CVPR 2015 Papers

This paper addresses the segmentation of videos with arbitrary motion, including dynamic textures, using novel motion features and a supervised learning approach. Dynamic textures are commonplace in natural scenes, and exhibit complex patterns of appearance and motion (e.g. water, smoke, swaying foliage). These are difficult for existing segmentation algorithms, often violate the brightness constancy assumption needed for optical flow, and have complex segment characteristics beyond uniform appearance or motion. Our solution uses custom spatiotemporal filters that capture texture and motion cues, along with a novel metric-learning framework that optimizes this representation for specific objects and scenes. This is used within a hierarchical, graph-based segmentation setting, yielding state-of-the-art results for dynamic texture segmentation. We also demonstrate the applicability of our approach to general object and motion segmentation, showing significant improvements over unsupervised segmentation and results comparable to the best task specific approaches.

Human actions capture a wide variety of interactions between people and objects. As a result, the set of possible actions is extremely large and it is difficult to obtain sufficient training examples for all actions. However, we could compensate for this sparsity in supervision by leveraging the rich semantic relationship between different actions. A single action is often composed of other smaller actions and is exclusive of certain others. We need a method which can reason about such relationships and extrapolate unobserved actions from known actions. Hence, we propose a novel neural network framework which jointly extracts the relationship between actions and uses them for training better action retrieval models. Our model incorporates linguistic, visual and logical consistency based cues to effectively identify these relationships. We train and test our model on a largescale image dataset of human actions. We show a significant improvement in mean AP compared to different baseline methods including the HEX-graph approach from Deng et al.

We consider the problem of designing a scene-specific pedestrian detector in a scenario where we have zero instances of real pedestrian data (i.e., no labeled real data or unsupervised real data). This scenario may arise when a new surveillance system is installed in a novel location and a scene-specific pedestrian detector must be trained prior to any observations of pedestrians. The key idea of our approach is to infer the potential appearance of pedestrians using geometric scene data and a customizable database of virtual simulations of pedestrian motion. We propose an efficient discriminative learning method that generates a spatially-varying pedestrian appearance model that takes into the account the perspective geometry of the scene. As a result, our method is able to learn a unique pedestrian classifier customized for every possible location in the scene. Our experimental results show that our proposed approach outperforms classical pedestrian detection models and hybrid synthetic-real models. Our results also yield a surprising result, that our method using purely synthetic data is able to outperform models trained on real scene-specific data when data is limited.

Multi-task learning is a natural approach for computer vision applications that require the simultaneous solution of several distinct but related problems, e.g. object detection, classification, tracking of multiple agents, or denoising, to name a few. The key idea is that exploring task relatedness (structure) can lead to better performances. In this paper, we propose and study a novel sparse, non-parametric approach exploiting the theory of Reproducing Kernel Hilbert Spaces for vector-valued functions. We develop a suitable regularization framework which can be formulated as a convex optimization problem, and is provably solvable using an alternating minimization approach. Empirical tests show that the proposed method compares favorably to state of the art techniques and further allows to recover interpretable structures, a problem of interest in its own right.

We develop a new approach to inferring lightness, the perceived reflectance of surfaces, from a single image. Classic methods view this problem from the perspective of intrinsic image decomposition, where an image is separated into reflectance and shading components. Rather than reason about reflectance and shading together, we learn to directly predict lightness differences between pixels. Large-scale training from human judgement data on relative reflectance, and patch representations built using deep networks, provide the foundation for our model. Benchmarked on the Intrinsic Images in the Wild dataset, our local lightness model achieves on-par performance with the state-of-the-art global lightness model, which incorporates multiple shading/reflectance priors and simultaneous reasoning between pairs of pixels in a dense conditional random field formulation.

We present a novel attribute learning framework named Hypergraph-based Attribute Predictor (HAP). In HAP, a hypergraph is leveraged to depict the attribute relations in the data. Then the attribute prediction problem is casted as a regularized hypergraph cut problem, in which a collection of attribute projections is jointly learnt from the feature space to a hypergraph embedding space aligned with the attributes. The learned projections directly act as attribute classifiers (linear and kernelized). This formulation leads to a very efficient approach. By considering our model as a multi-graph cut task, our framework can flexibly incorporate other available information, in particular class label. We apply our approach to attribute prediction, Zero-shot and N-shot learning tasks. The results on AWA, USAA and CUB databases demonstrate the value of our methods in comparison with the state-of-the-art approaches.

