CVPR 2016 Papers

Supervised contour detection methods usually require many labeled training images to obtain satisfactory performance. However, a large set of annotated data might be unavailable or extremely labor intensive. In this paper, we investigate the usage of semi-supervised learning (SSL) to obtain competitive detection accuracy with very limited training data (three labeled images). Specifically, we propose a semi-supervised structured ensemble learning approach for contour detection built on structured random forests (SRF). To allow SRF to be applicable to unlabeled data, we present an effective sparse representation approach to capture inherent structure in image patches by finding a compact and discriminative low-dimensional subspace representation in an unsupervised manner, enabling the incorporation of abundant unlabeled patches with their estimated structured labels to help SRF perform better node splitting. We re-examine the role of sparsity and propose a novel and fast sparse coding algorithm to boost the overall learning efficiency. To the best of our knowledge, this is the first attempt to apply SSL for contour detection. Extensive experiments on the BSDS500 segmentation dataset and the NYU Depth dataset demonstrate the superiority of the proposed method.

The state-of-the-art in semantic segmentation is currently represented by fully convolutional networks (FCNs). However, FCNs use large receptive fields and many pooling layers, both of which cause blurring and low spatial resolution in the deep layers. As a result FCNs tend to produce segmentations that are poorly localized around object boundaries. Prior work has attempted to address this issue in post-processing steps, for example using a color-based CRF on top of the FCN predictions. However, these approaches require additional parameters and low-level features that are difficult to tune and integrate into the original network architecture. Additionally, most CRFs use color-based pixel affinities, which are not well suited for semantic segmentation and lead to spatially disjoint predictions. To overcome these problems, we introduce a Boundary Neural Field (BNF), which is a global energy model integrating FCN predictions with boundary cues. The boundary information is used to enhance semantic segment coherence and to improve object localization. Specifically, we first show that the convolutional filters of semantic FCNs provide good features for boundary detection. We then employ the predicted boundaries to define pairwise potentials in our energy. Finally, we show that our energy decomposes semantic segmentation into multiple binary problems, which can be relaxed for efficient global optimization. We report extensive experiments demonstrating that minimization of our global boundary-based energy yields results superior to prior globalization methods, both quantitatively as well as qualitatively.

This supplementary material provides additional illustrations, visualizations and experiments. We start by showing the color coding and label mapping used for the semantic and instance label results in the paper. Then we provide more details about the 3D fold/curb detection and parameter settings that are used in the paper. Next, we provide additional quantitative and qualitative semi-dense inference results for both semantic and instance segmentation. Finally, we show the ability of our method to annotate 3D point clouds with semantic and instance labels which is a byproduct of our approach.

Deep convolutional neural networks (CNNs) are the backbone of state-of-art semantic image segmentation systems. Recent work has shown that complementing CNNs with fully-connected conditional random fields (CRFs) can significantly enhance their object localization accuracy, yet dense CRF inference is computationally expensive. We propose replacing the fully-connected CRF with domain transform (DT), a modern edge-preserving filtering method in which the amount of smoothing is controlled by a reference edge map. Domain transform filtering is several times faster than dense CRF inference and we show that it yields comparable semantic segmentation results, accurately capturing object boundaries. Importantly, our formulation allows learning the reference edge map from intermediate CNN features instead of using the image gradient magnitude as in standard DT filtering. This produces task-specific edges in an end-to-end trainable system optimizing the target semantic segmentation quality.

Edge-preserving image operations aim at smoothing an image without blurring the edges. Many excellent edge-preserving filtering techniques have been proposed recently to reduce the computational complexity or/and separate different scale structures. They normally adopt a user-selected scale measurement to control the detail/texture smoothing. However, natural photos contain objects of different sizes which cannot be described by a single scale measurement. On the other hand, edge/contour detection/analysis is closely related to edge-preserving filtering and has achieved significant progress recently. Nevertheless, most of the state-of-the-art filtering techniques ignore the success in this area. Inspired by the fact that learning-based edge detectors/classifiers significantly outperform traditional manually-designed detectors, this paper proposes a learning-based edge-preserving filtering technique. It synergistically combines the efficiency of the recursive filter and the effectiveness of the recent edge detector for scale-aware edge-preserving filtering. Unlike previous filtering methods, the propose filter can efficiently extract subjectively-meaningful structures from natural scenes containing multiple-scale objects.

