CVPR 2017 Papers

Statistical machine learning models that operate on manifold-valued data are being extensively studied in vision, motivated by applications in activity recognition, feature tracking and medical imaging. While non-parametric methods have been relatively well studied in the literature, efficient formulations for parametric models (which may offer benefits in small sample size regimes) have only emerged recently. So far, manifold-valued regression models (such as geodesic regression) are restricted to the analysis of cross-sectional data, i.e., the so-called "fixed effects" in statistics. But in most "longitudinal analysis" (e.g., when a participant provides multiple measurements, over time) the application of fixed effects models is problematic. In an effort to answer this need, this paper generalizes non-linear mixed effects model to the regime where the response variable is manifold-valued, i.e., f:R^d -> M. We derive the underlying model and estimation schemes and demonstrate the immediate benefits such a model can provide --- both for group level and individual level analysis --- on longitudinal brain imaging data. The direct consequence of our results is that longitudinal analysis of manifold-valued measurements (especially, the symmetric positive definite manifold) can be conducted in a computationally tractable manner.

In this paper, we propose an accurate edge detector using richer convolutional features (RCF). Since objects in natural images possess various scales and aspect ratios, learning the rich hierarchical representations is very critical for edge detection. CNNs have been proved to be effective for this task. In addition, the convolutional features in CNNs gradually become coarser with the increase of the receptive fields. According to these observations, we attempt to adopt richer convolutional features in such a challenging vision task. The proposed network fully exploits multiscale and multilevel information of objects to perform the image-to-image prediction by combining all the meaningful convolutional features in a holistic manner. Using VGG16 network, we achieve state-of-the-art performance on several available datasets. When evaluating on the well-known BSDS500 benchmark, we achieve ODS F-measure of 0.811 while retaining a fast speed (8 FPS). Besides, our fast version of RCF achieves ODS F-measure of 0.806 with 30 FPS.

Variable Projection (VarPro) is a framework to solve optimization problems efficiently by optimally eliminating a subset of the unknowns. It is in particular adapted for Separable Nonlinear Least Squares (SNLS) problems, a class of optimization problems including low-rank matrix factorization with missing data and affine bundle adjustment as instances. VarPro-based methods have received much attention over the last decade due to the experimentally observed large convergence basin for certain problem classes, where they have a clear advantage over standard methods based on Joint optimization over all unknowns. Yet no clear answers have been found in the literature as to why VarPro outperforms others and why Joint optimization, which has been successful in solving many computer vision tasks, fails on this type of problems. Also, the fact that VarPro has been mainly tested on small to medium-sized datasets has raised questions about its scalability. This paper intends to address these unsolved puzzles.

The success of many visual recognition tasks largely depends on a good similarity measure, and distance metric learning plays an important role in this regard. Meanwhile, Symmetric Positive Definite (SPD) matrix is receiving increased attention for feature representation in multiple computer vision applications. However, distance metric learning on SPD matrices has not been sufficiently researched. A few existing works approached this by learning either d^2 xp or d xk transformation matrix for dxd SPD matrices. Different from these methods, this paper proposes a new member to the family of distance metric learning for SPD matrices. It learns only d parameters to adjust the eigenvalues of the SPD matrices through an efficient optimisation scheme. Also, it is shown that the proposed method can be interpreted as learning a sample-specific transformation matrix, instead of the fixed transformation matrix learned for all the samples in the existing works. The optimised d parameters can be used to "massage" the SPD matrices for better discrimination while still keeping them in the original space. From this perspective, the proposed method complements, rather than competes with, the existing linear-transformation-based methods, as the latter can always be applied to the output of the former to perform distance metric learning in further. The proposed method has been tested on multiple SPD-based visual representation data sets used in the literature, and the results demonstrate its interesting properties and attractive performance.

We propose the residual expansion (RE) algorithm: a global (or near-global) optimization method for nonconvex least squares problems. Unlike most existing nonconvex optimization techniques, the RE algorithm is not based on either stochastic or multi-point searches; therefore, it can achieve fast global optimization. Moreover, the RE algorithm is easy to implement and successful in high-dimensional optimization. The RE algorithm exhibits excellent empirical performance in terms of k-means clustering, point-set registration, optimized product quantization, and blind image deblurring.

