CVPR 2016 Papers

Many recent delineation techniques owe much of their increased effectiveness to path classification algorithms that make it possible to distinguish promising paths from others. The downside of this development is that they require annotated training data, which is tedious to produce. In this paper, we propose an Active Learning approach that considerably speeds up the annotation process. Unlike standard ones, it takes advantage of the specificities of the delineation problem. It operates on a graph and can reduce the training set size by up to 80% without compromising the reconstruction quality. We will show that our approach outperforms conventional ones on various biomedical and natural image datasets, thus showing that it is broadly applicable.

We propose a semi-automatic method to obtain foreground object masks for a large set of related images. We develop a stagewise active approach to propagation: in each stage, we actively determine the images that appear most valuable for human annotation, then revise the foreground estimates in all unlabeled images accordingly. In order to identify images that, once annotated, will propagate well to other examples, we introduce an active selection procedure that operates on the joint segmentation graph over all images. It prioritizes human intervention for those images that are uncertain and influential in the graph, while also mutually diverse. We apply our method to obtain foreground masks for over 1 million images. Our method yields state-of-the-art accuracy on the ImageNet and MIT Object Discovery datasets, and it focuses human attention more effectively than existing propagation strategies.

What defines an action like "kicking ball"? We argue that the true meaning of an action lies in the change or transformation an action brings to the environment. In this paper, we propose a novel representation for actions by modeling an action as a transformation which changes the state of the environment before the action happens (precondition) to the state after the action (effect). Motivated by recent advancements of video representation using deep learning, we design a Siamese network which models the action as a transformation on a high-level feature space. We show that our model gives improvements on standard action recognition datasets including UCF101 and HMDB51. More importantly, our approach is able to generalize beyond learned action categories and shows significant performance improvement on cross-category generalization on our new ACT dataset.

Actionness was introduced to quantify the likelihood of containing a generic action instance at a specific location. Accurate and efficient estimation of actionness is important in video analysis and may benefit other relevant tasks such as action recognition and action detection. This paper presents a new deep architecture for actionness estimation, called hybrid fully convolutional network (H-FCN), which is composed of appearance FCN (A-FCN) and motion FCN (M-FCN). These two FCNs leverage the strong capacity of deep models to estimate actionness maps from the perspectives of static appearance and dynamic motion, respectively. In addition, the fully convolutional nature of H-FCN allows it to efficiently process videos with arbitrary sizes. Experiments are conducted on the challenging datasets of Stanford40, UCF Sports, and JHMDB to verify the effectiveness of H-FCN on actionness estimation, which demonstrate that our method achieves superior performance to previous ones. Moreover, we apply the estimated actionness maps on action proposal generation and action detection. Our actionness maps advance the current state-of-the-art performance of these tasks substantially.

This work presents an approach to category-based action recognition in video using sparse coding techniques. The proposed approach includes two main contributions: i) A new method to handle intra-class variations by decomposing each video into a reduced set of representative atomic action acts or key-sequences, and ii) A new video descriptor, ITRA: Inter-Temporal Relational Act Descriptor, that exploits the power of comparative reasoning to capture relative similarity relations among key-sequences. In terms of the method to obtain key-sequences, we introduce a loss function that, for each video, leads to the identification of a sparse set of representative key-frames capturing both, relevant particularities arising in the input video, as well as relevant generalities arising in the complete class collection. In terms of the method to obtain the ITRA descriptor, we introduce a novel scheme to quantify relative intra and inter-class similarities among local temporal patterns arising in the videos. The resulting ITRA descriptor demonstrates to be highly effective to discriminate among action categories. As a result, the proposed approach reaches remarkable action recognition performance on several popular benchmark datasets, outperforming alternative state-of-the-art techniques by a large margin.

