CVPR 2015 Papers

Dense 3D reconstruction still remains a hard task for a broad number of object classes which are not sufficiently textured or contain transparent and reflective parts. Shape priors are the tool of choice when the input data itself is not descriptive enough to get a faithful reconstruction. We propose a novel shape prior formulation that splits the object into multiple convex parts. The reconstruction problem is posed as a volumetric multi-label segmentation. Each of the transitions between labels is penalized with its individual anisotropic smoothness term. This powerful formulation allows us to represent a descriptive shape prior. For the object classes used in this paper the individual segments naturally correspond to different semantic parts of the object. This leads to a semantic segmentation as a side product of our shape prior formulation. We evaluate our method on several challenging real-world datasets. Our results show that we can resolve issues such as undesired holes and disconnected parts. Taking into account a segmentation of the free space, we show that we are able to reconstruct concavities, such as the interior of a mug.

We focus on the problem of semantic segmentation based on RGB-D data, with emphasis on analyzing cluttered indoor scenes containing many visual categories and instances. Our approach is based on a parametric figure-ground intensity and depth-constrained proposal process that generates spatial layout hypotheses at multiple locations and scales in the image followed by a sequential inference algorithm that produces a complete scene estimate. Our contributions can be summarized as follows: (1) a generalization of parametric max flow figure-ground proposal methodology to take advantage of intensity and depth information, in order to systematically and efficiently generate the breakpoints of an underlying spatial model in polynomial time, (2) new region description methods based on second-order pooling over multiple features constructed using both intensity and depth channels, (3) a principled search-based structured prediction inference and learning process that resolves conflicts in overlapping spatial partitions and selects regions sequentially towards complete scene estimates, and (4) extensive evaluation of the impact of depth, as well as the effectiveness of a large number of descriptors, both pre-designed and automatically obtained using deep learning, in a difficult RGB-D semantic segmentation problem with 92 classes. We report state of the art results in the challenging NYU Depth Dataset V2, extended for the RMRC 2013 and RMRC 2014 Indoor Segmentation Challenges, where currently the proposed model ranks first. Moreover, we show that by combining second-order and deep learning features, over 15% relative accuracy improvements can be additionally achieved. In a scene classification benchmark, our methodology further improves the state of the art by 24%.

This paper addresses the problem of pixel-level segmentation and classification of scene images with an entirely learning-based approach using Long Short Term Memory (LSTM) recurrent neural networks, which are commonly used for sequence classification. We investigate two-dimensional (2D) LSTM networks for natural scene images taking into account the complex spatial dependencies of labels. Prior methods generally have required separate classification and image segmentation stages and/or pre- and post-processing. In our approach, classification, segmentation, and context integration are all carried out by 2D LSTM networks, allowing texture and spatial model parameters to be learned within a single model. The networks efficiently capture local and global contextual information over raw RGB values and adapt well for complex scene images. Our approach, which has a much lower computational complexity than prior methods, achieved state-of-the-art performance over the Stanford Background and the SIFT Flow datasets. In fact, if no pre- or post-processing is applied, LSTM networks outperform other state-of-the-art approaches. Hence, only with a single-core Central Processing Unit (CPU), the running time of our approach is equivalent or better than the compared state-of-the-art approaches which use a Graphics Processing Unit (GPU). Finally, our networks' ability to visualize feature maps from each layer supports the hypothesis that LSTM networks are overall suited for image processing tasks.

With the help of a convolutional neural network~(CNN) trained to recognize objects, a scene image is represented as a bag of semantics (BoS). This involves classifying image patches using the network and considering the class posterior probability vectors as locally extracted semantic descriptors. The image BoS is summarized using a Fisher vector~(FV) embedding that exploits the properties of the space of these descriptors. The resulting representation is referred to as a semantic Fisher vector. Two implementations of a semantic FV are investigated. First involves modeling the BoS with a Dirichlet Mixture and computing the Fisher gradients for this model. Due to the difficulty of mixture modeling on a non-Euclidean probability simplex, this approach is shown to be unsuccessful. A second implementation is derived using the interpretation of semantic descriptors as parameters of a multinomial distribution. Like the parameters of any exponential family, these can be projected into their natural parameter space. For a CNN, this is shown equivalent to using inputs of its soft-max layer as patch descriptors. A semantic FV is then computed as a Gaussian Mixture FV in the space of these natural parameters. This representation is shown to outperform other alternatives such as FVs of features from the intermediate CNN layers or a classifier obtained by adapting (fine-tuning) the CNN. The proposed FV represents an embedding for object classification probabilities. As an image representation, therefore, it is complementary to the features obtained from a scene classification CNN. A combination of the two representations is shown to achieve state-of-the-art results on MIT Indoor scenes and SUN datasets.

