adaptive layers could better reflect the properties of the co-salient objects, especially their consistency among the image group; 2) looking wide to take advantage of the visually similar neighbors beyond a certain image group could effectively suppress the influence of the common background regions when formulating the intra-group consistency. In the proposed framework, the wide and deep information are explored for the object proposal windows extracted in each image, and the co-saliency scores are calculated by integrating the intra-image contrast and intra group consistency via a principled Bayesian formulation. Finally the window-level co-saliency scores are converted to the superpixel-level co-saliency maps through a foreground region agreement strategy. Comprehensive experiments on two benchmark datasets have demonstrated the consistent performance gain of the proposed approach.

We propose a novel method for establishing correspondences on deformable objects for single-target object tracking. The key ingredient is a dissimilarity measure between correspondences that takes into account their geometric compatibility, allowing us to separate inlier correspondences from outliers. We employ both static correspondences from the initial appearance of the object as well as adaptive correspondences from the previous frame to address the stability-plasticity dilemma. The geometric dissimilarity measure enables us to also disambiguate keypoints that are difficult to match. Based on these ideas we build a keypoint-based tracker that outputs rotated bounding boxes. We demonstrate in a rigorous empirical analysis that this tracker outperforms the state of the art on a dataset of 77 sequences.

This paper originally proposes the clique-graph and further presents a clique-graph matching method by preserving global and local structures. Especially, we formulate the objective function of clique-graph matching with respective to two latent variables, the clique information in the original graph and the pairwise clique correspondence constrained by the one-to-one matching. Since the objective function is not jointly convex to both latent variables, we decompose it into two consecutive steps for optimization: 1) clique-to-clique similarity measure by preserving local unary and pairwise correspondences; 2) graph-to-graph similarity measure by preserving global clique-to-clique correspondence. Extensive experiments on the synthetic data and real images show that the proposed method can outperform representative methods especially when both noise and outliers exist.

Traditional data-driven classifier learning approaches become limited when the training data is inadequate either in quantity or quality. To address this issue, in this paper we propose to combine hidden information and data to enhance classifier learning. Hidden information represents information that is only available during training but not available during testing. It often exists in many applications yet has not been thoroughly exploited, and existing methods to utilize hidden information are still limited. To this end, we propose two general approaches to exploit different types of hidden information to improve different classifiers. We also extend the proposed methods to deal with incomplete hidden information. Experimental results on different applications demonstrate the effectiveness of the proposed methods for exploiting hidden information and their superior performance to existing methods.

Video segmentation has become an important and active research area with a large diversity of proposed approaches. Graph-based methods, enabling topperformance on recent benchmarks, consist of three essential components: 1. powerful features account for object appearance and motion similarities; 2. spatio-temporal neighborhoods of pixels or superpixels (the graph edges) are modeled using a combination of those features; 3. video segmentation is formulated as a graph partitioning problem. While a wide variety of features have been explored and various graph partition algorithms have been proposed, there is surprisingly little research on how to construct a graph to obtain the best video segmentation performance. This is the focus of our paper. We propose to combine features by means of a classifier, use calibrated classifier outputs as edge weights and define the graph topology by edge selection. By learning the graph (without changes to the graph partitioning method), we improve the results of the best performing video segmentation algorithm by 6% on the challenging VSB100 benchmark, while reducing its runtime by 55%, as the learnt graph is much sparser.

Classifiers for object categorization are usually evaluated by their accuracy on a set of i.i.d. test examples. This provides us with an estimate of the expected error when applying the classifiers to a single new image. In real application, however, classifiers are rarely only used for a single image and then discarded. Instead, they are applied sequentially to many images, and these are typically not i.i.d. samples from a fixed data distribution, but they carry dependencies and their class distribution varies over time. In this work, we argue that the phenomenon of correlated data at prediction time is not a nuisance, but a blessing in disguise. We describe a probabilistic method for adapting classifiers at prediction time without having to retrain them. We also introduce a framework for creating realistically distributed image sequences, which offers a way to benchmark classifier adaptation methods, such as the one we propose. Experiments on the ILSVRC2010 and ILSVRC2012 datasets show that adapting object classification systems at prediction time can significantly reduce their error rate, even with no additional human feedback.

Many existing recognition algorithms combine different modalities based on training accuracy but do not consider the possibility of noise at test time. We describe an algorithm that perturbs test features so that all modalities predict the same class. We enforce this perturbation to be as small as possible via a quadratic program (QP) for continuous features, and a mixed integer program (MIP) for binary features. To efficiently solve the MIP, we provide a greedy algorithm and empirically show that its solution is very close to that of a state-of-the-art MIP solver. We evaluate our algorithm on several datasets and show that the method outperforms existing approaches.

Occlusion relations inform the partition of the image domain into ``objects'' but are difficult to determine from a single image or short-baseline video. We show how long-term occlusion relations can be robustly inferred from video, and used within a convex optimization framework to segment the image domain into regions. We highlight the challenges in determining these occluder/occluded relations and ensuring regions remain temporally consistent, propose strategies to overcome them, and introduce an efficient numerical scheme to perform the partition directly on the pixel grid, without the need for superpixelization or other preprocessing steps.

Object reconstruction from a single image -- in the wild -- is a problem where we can make progress and get meaningful results today. This is the main message of this paper, which introduces an automated pipeline with pixels as inputs and 3D surfaces of various rigid categories as outputs in images of realistic scenes. At the core of our approach are deformable 3D models that can be learned from 2D annotations available in existing object detection datasets, that can be driven by noisy automatic object segmentations and which we complement with a bottom-up module for recovering high-frequency shape details. We perform a comprehensive quantitative analysis and ablation study of our approach using the recently introduced PASCAL 3D+ dataset and show very encouraging automatic reconstructions on PASCAL VOC.