We present a highly accurate single image superresolution (SR) method. Our method uses a very deep convolutional network inspired by VGG-net used for ImageNet classification [19]. We find increasing our network depth shows a significant improvement in accuracy. Our final model uses 20 weight layers. By cascading small filters many times in a deep network structure, contextual information over large image regions is exploited in an efficient way. With very deep networks, however, convergence speed becomes a critical issue during training. We propose a simple yet effective training procedure. We learn residuals only and use extremely high learning rates (104 times higher than SRCNN [6]) enabled by adjustable gradient clipping. Our proposed method performs better than existing methods in accuracy and visual improvements in our results are easily noticeable.

This paper proposes a novel local descriptor through accumulated stability voting (ASV). The stability of feature dimensions is measured by their differences across scales. To be more robust to noise, the stability is further quantized by thresholding. The principle of maximum entropy is utilized for determining the best thresholds for maximizing discriminant power of the resultant descriptor. Accumulating stability renders a real-valued descriptor and it can be converted into a binary descriptor by an additional thresholding process. The real-valued descriptor attains high matching accuracy while the binary descriptor makes a good compromise between storage and accuracy. Our descriptors are simple yet effective, and easy to implement. In addition, our descriptors require no training. Experiments on popular benchmarks demonstrate the effectiveness of our descriptors and their superiority to the state-of-the-art descriptors.

Finding meaningful, structured representations of 3D point cloud data (PCD) has become a core task for spatial perception applications. In this paper we introduce a method for constructing compact generative representations of PCD at multiple levels of detail. As opposed to deterministic structures such as voxel grids or octrees, we propose probabilistic subdivisions of the data through local mixture modeling, and show how these subdivisions can provide a maximum likelihood segmentation of the data. The final representation is hierarchical, compact, parametric, and statistically derived, facilitating run-time occupancy calculations through stochastic sampling. Unlike traditional deterministic spatial subdivision methods, our technique enables dynamic creation of voxel grids according the application's best needs. In contrast to other generative models for PCD, we explicitly enforce sparsity among points and mixtures, a technique which we call expectation sparsification. This leads to a highly parallel hierarchical Expectation Maximization (EM) algorithm well-suited for the GPU and real-time execution. We explore the trade-offs between model fidelity and model size at various levels of detail, our tests showing favorable performance when compared to octree and NDT-based methods.

Deep learning using convolutional neural networks (CNNs) is quickly becoming the state-of-the-art for challenging computer vision applications. However, deep learning's power consumption and bandwidth requirements currently limit its application in embedded and mobile systems with tight energy budgets. In this paper, we explore the energy savings of optically computing the first layer of CNNs. To do so, we utilize bio-inspired Angle Sensitive Pixels (ASPs), custom CMOS diffractive image sensors which act similar to Gabor filter banks in the V1 layer of the human visual cortex. ASPs replace both image sensing and the first layer of a conventional CNN by directly performing optical edge filtering, saving sensing energy, data bandwidth, and CNN FLOPS to compute. Our experimental results (both on synthetic data and a hardware prototype) for a variety of vision tasks such as digit recognition, object recognition, and face identification demonstrate 97% reduction in image sensor power consumption and 90% reduction in data bandwidth from sensor to CPU, while achieving similar performance compared to traditional deep learning pipelines.

We propose a weighted variational model to estimate both the reflectance and the illumination from an observed image. We show that, though it is widely adopted for ease of modeling, the log-transformed image for this task is not ideal. Based on the previous investigation of the logarithmic transformation, a new weighted variational model is proposed for better prior representation, which is imposed in the regularization terms. Different from conventional variational models, the proposed model can preserve the estimated reflectance with more details. Moreover, the proposed model can suppress noise to some extent. An alternating minimization scheme is adopted to solve the proposed model. Experimental results demonstrate the effectiveness of the proposed model with its algorithm. Compared with other variational methods, the proposed method yields comparable or better results on both subjective and objective assessments.