Videos consisting of thousands of high resolution frames are challenging for existing structure from motion (SfM) and simultaneous-localization and mapping (SLAM) techniques. We present a new approach for simultaneously computing extrinsic camera poses and 3D scene structure that is capable of handling such large volumes of image data. The key insight behind this paper is to effectively exploit coherence in densely sampled video input. Our technical contributions include robust tracking and selection of confident video frames, a novel window bundle adjustment, frame-to-structure verification for globally consistent reconstructions with multi-loop closing, and utilizing efficient global linear camera pose estimation in order to link both consecutive and distant bundle adjustment windows. To our knowledge we describe the first system that is capable of handling high resolution, high frame-rate video data with close to realtime performance. In addition, our approach can robustly integrate data from different video sequences, allowing multiple video streams to be simultaneously calibrated in an efficient and globally optimal way. We demonstrate high quality alignment on large scale challenging datasets, e.g., 2-20 megapixel resolution at frame rates of 25-120 Hz with thousands of frames.

For practical applications, an object detection system requires huge number of classes to meet real world needs. Many successful object detection systems use part-based model which trains several filters (classifiers) for each class to perform multiclass object detection. However, these methods have linear computational complexity in regard to the number of classes and may lead to huge computing time. To solve the problem, some works learn a codebook for the filters and conduct operations only on the codebook to make computational complexity sublinear in regard to the number of classes. But the past studies missed to consider filter characteristics, e.g., filters are weights trained by Support Vector Machine, and rather they applied method such as sparse coding for visual signals' optimization. This misuse results in huge accuracy loss when a large speedup is required. To remedy this shortcoming, we have developed a new method called Regularized Sparse Coding which is designed to reconstruct filter functionality. That is, it reconstructs the ability of filter to produce accurate score for classification. Our method can reconstruct filters by minimizing score map error, while sparse coding reconstructs filters by minimizing appearance error. This different optimization strategy makes our method be able to have small accuracy loss when a large speedup is achieved. On the ILSVRC 2013 dataset, which has 200 classes, this work represents a 16 times speedup using only 1.25% memory on single CPU with 0.04 mAP drop when compared with the original Deformable Part Model. Moreover, parallel computing on GPUs is also applicable for our work to achieve more speedup.

Specular reflection generally decreases the saturation of surface colors, which will be possibly confused with other colors that have the same hue but lower saturation. Traditional methods for specular reflection separation suffer this problem of hue-saturation ambiguity, producing over-saturated specular-free images quite often. We proposed a two-step approach to solve this problem. In the first step, we produce an over-saturated specular-free image by global chromaticity propagation from specular-free pixels to highlighted ones. Then we recover the saturation based on priors of the piecewise constancy of diffuse chromaticity as well as the spatial sparsity and smoothness of specular reflection. We achieve this through increasing the achromatic component of diffuse chromaticity, while the magnitudes of increments are determined by linear programming under the constraints derived from the priors. Experiments on both laboratory and natural images show that our method can separate the specular reflection while preserving the saturation of the underlying surface colors.

People can immediately and precisely identify 1, 2, 3 or 4 items by a simple glance. The phenomenon, known as Subitizing, inspires us to pursue the task of Salient Object Subitizing (SOS), i.e. predicting the existence and the number of salient objects in a scene using holistic cues. To study this problem, we propose a new image dataset annotated by Amazon Mechanical Turk. We show that for a substantial proportion of our dataset, there is a high labeling consistency among different subjects, even when a very limited viewing time (0.5s) is given. On our dataset, the baseline method using the global Convolutional Neural Network (CNN) feature achieves 94% recall rate in detecting the existence of salient objects, and 42-82% recall rate (chance is 20%) in predicting the number of salient objects (1, 2, 3, and 4+), without resorting to any object localization process. Finally, we demonstrate the usefulness of the proposed subitizing technique in two computer vision applications: salient object detection and object proposal.

We propose a bootstrap learning algorithm for salient object detection in which both weak and strong models are exploited. First, a weak saliency map is constructed based on image priors to generate training samples for a strong model. Second, a strong classifier based on samples directly from an input image is learned to detect salient pixels. Results from multiscale saliency maps are integrated to further improve the detection performance. Extensive experiments on five benchmark datasets demonstrate that the proposed bootstrap learning algorithm performs favorably against the state-of-the-art saliency detection methods. Furthermore, we show that the proposed bootstrap learning approach can be easily applied to other bottom-up saliency models for significant improvement.