We propose a new deep network architecture for removing rain streaks from individual images based on the deep convolutional neural network (CNN). Inspired by the deep residual network (ResNet) that simplifies the learning process by changing the mapping form, we propose a deep detail network to directly reduce the mapping range from input to output, which makes the learning process easier. To further improve the de-rained result, we use a priori image domain knowledge by focusing on high frequency detail during training, which removes background interference and focuses the model on the structure of rain in images. This demonstrates that a deep architecture not only has benefits for high-level vision tasks but also can be used to solve low-level imaging problems. Though we train the network on synthetic data, we find that the learned network generalizes well to real-world test images. Experiments show that the proposed method significantly outperforms state-of-the-art methods on both synthetic and real-world images in terms of both qualitative and quantitative measures. We discuss applications of this structure to denoising and JPEG artifact reduction at the end of the paper.

Past research on facial expressions have used relatively limited datasets, which makes it unclear whether current methods can be employed in real world. In this paper, we present a novel database, RAF-DB, which contains about 30000 facial images from thousands of individuals. Each image has been individually labeled about 40 times, then EM algorithm was used to filter out unreliable labels. Crowdsourcing reveals that real-world faces often express compound emotions, or even mixture ones. For all we know, RAF-DB is the first database that contains compound expressions in the wild. Our cross-database study shows that the action units of basic emotions in RAF-DB are much more diverse than, or even deviate from, those of lab-controlled ones. To address this problem, we propose a new DLP-CNN (Deep Locality-Preserving CNN) method, which aims to enhance the discriminative power of deep features by preserving the locality closeness while maximizing the inter-class scatters. The benchmark experiments on the 7-class basic expressions and 11-class compound expressions, as well as the additional experiments on SFEW and CK+ databases, show that the proposed DLP-CNN outperforms the state-of-the-art handcrafted features and deep learning based methods for the expression recognition in the wild.

Image scene understanding requires learning the relationships between objects in the scene. A scene with many objects may have only a few individual interacting objects (e.g., in a party image with many people, only a handful of people might be speaking with each other). To detect all relationships, it would be inefficient to first detect all individual objects and then classify all pairs; not only is the number of all pairs quadratic, but classification requires limited object categories, which is not scalable for real-world images. In this paper we address these challenges by using pairs of related regions in images to train a relationship proposer that at test time produces a manageable number of related regions. We name our model the Relationship Proposal Network (Rel-PN). Like object proposals, our Rel-PN is class-agnostic and thus scalable to an open vocabulary of objects. We demonstrate the ability of our Rel-PN to localize relationships with only a few thousand proposals. We demonstrate its performance on the Visual Genome dataset and compare to other baselines that we designed. We also conduct experiments on a smaller subset of 5,000 images with over 37,000 related regions and show promising results.

The 3D shapes of faces are well known to be discriminative. Yet despite this, they are rarely used for face recognition and always under controlled viewing conditions. We claim that this is a symptom of a serious but often overlooked problem with existing methods for single view 3D face reconstruction: when applied "in the wild", their 3D estimates are either unstable and change for different photos of the same subject or they are over-regularized and generic. In response, we describe a robust method for regressing discriminative 3D morphable face models (3DMM). We use a convolutional neural network (CNN) to regress 3DMM shape and texture parameters directly from an input photo. We overcome the shortage of training data required for this purpose by offering a method for generating huge numbers of labeled examples. The 3D estimates produced by our CNN surpass state of the art accuracy on the MICC data set. Coupled with a 3D-3D face matching pipeline, we show the first competitive face recognition results on the LFW, YTF and IJB-A benchmarks using 3D face shapes as representations, rather than the opaque deep feature vectors used by other modern systems.