We propose a system that finds text in natural scenes using a variety of cues. Our novel data-driven method incorporates coarse-to-fine detection of character pixels using convolutional features (Text-Conv), followed by extracting connected components (CCs) from characters using edge and color features, and finally performing a graph-based segmentation of CCs into words (Word-Graph). For Text-Conv, the initial detection is based on convolutional feature maps similar to those used in Convolutional Neural Networks (CNNs), but learned using Convolutional k-means. Convolution masks defined by local and neighboring patch features are used to improve detection accuracy. The Word-Graph algorithm uses contextual information to both improve word segmentation and prune false character/word detections. Different definitions for foreground (text) regions are used to train the detection stages, some based on bounding box intersection, and others on bounding box and pixel intersection. Our system obtains pixel, character, and word detection f-measures of 93.14%, 90.26%, and 86.77% respectively for the ICDAR 2015 Robust Reading Focused Scene Text dataset, out-performing state-of-the-art systems. This approach may work for other detection targets with homogenous color in natural scenes.

With the recent progress in visual recognition, we have already started to see a surge of vision related real-world applications. These applications, unlike general scene understanding, are task oriented and require specific information from visual data. Considering the current growth in new sensory devices, feature designs, feature learning methods, and algorithms, the search in the space of features and models becomes combinatorial. In this paper, we propose a novel cost-sensitive task-oriented recognition method that is based on a combination of linguistic semantics and visual cues. Our task-oriented framework is able to generalize to unseen tasks for which there is no training data and outperforms state-of-the-art cost-based recognition baselines on our new task-based dataset.

Multilinear models are widely used to represent the statistical variations of 3D human faces as they decouple shape changes due to identity and expression. Existing methods to learn a multilinear face model degrade if not every person is captured in every expression, if face scans are noisy or partially occluded, if expressions are erroneously labeled, or if the vertex correspondence is inaccurate. These limitations impose requirements on the training data that disqualify large amounts of available 3D face data from being usable to learn a multilinear model. To overcome this, we introduce the first framework to robustly learn a multilinear model from 3D face databases with missing data, corrupt data, wrong semantic correspondence, and inaccurate vertex correspondence. To achieve this robustness to erroneous training data, our framework jointly learns a multilinear model and fixes the data. We evaluate our framework on two publicly available 3D face databases, and show that our framework achieves a data completion accuracy that is comparable to state-of-the-art tensor completion methods. Our method reconstructs corrupt data more accurately than state-of-the-art methods, and improves the quality of the learned model significantly for erroneously labeled expressions.

In recent years, sensors capable of measuring both color and depth information have become increasingly popular. Despite the abundance of colored point set data, state-of-the-art probabilistic registration techniques ignore the available color information. In this paper, we propose a probabilistic point set registration framework that exploits available color information associated with the points. Our method is based on a model of the joint distribution of 3D-point observations and their color information. The proposed model captures discriminative color information, while being computationally efficient. We derive an EM algorithm for jointly estimating the model parameters and the relative transformations. Comprehensive experiments are performed on the Stanford Lounge dataset, captured by an RGB-D camera, and two point sets captured by a Lidar sensor. Our results demonstrate a significant gain in robustness and accuracy when incorporating color information. On the Stanford Lounge dataset, our approach achieves a relative reduction of the failure rate by 78% compared to the baseline. Furthermore, our proposed model outperforms standard strategies for combining color and 3D-point information, leading to state-of-the-art results.

Conventional representation based classifiers, ranging from the classical nearest neighbor classifier and nearest subspace classifier to the recently developed sparse representation based classifier (SRC) and collaborative representation based classifier (CRC), are essentially distance based classifiers. Though SRC and CRC have shown interesting classification results, their intrinsic classification mechanism remains unclear. In this paper we propose a probabilistic collaborative representation framework, where the probability that a test sample belongs to the collaborative subspace of all classes can be well defined and computed. Consequently, we present a probabilistic collaborative representation based classifier (ProCRC), which jointly maximizes the likelihood that a test sample belongs to each of the multiple classes. The final classification is performed by checking which class has the maximum likelihood. The proposed ProCRC has a clear probabilistic interpretation, and it shows superior performance to many popular classifiers, including SRC, CRC and SVM. Coupled with the CNN features, it also leads to state-of-the-art classification results on a variety of challenging visual datasets.