Exploiting label dependency for multi-label image classification can significantly improve classification performance. Probabilistic Graphical Models are one of the primary methods for representing such dependencies. The structure of graphical models, however, is either determined heuristically or learned from very limited information. Moreover, neither of these approaches scales well to large or complex graphs. We propose a principled way to learn the structure of a graphical model by considering input features and labels, together with loss functions. We formulate this problem into a max-margin framework initially, and then transform it into a convex programming problem. Finally, we propose a highly scalable procedure that activates a set of cliques iteratively. Our approach exhibits both strong theoretical properties and a significant performance improvement over state-of-the-art methods on both synthetic and real-world data sets.

Large-scale supervised datasets are crucial to train convolutional neural networks (CNNs) for various computer vision problems. However, obtaining a massive amount of well-labeled data is usually very expensive and time consuming. In this paper, we introduce a general framework to train CNNs with only a limited number of clean labels and millions of easily obtained noisy labels. We model the relationships between images, class labels and label noises with a probabilistic graphical model and further integrate it into an end-to-end deep learning system. To demonstrate the effectiveness of our approach, we collect a large-scale real-world clothing classification dataset with both noisy and clean labels. Experiments on this dataset indicate that our approach can better correct the noisy labels and improves the performance of trained CNNs.

Detecting poorly textured objects and estimating their 3D pose reliably is still a very challenging problem. We introduce a simple but powerful approach to computing descriptors for object views that efficiently capture both the object identity and 3D pose. By contrast with previous manifold-based approaches, we can rely on the Euclidean distance to evaluate the similarity between descriptors, and therefore use scalable Nearest Neighbor search methods to efficiently handle a large number of objects under a large range of poses. To achieve this, we train a Convolutional Neural Network to compute these descriptors by enforcing simple similarity and dissimilarity constraints between the descriptors. We show that our constraints nicely untangle the images from different objects and different views into clusters that are not only well-separated but also structured as the corresponding sets of poses: The Euclidean distance between descriptors is large when the descriptors are from different objects, and directly related to the distance between the poses when the descriptors are from the same object. These important properties allow us to outperform state-of-the-art object views representations on challenging RGB and RGB-D data.

The recent availability of geo-tagged images and rich geospatial data has inspired a number of algorithms for image based geolocalization. Most approaches predict the location of a query image by matching to ground-level images with known locations (e.g., street-view data). However, most of the Earth does not have ground-level reference photos available. Fortunately, more complete coverage is provided by oblique aerial or "bird's eye" imagery. In this work, we localize a ground-level query image by matching it to a reference database of aerial imagery. We use publicly available data to build a dataset of 78K aligned cross- view image pairs. The primary challenge for this task is that traditional computer vision approaches cannot handle the wide baseline and appearance variation of these cross- view pairs. We use our dataset to learn a feature representation in which matching views are near one another and mismatched views are far apart. Our proposed approach, Where-CNN, is inspired by deep learning success in face verification and achieves significant improvements over traditional hand-crafted features and existing deep features learned from other large-scale databases. We show the effectiveness of Where-CNN in finding matches between street view and aerial view imagery and demonstrate the ability of our learned features to generalize to novel locations.

Part model-based methods have been successfully applied to object detection and scene classification and have achieved state-of-the-art results. More recently the "sparselets" work [1-3] were introduced to serve as a universal set of shared basis learned from a large number of part detectors, resulting in notable speedup. Inspired by this framework, in this paper, we propose a novel scheme to train more effective sparselets with a coarse-to-fine framework. Specifically, we first train coarse sparselets to exploit the redundancy existing among part detectors by using an unsupervised single-hidden layer auto-encoder. Then, we simultaneously train fine sparselets and activation vectors using a supervised single-hidden-layer neural network, in which sparselets training and discriminative activation vectors learning are jointly embedded into a unified framework. In order to adequately explore the discriminative information hidden in the part detectors and to achieve sparsity, we propose to optimize a new discriminative objective function by imposing L0-norm sparsity constraint on the activation vectors. By using the proposed framework, promising results for multi-class object detection and scene classification are achieved on PASCAL VOC 2007, MIT Scene-67, and UC Merced Land Use datasets, compared with the existing sparselets baseline methods.