We introduce an unsupervised, geodesic distance based, salient video object segmentation method. Unlike traditional methods, our method incorporates saliency as prior for object via the computation of robust geodesic measurement. We consider two discriminative visual features: spatial edges and temporal motion boundaries as indicators of foreground object locations. We first generate frame-wise spatiotemporal saliency maps using geodesic distance from these indicators. Building on the observation that foreground areas are surrounded by the regions with high spatiotemporal edge values, geodesic distance provides an initial estimation for foreground and background. Then, high-quality saliency results are produced via the geodesic distances to background regions in the subsequent frames. Through the resulting saliency maps, we build global appearance models for foreground and background. By imposing motion continuity, we establish a dynamic location model for each frame. Finally, the spatiotemporal saliency maps, appearance models and dynamic location models are combined into an energy minimization framework to attain both spatially and temporally coherent object segmentation. Extensive quantitative and qualitative experiments on benchmark video dataset demonstrate the superiority of the proposed method over the state-of-the-art algorithms.

Saliency propagation has been widely adopted for identifying the most attractive object in an image. The propagation sequence generated by existing saliency detection methods is governed by the spatial relationships of image regions, i.e., the saliency value is transmitted between two adjacent regions. However, for the inhomogeneous difficult adjacent regions, such a sequence may incur wrong propagations. In this paper, we attempt to manipulate the propagation sequence for optimizing the propagation quality. Intuitively, we postpone the propagations to difficult regions and meanwhile advance the propagations to less ambiguous simple regions. Inspired by the theoretical results in educational psychology, a novel propagation algorithm employing the teaching-to-learn and learning-to-teach strategies is proposed to explicitly improve the propagation quality. In the teaching-to-learn step, a teacher is designed to arrange the regions from simple to difficult and then assign the simplest regions to the learner. In the learning-to-teach step, the learner delivers its learning confidence to the teacher to assist the teacher to choose the subsequent simple regions. Due to the interactions between the teacher and learner, the uncertainty of original difficult regions is gradually reduced, yielding manifest salient objects with optimized background suppression. Extensive experimental results on benchmark saliency datasets demonstrate the superiority of the proposed algorithm over twelve representative saliency detectors.

In this paper, we introduce Cellular Automata--a dynamic evolution model to intuitively detect the salient object. First, we construct a background-based map using color and space contrast with the clustered boundary seeds. Then, a novel propagation mechanism dependent on Cellular Automata is proposed to exploit the intrinsic relevance of similar regions through interactions with neighbors. Impact factor matrix and coherence matrix are constructed to balance the influential power towards each cell's next state. The saliency values of all cells will be renovated simultaneously according to the proposed updating rule. It's surprising to find out that parallel evolution can improve all the existing methods to a similar level regardless of their original results. Finally, we present an integration algorithm in the Bayesian framework to take advantage of multiple saliency maps. Extensive experiments on six public datasets demonstrate that the proposed algorithm outperforms state-of-the-art methods.

Low-level saliency cues or priors do not produce good enough saliency detection results especially when the salient object presents in a low-contrast background with confusing visual appearance. This issue raises a serious problem for conventional approaches. In this paper, we tackle this problem by proposing a multi-context deep learning framework for salient object detection. We employ deep Convolutional Neural Networks to model saliency of objects in images. Global context and local context are both taken into account, and are jointly modeled in a unified multi-context deep learning framework. To provide a better initialization for training the deep neural networks, we investigate different pre-training strategies, and a task-specific pre-training scheme is designed to make the multi-context modeling suited for saliency detection. Furthermore, recently proposed contemporary deep models in the ImageNet Image Classification Challenge are tested, and their effectiveness in saliency detection are investigated. Our approach is extensively evaluated on five public datasets, and experimental results show significant and consistent improvements over the state-of-the-art methods.

We study the simultaneous detection of multiple structures in the presence of overwhelming number of outliers in a large population of points. Our approach reduces the problem to sampling an extremely sparse subset of the original population of data in one grab, followed by an unsupervised clustering of the population based on a set of instantiated models from this sparse subset. We show that the problem can be modeled using a multivariate hypergeometric distribution, and derive accurate mathematical bounds to determine a tight approximation to the sample size, leading thus to a sparse sampling strategy. We evaluate the method thoroughly in terms of accuracy, its behavior against varying input parameters, and comparison against existing methods, including the state of the art. The key features of the proposed approach are: (i) sparseness of the sampled set, where the level of sparseness is independent of the population size and the distribution of data, (ii) robustness in the presence of overwhelming number of outliers, and (iii) unsupervised detection of all model instances, i.e. without requiring any prior knowledge of the number of embedded structures. To demonstrate the generic nature of the proposed method, we show experimental results on different computer vision problems, such as detection of physical structures e.g. lines, planes, etc., as well as more abstract structures such as fundamental matrices, and homographies in multi-body structure from motion.