In many sub-fields, researchers collect datasets of human ground truth that are used to create a new algorithm. For example, in research on image perception, datasets have been collected for topics such as what makes an image aesthetic or memorable. Despite high costs for human data collection, datasets are infrequently reused beyond their own fields of interest. Moreover, the algorithms built from them are domain-specific (predict a small set of attributes) and usually unconnected to one another. In this paper, we present a paradigm for building generalized and expandable models of human image perception. First, we fuse multiple fragmented and partially-overlapping datasets through data imputation. We then create a theoretically-structured statistical model of human image perception that is fit to the fused datasets. The resulting model has many advantages. (1) It is generalized, going beyond the content of the constituent datasets, and can be easily expanded by fusing additional datasets. (2) It provides a new ontology usable as a network to expand human data in a cost-effective way. (3) It can guide the design of a generalized computational algorithm for multi-dimensional visual perception. Indeed, experimental results show that a model-based algorithm outperforms state-of-the-art methods on predicting visual sentiment, visual realism and interestingness. Our paradigm can be used in various visual tasks (e.g., video summarization).

Regression is an essential tool in Statistical analysis of data with many applications in Computer Vision, Machine Learning, Medical Imaging and various disciplines of Science and Engineering. Linear and nonlinear regression in a vector space setting has been well studied in literature. However, generalizations to manifold-valued data are only recently gaining popularity. With the exception of a few, most existing methods of regression for manifold valued data are limited to geodesic regression which is a generalization of the linear regression in vector-spaces. In this paper, we present a novel nonlinear kernel-based regression method that is applicable to manifold valued data. Our method is applicable to cases when the independent and dependent variables in the regression model are both manifold-valued or one is manifold-valued and the other is vector or scalar valued. Further, unlike most methods, our method does not require any imposed ordering on the manifold-valued data. The performance of our model is tested on a large number of real data sets acquired from Alzhiemers and movement disorder (Parkinsons and Essential Tremor) patients. We present an extensive set of results along with statistical validation and comparisons.

In this paper, we propose a new curvature penalized minimal path model for image segmentation via closed contour detection based on the weighted Euler elastica curves, firstly introduced to the field of computer vision in [22]. Our image segmentation method extracts a collection of curvature penalized minimal geodesics, concatenated to form a closed contour, by connecting a set of user-specified points. Globally optimal minimal paths can be computed by solving an Eikonal equation. This first order PDE is traditionally regarded as unable to penalize curvature, which is related to the path acceleration in active contour models. We introduce here a new approach that enables finding a global minimum of the geodesic energy including a curvature term. We achieve this through the use of a novel Finsler metric adding to the image domain the orientation as an extra space dimension. This metric is non-Riemannian and asymmetric, defined on an orientation lifted space, incorporating the curvature penalty in the geodesic energy. Experiments show that the proposed Finsler minimal path model indeed outperforms state-of-the-art minimal path models in both synthetic and real images.

We present a multi-stream bi-directional recurrent neural network for fine-grained action detection. Recently, two-stream convolutional neural networks (CNNs) trained on stacked optical flow and image frames have been successful for action recognition in videos. Our system uses a tracking algorithm to locate a bounding box around the person, which provides a frame of reference for appearance and motion and also suppresses background noise that is not within the bounding box. We train two additional streams on motion and appearance cropped to the tracked bounding box, along with full-frame streams. Our motion streams use pixel trajectories of a frame as raw features, in which the displacement values corresponding to a moving scene point are at the same spatial position across several frames. To model long-term temporal dynamics within and between actions, the multi-stream CNN is followed by a bi-directional Long Short-Term Memory (LSTM) layer. We show that our bi-directional LSTM network utilizes about 8 seconds of the video sequence to predict an action label. We test on two action detection datasets: the MPII Cooking 2 Dataset, and a new MERL Shopping Dataset that we introduce and make available to the community with this paper. The results demonstrate that our method significantly outperforms state-of-the-art action detection methods on both datasets.