When human annotators are given a choice about what to label in an image, they apply their own subjective judgments on what to ignore and what to mention. We refer to these noisy "human-centric" annotations as exhibiting human reporting bias. Examples of such annotations include image tags and keywords found on photo sharing sites, or in datasets containing image captions. In this paper, we use these noisy annotations for learning visually correct image classifiers. Such annotations do not use consistent vocabulary, and miss a significant amount of the information present in an image; however, we demonstrate that the noise in these annotations exhibits structure and can be modeled. We propose an algorithm to decouple the human reporting bias from the correct visually grounded labels. Our results are highly interpretable for reporting "what's in the image" versus "what's worth saying." We demonstrate the algorithm's efficacy along a variety of metrics and datasets, including MS COCO and Yahoo Flickr 100M. We show significant improvements over traditional algorithms for both image classification and image captioning, doubling the performance of existing methods in some cases.

We introduce the novel problem of identifying the photographer behind a photograph. To explore the feasibility of current computer vision techniques to address this problem, we created a new dataset of over 180,000 images taken by 41 well-known photographers. Using this dataset, we examined the effectiveness of a variety of features (low and high-level, including CNN features) at identifying the photographer. We also trained a new deep convolutional neural network for this task. Our results show that high-level features greatly outperform low-level features. We provide qualitative results using these learned models that give insight into our method's ability to distinguish between photographers, and allow us to draw interesting conclusions about what specific photographers shoot. We also demonstrate two applications of our method.

Since scenes are composed in part of objects, accurate recognition of scenes requires knowledge about both scenes and objects. In this paper we address two related problems: 1) scale induced dataset bias in multi-scale convolutional neural network (CNN) architectures, and 2) how to combine effectively scene-centric and object-centric knowledge (i.e. Places and ImageNet) in CNNs. An earlier attempt, Hybrid-CNN, showed that incorporating ImageNet did not help much. Here we propose an alternative method taking the scale into account, resulting in significant recognition gains. By analyzing the response of ImageNet-CNNs and Places-CNNs at different scales we find that both operate in different scale ranges, so using the same network for all the scales induces dataset bias resulting in limited performance. Thus, adapting the feature extractor to each particular scale (i.e. scale-specific CNNs) is crucial to improve recognition, since the objects in the scenes have their specific range of scales. Experimental results show that the recognition accuracy highly depends on the scale, and that simple yet carefully chosen multi-scale combinations of ImageNet-CNNs and Places-CNNs, can push the state-of-the-art recognition accuracy in SUN397 up to 66.26% (and even 70.17% with deeper architectures, comparable to human performance).

Scene labeling task is to segment the image into meaningful regions and categorize them into classes of objects which comprised the image. Commonly used methods typically find the local features for each segment and label them using classifiers. Afterwards, labeling is smoothed in order to make sure that neighboring regions receive similar labels. However, these methods ignore expressive connections between labels and non-local dependencies among regions. In this paper, we propose to use a sparse estimation of precision matrix (also called concentration matrix), which is the inverse of covariance matrix of data obtained by graphical lasso to find interaction between labels and regions. To do this, we formulate the problem as an energy minimization over a graph, whose structure is captured by applying sparse constraint on the elements of the precision matrix. This graph encodes (or represents) only significant interactions and avoids a fully connected graph, which is typically used to reflect the long distance associations. We use local and global information to achieve better labeling. We assess our approach on three datasets and obtained promising results.