Although RGB-D sensors have enabled major breakthroughs for several vision tasks, such as 3D reconstruction, we have not attained the same level of success in high-level scene understanding. Perhaps one of the main reasons is the lack of a large-scale benchmark with 3D annotations and 3D evaluation metrics. In this paper, we introduce an RGB-D benchmark suite for the goal of advancing the state-of-the-arts in all major scene understanding tasks. Our dataset is captured by four different sensors and contains 10,335 RGB-D images, at a similar scale as PASCAL VOC. The whole dataset is densely annotated and includes 146,617 2D polygons and 64,595 3D bounding boxes with accurate object orientations, as well as a 3D room layout and scene category for each image. This dataset enables us to train data-hungry algorithms for scene-understanding tasks, evaluate them using meaningful 3D metrics, avoid overfitting to a small testing set, and study cross-sensor bias.

Image quality assessment (IQA) tries to estimate human perception based image visual quality in an objective manner. Existing approaches target this problem with or without reference images. For no-reference image quality assessment, there is no given reference image or any knowledge of the distortion type of the image. Previous approaches measure the image quality from signal level rather than semantic analysis. They typically depend on various features to represent local characteristic of an image. In this paper we propose a new no-reference (NR) image quality assessment (IQA) framework based on semantic obviousness. We discover that semantic-level factors affect human perception of image quality. With such observation, we explore semantic obviousness as a metric to perceive objects of an image. We propose to extract two types of features, one to measure the semantic obviousness of the image and the other to discover local characteristic. Then the two kinds of features are combined for image quality estimation. The principles proposed in our approach can also be incorporated with many existing IQA algorithms to boost their performance. We evaluate our approach on the LIVE dataset. Our approach is demonstrated to be superior to the existing NR-IQA algorithms and comparable to the state-of-the-art full-reference IQA (FR-IQA) methods. Cross-dataset experiments show the generalization ability of our approach.

This paper investigates how to perform robust and efficient unsupervised video segmentation while suppressing the effects of data noises and/or corruptions. We propose a general algorithm, called Sub-Optimal Low-rank Decomposition (SOLD), which pursues the low-rank representation for video segmentation. Given the supervoxels affinity matrix of an observed video sequence, SOLD seeks a sub-optimal solution by making the matrix rank explicitly determined. In particular, the affinity matrix with the rank fixed can be decomposed into two sub-matrices of low rank, and then we iteratively optimize them with closed-form solutions. Moreover, we incorporate a discriminative replication prior into our framework based on the obervation that small-size video patterns tend to recur frequently within the same object. The video can be segmented into several spatio-temporal regions by applying the Normalized-Cut (NCut) algorithm with the solved low-rank representation. To process the streaming videos, we apply our algorithm sequentially over a batch of frames over time, in which we also develop several temporal consistent constraints improving the robustness. Extensive experiments on the public benchmarks demonstrate superior performance of our framework over other state-of-the-art approaches.

Saliency in Context (SALICON) is an ongoing effort that aims at understanding and predicting visual attention. This paper presents a new method to collect large-scale human data during natural explorations on images. While current datasets present a rich set of images and task-specific annotations such as category labels and object segments, this work focuses on recording and logging how humans shift their attention during visual exploration. The goal is to offer new possibilities to (1) complement task-specific annotations to advance the ultimate goal in visual understanding, and (2) understand visual attention and learn saliency models, all with human attentional data at a much larger scale. We designed a mouse-contingent multi-resolutional paradigm based on neurophysiological and psychophysical studies of peripheral vision, to simulate the natural viewing behavior of humans. The new paradigm allowed using a general-purpose mouse instead of an eye tracker to record viewing behaviors, thus enabling large-scale data collection. The paradigm was validated with controlled laboratory as well as large-scale online data. We report in this paper a proof-of-concept SALICON dataset of human "free-viewing" data on 10,000 images from the Microsoft COCO (MS COCO) dataset with rich contextual information. We evaluated the use of the collected data in the context of saliency prediction, and demonstrated them a good source as ground truth for the evaluation of saliency algorithms.

Face recognition under viewpoint and illumination changes is a difficult problem, so many researchers have tried to solve this problem by producing the pose- and illumination- invariant feature. Zhu et al. [26] changed all arbitrary pose and illumination images to the frontal view image to use for the invariant feature. In this scheme, preserving identity while rotating pose image is a crucial issue. This paper proposes a new deep architecture based on a novel type of multitask learning, which can achieve superior performance in rotating to a target-pose face image from an arbitrary pose and illumination image while preserving identity. The target pose can be controlled by the user's intention. This novel type of multi-task model significantly improves identity preservation over the single task model. By using all the synthesized controlled pose images, called Controlled Pose Image (CPI), for the pose- illumination- invariant feature and voting among the multiple face recognition results, we clearly outperform the state-of-the-art algorithms by more than 4~6% on the MultiPIE dataset.