As consumer depth sensors become widely available, estimating scene flow from RGBD sequences has received increasing attention. Although the depth information allows the recovery of 3D motion from a single view, it poses new challenges. In particular, depth boundaries are not well-aligned with RGB image edges and therefore not reliable cues to localize 2D motion boundaries. In addition, methods that extend the 2D optical flow formulation to 3D still produce large errors in occlusion regions. To better use depth for occlusion reasoning, we propose a layered RGBD scene flow method that jointly solves for the scene segmentation and the motion. Our key observation is that the noisy depth is sufficient to decide the depth ordering of layers, thereby avoiding a computational bottleneck for RGB layered methods. Furthermore, the depth enables us to estimate a per-layer 3D rigid motion to constrain the motion of each layer. Experimental results on both the Middlebury and real-world sequences demonstrate the effectiveness of the layered approach for RGBD scene flow estimation.

This paper is about estimating intensity levels of Facial Action Units (FAUs) in videos as an important and challenging step toward interpreting facial expressions. To address uncertainty in detections of facial landmark points, used as out input features, we formulate a new generative framework comprised of a graphical model, inference, and algorithms for learning both model parameters and structure. Our model is a latent tree (LT) that represents input features of facial landmark points and FAU intensities as leaf nodes, and encodes their higher-order dependencies with latent nodes at tree levels closer to the root. No other restrictions are placed on the model structure beyond that it is a tree. We specify a new algorithm for efficient learning of model structure that iteratively builds LT by adding either new edge or new hidden node to LT, whichever of these two graph-edit operations gives the highest increase of the joint likelihood. Our structure learning efficiently computes likelihood increase and selects an optimal graph revision without considering all possible structural changes. For FAU intensity estimation, we derive closed-form expressions of posterior marginals of all variables in LT, and specify an efficient inference of in two passes -- bottom-up and top-down. Our evaluation on the benchmark DISFA and ShoulderPain datasets, in subject-independent setting, demonstrate that we outperform the state of the art, even in the presence of significant noise in locations of facial landmark points. We demonstrate our correct learning of model structure by probabilistically sampling facial landmark points, conditioned on a given FAU intensity, and thus generating plausible facial expressions.

Depth cameras have helped commoditize 3D digitization of the real-world. It is now feasible to use a single Kinect-like camera to scan in an entire building or other large-scale scenes. At large scales, however, there is an inherent challenge of dealing with distortions and drift due to accumulated pose estimation errors. Existing techniques suffer from one or more of the following: a) requiring an expensive offline global optimization step taking hours to compute; b) needing a full second pass over the input depth frames to correct for accumulated errors; c) relying on RGB data alongside depth data to optimize poses; or d) requiring the user to create explicit loop closures to allow gross alignment errors to be resolved. In this paper, we present a method that addresses all of these issues. Our method supports online model correction, without needing to reprocess or store any input depth data. Even while performing global correction of a large 3D model, our method takes only minutes rather than hours to compute. Our model does not require any explicit loop closures to be detected and, finally, relies on depth data alone, allowing operation in low-lighting conditions. We show qualitative results on many large scale scenes, highlighting the lack of error and drift in our reconstructions. We compare to state of the art techniques and demonstrate large-scale dense surface reconstruction "in the dark", a capability not offered by RGB-D techniques.

Satellite imagery is a valuable source of information for assessing damages in distressed areas undergoing a calamity, such as an earthquake or an armed conflict. However, the sheer amount of data required to be inspected for this assessment makes it impractical to do it manually. To address this problem, we present a semi-supervised learning framework for large-scale damage detection in satellite imagery. We present a comparative evaluation of our framework using over 88 million images collected from 4,665 square kilometers from 12 different locations around the world. To enable accurate and efficient damage detection, we introduce a novel use of hierarchical shape features in the bags-of-visual words setting. We analyze how practical factors such as sun, sensor-resolution, and satellite-angle differences impact the effectiveness of our proposed representation, and compare it to five alternative features in multiple learning settings. Finally, we demonstrate through a user-study that our semi-supervised framework results in a ten-fold reduction in human annotation time at a minimal loss in detection accuracy compared to an exhaustive manual inspection.