Image segmentation is a key component in many computer vision systems, and it is recovering a prominent spot in the literature as methods improve and overcome their limitations. The outputs of most recent algorithms are in the form of a hierarchical segmentation, which provides segmentation at different scales in a single tree-like structure. Commonly, these hierarchical methods start from some low-level features, and are not aware of the scale information of the different regions in them. As such, one might need to work on many different levels of the hierarchy to find the objects in the scene. This work tries to modify the existing hierarchical algorithm by improving their alignment, that is, by trying to modify the depth of the regions in the tree to better couple depth and scale. To do so, we first train a regressor to predict the scale of regions using mid-level features. We then define the anchor slice as the set of regions that better balance between over-segmentation and under-segmentation. The output of our method is an improved hierarchy, re-aligned by the anchor slice. To demonstrate the power of our method, we perform comprehensive experiments, which show that our method, as a post-processing step, can significantly improve the quality of the hierarchical segmentation representations, and ease the usage of hierarchical image segmentation to high-level vision tasks such as object segmentation. We also prove that the improvement generalizes well across different algorithms and datasets, with a low computational cost.

Person re-identification addresses the problem of matching people across disjoint camera views and extensive efforts have been made to seek either the robust feature representation or the discriminative matching metrics. However, most existing approaches focus on learning a fixed distance metric for all instance pairs, while ignoring the individuality of each person. In this paper, we formulate the person re-identification problem as an imbalanced classification problem and learn a classifier specifically for each pedestrian such that the matching model is highly tuned to the individual's appearance. To establish correspondence between feature space and classifier space, we propose a Least Square Semi-Coupled Dictionary Learning (LSSCDL) algorithm to learn a pair of dictionaries and a mapping function efficiently. Extensive experiments on a series of challenging databases demonstrate that the proposed algorithm performs favorably against the state-of-the-art approaches, especially on the rank-1 recognition rate.

Scene parsing has attracted a lot of attention in computer vision. While parametric models have proven effective for this task, they cannot easily incorporate new training data. By contrast, nonparametric approaches, which bypass any learning phase and directly transfer the labels from the training data to the query images, can readily exploit new labeled samples as they become available. Unfortunately, because of the computational cost of their label transfer procedures, state-of-the-art nonparametric methods typically filter out most training images to only keep a few relevant ones to label the query. As such, these methods throw away many images that still contain valuable information and generally obtain an unbalanced set of labeled samples. In this paper, we introduce a nonparametric approach to scene parsing that follows a sample-and-filter strategy. More specifically, we propose to sample labeled superpixels according to an image similarity score, which allows us to obtain a balanced set of samples. We then formulate label transfer as an efficient filtering procedure, which lets us exploit more labeled samples than existing techniques. Our experiments evidence the benefits of our approach over state-of-the-art nonparametric methods on two benchmark datasets.

Human eye fixations often correlate with locations of salient objects in the scene. However, only a handful of approaches have attempted to simultaneously address the related aspects of eye fixations and object saliency. In this work, we propose a deep convolutional neural network (CNN) capable of predicting eye fixations and segmenting salient objects in a unified framework. We design the initial network layers, shared between both the tasks, such that they capture the object level semantics and the global contextual aspects of saliency, while the deeper layers of the network address task specific aspects. In addition, our network captures saliency at multiple scales via inception-style convolution blocks. Our network shows a significant improvement over the current state-of-the-art for both eye fixation prediction and salient object segmentation across a number of challenging datasets.

In this paper, we propose a novel method to perform weakly-supervised image parsing based on the dictionary learning framework. To deal with the challenges caused by the label ambiguities, we design a saliency guided weight assignment scheme to boost the discriminative dictionary learning. More specifically, with a collection of tagged images, the proposed method first conducts saliency detection and automatically infers the confidence for each semantic class to be foreground or background. These clues are then incorporated to learn the dictionaries, the weights, as well as the sparse representation coefficients in the meanwhile. Once obtained the coefficients of a superpixel, we use a sparse representation classifier to determine its semantic label. The approach is validated on the MSRC21, PASCAL VOC07, and VOC12 datasets. Experimental results demonstrate the encouraging performance of our approach in comparison with some state-of-the-arts.

Due to the limited amount of training samples, fine-tuning pre-trained deep models online is prone to over-fitting. In this paper, we propose a sequential training method for convolutional neural networks (CNNs) to effectively transfer pre-trained deep features for online applications. We regard a CNN as an ensemble with each channel of the output feature map as an individual base learner. Each base learner is trained using different loss criterions to reduce correlation and avoid over-training. To achieve the best ensemble online, all the base learners are sequentially sampled into the ensemble via important sampling. To further improve the robustness of each base learner, we propose to train the convolutional layers with random binary masks, which serves as a regularization to enforce each base learner to focus on different input features. The proposed online training method is applied to visual tracking problem by transferring deep features trained on massive annotated visual data and is shown to significantly improve tracking performance. Extensive experiments are conducted on two challenging benchmark data set and demonstrate that our tracking algorithm can outperform state-of-the-art methods with a considerable margin.