The images taken through glass often capture a target transmitted scene as well as undesired reflected scenes. In this paper, we propose a low-rank matrix completion algorithm to remove reflection artifacts automatically from multiple glass images taken at slightly different camera locations. We assume that the transmitted scenes are more dominant than the reflected scenes in typical glass images. We first warp the multiple glass images to a reference image, where the gradients are consistent in the transmission images while the gradients are varying across the reflection images. Based on this observation, we compute a gradient reliability such that the pixels belonging to the salient edges of the transmission image are assigned high reliability. Then we suppress the gradients of the reflection images and recover the gradients of the transmission images only, by solving a low-rank matrix completion problem in gradient domain. We reconstruct an original transmission image using the resulting optimal gradient map. Experimental results show that the proposed algorithm removes the reflection artifacts from glass images faithfully and outperforms the existing algorithms on typical glass images.

Separating an image into reflectance and shading layers poses a challenge for learning approaches because no large corpus of precise and realistic ground truth decompositions exists. The Intrinsic Images in the Wild (IIW) dataset provides a sparse set of relative human reflectance judgments, which serves as a standard benchmark for intrinsic images. A number of methods use IIW to learn statistical dependencies between the images and their reflectance layer. Although learning plays an important role for high performance, we show that a standard signal processing technique achieves performance on par with current state-of-the-art. We propose a loss function for CNN learning of dense reflectance predictions. Our results show a simple pixel-wise decision, without any context or prior knowledge, is sufficient to provide a strong baseline on IIW. This sets a competitive baseline which only two other approaches surpass. We then develop a joint bilateral filtering method that implements strong prior knowledge about reflectance constancy. This filtering operation can be applied to any intrinsic image algorithm and we improve several previous results achieving a new state-of-the-art on IIW. Our findings suggest that the effect of learning-based approaches may have been over-estimated so far. Explicit prior knowledge is still at least as important to obtain high performance in intrinsic image decompositions.

Recently, very deep convolutional neural networks (CNNs) have shown outstanding performance in object recognition and have also been the first choice for dense classification problems such as semantic segmentation. However, repeated subsampling operations like pooling or convolution striding in deep CNNs lead to a significant decrease in the initial image resolution. Here, we present RefineNet, a generic multi-path refinement network that explicitly exploits all the information available along the down-sampling process to enable high-resolution prediction using long-range residual connections. In this way, the deeper layers that capture high-level semantic features can be directly refined using fine-grained features from earlier convolutions. The individual components of RefineNet employ residual connections following the identity mapping mindset, which allows for effective end-to-end training. Further, we introduce chained residual pooling, which captures rich background context in an efficient manner. We carry out comprehensive experiments and set new state-of-the-art results on seven public datasets. In particular, we achieve an intersection-over-union score of 83.4 on the challenging PASCAL VOC 2012 dataset, which is the best reported result to date.

Modeling of high order interactional context, e.g., group interaction, lies in the central of collective/group activity recognition. However, most of the previous activity recognition methods do not offer a flexible and scalable scheme to handle the high order context modeling problem. To explicitly address this fundamental bottleneck, we propose a recurrent interactional context modeling scheme based on LSTM network. By utilizing the information propagation/aggregation capability of LSTM, the proposed scheme unifies the interactional feature modeling process for single person dynamics, intra-group (e.g., persons within a group) and inter-group(e.g., group to group)interactions. The proposed high order context modeling scheme produces more discriminative/descriptive interactional features. It is very flexible to handle a varying number of input instances (e.g., different number of persons in a group or different number of groups) and linearly scalable to high order context modeling problem. Extensive experiments on two benchmark collective/group activity datasets demonstrate the effectiveness of the proposed method.

This work presents a weakly supervised framework with deep neural networks for vision-based continuous sign language recognition, where the ordered gloss labels but no exact temporal locations are available with the video of sign sentence, and the amount of labeled sentences for training is limited. Our approach addresses the mapping of video segments to glosses by introducing recurrent convolutional neural network for spatio-temporal feature extraction and sequence learning. We design a three-stage optimization process for our architecture. First, we develop an end-to-end sequence learning scheme and employ connectionist temporal classification (CTC) as the objective function for alignment proposal. Second, we take the alignment proposal as stronger supervision to tune our feature extractor. Finally, we optimize the sequence learning model with the improved feature representations, and design a weakly supervised detection network for regularization. We apply the proposed approach to a real-world continuous sign language recognition benchmark, and our method, with no extra supervision, achieves results comparable to the state-of-the-art.