This paper presents a method for stitching stereoscopic panoramas from stereo images casually taken using a stereo camera. This method addresses three challenges of stereoscopic image stitching: how to handle parallax, how to stitch the left- and right-view panorama consistently, and how to take care of disparity during stitching. This method addresses these challenges using a three-step approach. First, we employ a state-of-the-art stitching algorithm that handles parallax well to stitch the left views of input stereo images and create the left view of the final stereoscopic panorama. Second, we stitch the input disparity maps to obtain the target disparity map for the stereoscopic panorama by solving a Poisson's equation. This target disparity map is optimized such that there are no vertical disparities and the original perceived depth distribution is preserved. Finally, we warp the right views of the input stereo images and stitch them into the right view of the final stereoscopic panorama according to the target disparity map. The stitching of the right views is formulated as a labeling problem that is constrained by the stitching of the left views to make the left- and right-view panorama consistent to avoid retinal rivalry. Our experiments show that our method can effectively stitch casually taken stereo images and produce high-quality stereoscopic panoramas that deliver a pleasant stereoscopic 3D viewing experience.

We extends the previous 2D cascaded object pose regression work [9] in two aspects so that it works better for 3D articulated objects. Our first contribution is 3D pose-indexed features that generalize the previous 2D parameterized features and achieve better invariance to 3D transformations. Our second contribution is a principled hierarchical regression that is adapted to the articulated object structure. It is therefore more accurate and faster. Comprehensive experiments verify the state-of-the-art accuracy and efficiency of the proposed approach on the challenging 3D hand pose estimation problem, on a public dataset and our new dataset.

Can humans fly? Emphatically no. Can cars eat? Again, absolutely not. Yet, these absurd inferences result from the current disregard for particular types of actors in action understanding. There is no work we know of on simultaneously inferring actors and actions in the video, not to mention a dataset to experiment with. Our paper hence marks the first effort in the computer vision community to jointly consider various types of actors undergoing various actions. To start with the problem, we collect a dataset of 3782 videos from YouTube and label both pixel-level actors and actions in each video. We formulate the general actor-action understanding problem and instantiate it at various granularities: both video-level single- and multiple-label actor-action recognition and pixel-level actor-action semantic segmentation. Our experiments demonstrate that inference jointly over actors and actions outperforms inference independently over them, and hence concludes our argument of the value of explicit consideration of various actors in comprehensive action understanding.

This paper presents a reliable non-blind method to measure intrinsic lens blur. We first introduce an accurate camera-scene alignment framework that avoids erroneous homography estimation and camera tone curve estimation. This alignment is used to generate a sharp correspondence of a target pattern captured by the camera. Second, we introduce a Point Spread Function (PSF) estimation approach where information about the frequency spectrum of the target image is taken into account. As a result of these steps and the ability to use multiple target images in this framework, we achieve a PSF estimation method robust against noise and suitable for mobile devices. Experimental results show that the proposed method results in PSFs with more than 10 dB higher accuracy in noisy conditions compared with the PSFs generated using state-of-the-art techniques.

Automatically describing an image with a sentence is a long-standing challenge in computer vision and natural language processing. Due to recent progress in object detection, attribute classification, action recognition, etc., there is renewed interest in this area. However, evaluating the quality of descriptions has proven to be challenging. We propose a novel paradigm for evaluating image descriptions that uses human consensus. This paradigm consists of three main parts: a new triplet-based method of collecting human annotations to measure consensus, a new automated metric that captures consensus, and two new datasets: PASCAL-50S and ABSTRACT-50S that contain 50 sentences describing each image. Our simple metric captures human judgment of consensus better than existing metrics across sentences generated by various sources. We also evaluate five state-of-the-art image description approaches using this new protocol and provide a benchmark for future comparisons. A version of CIDEr named CIDEr-D is available as a part of MS COCO evaluation server to enable systematic evaluation and benchmarking.

Numerous recent approaches attempt to remove image blur due to camera shake, either with one or multiple input images, by explicitly solving an inverse and inherently ill-posed deconvolution problem. If the photographer takes a burst of images, a modality available in virtually all modern digital cameras, we show that it is possible to combine them to get a clean sharp version. This is done without explicitly solving any blur estimation and subsequent inverse problem. The proposed algorithm is strikingly simple: it performs a weighted average in the Fourier domain, with weights depending on the Fourier spectrum magnitude. The method's rationale is that camera shake has a random nature and therefore each image in the burst is generally blurred differently. Experiments with real camera data show that the proposed Fourier Burst Accumulation algorithm achieves state-of-the-art results an order of magnitude faster, with simplicity for on-board implementation on camera phones.

We introduce tools and methodologies to collect high quality, large scale fine-grained computer vision datasets using citizen scientists -- crowd annotators who are passionate and knowledgeable about specific domains such as birds or airplanes. We worked with citizen scientists and domain experts to collect NABirds, a new high quality dataset containing 48,562 images of North American birds with 555 categories, part annotations and bounding boxes. We find that citizen scientists are significantly more accurate than Mechanical Turkers at zero cost. We worked with bird experts to measure the quality of popular datasets like CUB-200-2011 and ImageNet and found class label error rates of at least 4%. Nevertheless, we found that learning algorithms are surprisingly robust to annotation errors and this level of training data corruption can lead to an acceptably small increase in test error if the training set has sufficient size. At the same time, we found that an expert-curated high quality test set like NABirds is necessary to accurately measure the performance of fine-grained computer vision systems. We used NABirds to train a publicly available bird recognition service deployed on the web site of the Cornell Lab of Ornithology.