Most convolutional neural networks (CNNs) lack midlevel layers that model semantic parts of objects. This limits CNN-based methods from reaching their full potential in detecting and utilizing small semantic parts in recognition. Introducing such mid-level layers can facilitate the extraction of part-specific features which can be utilized for better recognition performance. This is particularly important in the domain of fine-grained recognition. In this paper, we propose a new CNN architecture that integrates semantic part detection and abstraction (SPDA-CNN) for fine-grained classification. The proposed network has two sub-networks: one for detection and one for recognition. The detection sub-network has a novel top-down proposal method to generate small semantic part candidates for detection. The classification sub-network introduces novel part layers that extract features from parts detected by the detection sub-network, and combine them for recognition. As a result, the proposed architecture provides an end-to-end network that performs detection, localization of multiple semantic parts, and whole object recognition within one framework that shares the computation of convolutional filters. Our method outperforms state-of-the-art methods with a large margin for small parts detection (e.g. our precision of 93.40% vs the best previous precision of 74.00% for detecting the head on CUB-2011). It also compares favorably to the existing state-of-the-art on fine-grained classification, e.g. it achieves 85.14% accuracy on CUB-2011.

Light fields (LFs) are image-based representation that records the radiance along all rays along every direction through every point in space. Traditionally LFs are acquired by using a 2D grid of evenly spaced pinhole cameras or by translating a pinhole camera along the 2D grid using a robot arm. In this paper, we present a novel LF sampling scheme by exploiting a special non-centric camera called the crossed-slit or XSlit camera. An XSlit camera acquires rays that simultaneously pass through two oblique slits. We show that, instead of translating the camera as in the pinhole case, we can effectively sample the LF by rotating individual or both slits while keeping the camera fixed. This leads a "fixed-location" LF acquisition scheme. We further show through theoretical analysis and experiments that the resulting XSlit LFs provide several advantages: they provide more dense spatial-angular sampling, are amenable multi-view stereo matching and volumetric reconstruction, and can synthesize unique refocusing effects.

The vast majority of modern consumer-grade cameras employ a rolling shutter mechanism. In dynamic geometric computer vision applications such as visual SLAM, the so-called rolling shutter effect therefore needs to be properly taken into account. A dedicated relative pose solver appears to be the first problem to solve, as it is of eminent importance to bootstrap any derivation of multi-view geometry. However, despite its significance, it has received inadequate attention to date. This paper presents a detailed investigation of the geometry of the rolling shutter relative pose problem. We introduce the rolling shutter essential matrix, and establish its link to existing models such as the push-broom cameras, summarized in a clean hierarchy of multi-perspective cameras. The generalization of well-established concepts from epipolar geometry is completed by a definition of the Sampson distance in the rolling shutter case. The work is concluded with a careful investigation of the introduced epipolar geometry for rolling shutter cameras on several dedicated benchmarks.

We present a solution to the rolling shutter (RS) absolute camera pose problem with known vertical direction. Our new solver, R5Pup, is an extension of the general minimal solution R6P, which uses a double linearized RS camera model initialized by the standard perspective P3P. Here, thanks to using known vertical directions, we avoid double linearization and can get the camera absolute pose directly from the RS model without the initialization by a standard P3P. Moreover, we need only five 2D-to-3D matches while R6P needed six such matches. We demonstrate in simulated and real experiments that our new R5Pup is robust, fast and a very practical method for absolute camera pose computation for modern cameras on mobile devices. We compare our R5Pup to the state of the art RS and perspective methods and demonstrate that it outperforms them when vertical direction is known in the range of accuracy available on modern mobile devices. We also demonstrate that when using R5Pup solver in structure from motion (SfM) pipelines, it is better to transform already reconstructed scenes into the standard position, rather than using hard constraints on the verticality of up vectors.