Convolutional neural networks have enabled accurate image super-resolution in real-time. However, recent attempts to benefit from temporal correlations in video super-resolution have been limited to naive or inefficient architectures. In this paper, we introduce spatio-temporal sub-pixel convolution networks that effectively exploit temporal redundancies and improve reconstruction accuracy while maintaining real-time speed. Specifically, we discuss the use of early fusion, slow fusion and 3D convolutions for the joint processing of multiple consecutive video frames. We also propose a novel joint motion compensation and video super-resolution algorithm that is orders of magnitude more efficient than competing methods, relying on a fast multi-resolution spatial transformer module that is end-to-end trainable. These contributions provide both higher accuracy and temporally more consistent videos, which we confirm qualitatively and quantitatively. Relative to single-frame models, spatio-temporal networks can either reduce the computational cost by 30% whilst maintaining the same quality or provide a 0.2dB gain for a similar computational cost. Results on publicly available datasets demonstrate that the proposed algorithms surpass current state-of-the-art performance in both accuracy and efficiency.

Recent research endeavors have shown the potential of using feed-forward convolutional neural networks to accomplish fast style transfer for images. In this work, we take one step further to explore the possibility of exploiting a feed-forward network to perform style transfer for videos and simultaneously maintain temporal consistency among stylized video frames. Our feed-forward network is trained by enforcing the outputs of consecutive frames to be both well stylized and temporally consistent. More specifically, a hybrid loss is proposed to capitalize on the content information of input frames, the style information of a given style image, and the temporal information of consecutive frames. To calculate the temporal loss during the training stage, a novel two-frame synergic training mechanism is proposed. Compared with directly applying an existing image style transfer method to videos, our proposed method employs the trained network to yield temporally consistent stylized videos which are much more visually pleasant. In contrast to the prior video style transfer method which relies on time-consuming optimization on the fly, our method runs in real time while generating competitive visual results.

We present a novel method to track 3D models in color and depth data. To this end, we introduce approximations that accelerate the state-of-the-art in region-based tracking by an order of magnitude while retaining similar accuracy. Furthermore, we show how the method can be made more robust in the presence of depth data and consequently formulate a new joint contour and ICP tracking energy. We present better results than the state-of-the-art while being much faster then most other methods and achieving all of the above on a single CPU core.

When considering person re-identification (re-ID) as a retrieval process, re-ranking is a critical step to improve its accuracy. Yet in the re-ID community, limited effort has been devoted to re-ranking, especially those fully automatic, unsupervised solutions. In this paper, we propose a k-reciprocal encoding method to re-rank the re-ID results. Our hypothesis is that if a gallery image is similar to the probe in the k-reciprocal nearest neighbors, it is more likely to be a true match. Specifically, given an image, a k-reciprocal feature is calculated by encoding its k-reciprocal nearest neighbors into a single vector, which is used for re-ranking under the Jaccard distance. The final distance is computed as the combination of the original distance and the Jaccard distance. Our re-ranking method does not require any human interaction or any labeled data, so it is applicable to large-scale datasets. Experiments on the large-scale Market-1501, CUHK03, MARS, and PRW datasets confirm the effectiveness of our method.

We present a method for radiometric calibration of cameras from a single image that contains a human face. This technique takes advantage of a low-rank property that exists among certain skin albedo gradients because of the pigments within the skin. This property becomes distorted in images that are captured with a non-linear camera response function, and we perform radiometric calibration by solving for the inverse response function that best restores this low-rank property in an image. Although this work makes use of the color properties of skin pigments, we show that this calibration is unaffected by the color of scene illumination or the sensitivities of the camera's color filters. Our experiments validate this approach on a variety of images containing human faces, and show that faces can provide an important source of calibration data in images where existing radiometric calibration techniques perform poorly.