Domain adaptation (DA) has gained a lot of success in the recent years in computer vision to deal with situations where the learning process has to transfer knowledge from a source to a target domain. In this paper, we introduce a novel unsupervised DA approach based on both subspace alignment and selection of landmarks similarly distributed between the two domains. Those landmarks are selected so as to reduce the discrepancy between the domains and then are used to non linearly project the data in the same space where an efficient subspace alignment (in closed-form) is performed. We carry out a large experimental comparison in visual domain adaptation showing that our new method outperforms the most recent unsupervised DA approaches.

Recently, an increasing number of works have proposed to learn visual saliency by leveraging human fixations. However, the collection of human fixations is time consuming and the existing eye tracking datasets are generally small when compared with other domains. Thus, it contains a certain degree of dataset bias due to the large image variations (e.g., outdoor scenes vs. emotion-evoking images). In the learning based saliency prediction literature, most models are trained and evaluated within the same dataset and cross dataset validation is not yet a common practice. Instead of directly applying model learned from another dataset in cross dataset fashion, it is better to transfer the prior knowledge obtained from one dataset to improve the training and prediction on another. In addition, since new datasets are built and shared in the community from time to time, it would be good not to retrain the entire model when new data are added. To address these problems, we proposed a new learning based saliency model, namely Label Consistent Quadratic Surrogate algorithm, which employs an iterative online algorithm to learn a sparse dictionary with label consistent constraint. The advantages of the proposed model are three-folds: (1) the quadratic surrogate function guarantees convergence at each iteration, (2) the label consistent constraint enforces the predicted sparse code to be discriminative, and (3) the online properties enable the proposed algorithm to adapt existing model with new data without retraining. As shown in this work, the proposed saliency model achieves better performance than the state-of-the-art saliency models.

This paper investigates the problem of registering a scanned scene, represented by 3D Euclidean point coordinates, and two or more uncalibrated cameras. An unknown subset of the scanned points have their image projections detected and matched across images. The proposed approach assumes the cameras only known in some arbitrary projective frame and no calibration or autocalibration is required. The devised solution is based on a Linear Matrix Inequality (LMI) framework that allows simultaneously estimating the projective transformation relating the cameras to the scene and establishing 2D-3D correspondences without triangulating image points. The proposed LMI framework allows both deriving triangulation-free LMI cheirality conditions and establishing putative correspondences between 3D volumes (boxes) and 2D pixel coordinates. Two registration algorithms, one exploiting the scene's structure and the other concerned with robustness, are presented. Both algorithms employ the Branch-and-Prune paradigm and guarantee convergence to a global solution under mild initial bound conditions. The results of our experiments are presented and compared against other approaches.

Total Variation (TV) is an effective and popular prior model in the field of regularization- image processing. This paper focuses on TV for image restoration in the presence of impulse noise. This type of noise frequently arises in data acquisition and transmission due to many reasons, e.g. a faulty sensor or analog-to-digital converter errors. Removing this noise is an important task in image restoration. State-of-the-art methods such as Adaptive Outlier Pursuit(AOP) [42], which is based on TV with L02-norm data fidelity, only give sub-optimal performance. In this paper, we propose a new method, called L0TV-PADMM, which solves the TV-based restoration problem with L0-norm data fidelity. To effectively deal with the resulting non-convex non-smooth optimization problem, we first reformulate it as an equivalent MPEC (Mathematical Program with Equilibrium Constraints), and then solve it using a proximal Alternating Direction Method of Multipliers (PADMM). Our L0TV-PADMM method finds a desirable solution to the original L0-norm optimization problem and is proven to be convergent under mild conditions. We apply L0TV-PADMM to the problems of image denoising and deblurring in the presence of impulse noise. Our extensive experiments demonstrate that L0TV-PADMM outperforms state-of-the-art image restoration methods.