CVPR 2017 Papers

Feature pyramids are a basic component in recognition systems for detecting objects at different scales. But pyramid representations have been avoided in recent object detectors that are based on deep convolutional networks, partially because they are slow to compute and memory intensive. In this paper, we exploit the inherent multi-scale, pyramidal hierarchy of deep convolutional networks to construct feature pyramids with marginal extra cost. A top-down architecture with lateral connections is developed for building high-level semantic feature maps at all scales. This architecture, called a Feature Pyramid Network (FPN), shows significant improvement as a generic feature extractor in several applications. Using a basic Faster R-CNN system, our method achieves state-of-the-art single-model results on the COCO detection benchmark without bells and whistles, surpassing all existing single-model entries including those from the COCO 2016 challenge winners. In addition, our method can run at 5 FPS on a GPU and thus is a practical and accurate solution to multi-scale object detection. Code will be made publicly available.

Objects appear to scale differently in natural images. This fact requires methods dealing with object-centric tasks (e.g. object proposal) to have robust performance over variances in object scales. In the paper, we present a novel segment proposal framework, namely FastMask, which takes advantage of hierarchical features in deep convolutional neural networks to segment multi-scale objects in one shot. Innovatively, we adapt segment proposal network into three different functional components (body, neck and head). We further propose a weight-shared residual neck module as well as a scale-tolerant attentional head module for efficient one-shot inference. On MS COCO benchmark, the proposed FastMask outperforms all state-of-the-art segment proposal methods in average recall being 2 5 times faster. Moreover, with a slight trade-off in accuracy, FastMask can segment objects in near real time ( 13 fps) with 800*600 resolution images, demonstrating its potential in practical applications. Our implementation is available on https://github.com/voidrank/FastMask.

Fast-AT is an automatic thumbnail generation system based on deep neural networks. It is a fully-convolutional deep neural network, which learns specific filters for thumbnails of different sizes and aspect ratios. During inference, the appropriate filter is selected depending on the dimensions of the target thumbnail. Unlike most previous work, Fast-AT does not utilize saliency but addresses the problem directly. In addition, it eliminates the need to conduct region search over the saliency map. The model generalizes to thumbnails of different sizes including those with extreme aspect ratios and can generate thumbnails in real time. A data set of more than 70,000 thumbnail annotations was collected to train Fast-AT. We show competitive results in comparison to existing techniques.

Recent advances have enabled "oracle" classifiers that can classify across many classes and input distributions with high accuracy without retraining. However, these classifiers are relatively heavyweight, so that applying them to classify video is costly. We show that day-to-day video exhibits highly skewed class distributions over the short term, and that these distributions can be classified by much simpler models. We formulate the problem of detecting the short-term skews online and exploiting models based on it as a new sequential decision making problem dubbed the Online Bandit Problem, and present a new algorithm to solve it. When applied to recognizing faces in TV shows and movies, we realize end-to-end classification speedups of 2.4-7.8x/2.6-11.2x (on GPU/CPU) relative to a state-of-the-art convolutional neural network, at competitive accuracy.

Numerous methods have been proposed for person re-identification, most of which however neglect the matching efficiency. Recently, several hashing based approaches have been developed to make re-identification more scalable for large-scale gallery sets. Despite their efficiency, these works ignore cross-camera variations, which severely deteriorate the final matching accuracy. To address the above issues, we propose a novel hashing based method for fast person re-identification, namely Cross-camera Semantic Binary Transformation (CSBT). CSBT aims to transform original high-dimensional feature vectors into compact identity-preserving binary codes. To this end, CSBT first employs a subspace projection to mitigate cross-camera variations, by maximizing intra-person similarities and inter-person discrepancies. Subsequently, a binary coding scheme is proposed via seamlessly incorporating both the semantic pairwise relationships and local affinity information. Finally, a joint learning framework is proposed for simultaneous subspace projection learning and binary coding based on discrete alternating optimization. Experimental results on four benchmarks clearly demonstrate the superiority of CSBT over the state-of-the-art methods.

We propose a new multi-frame method for efficiently computing scene flow (dense depth and optical flow) and camera ego-motion for a dynamic scene observed from a moving stereo camera rig. Our technique also segments out moving objects from the rigid scene. In our method, we first estimate the disparity map and the 6-DOF camera motion using stereo matching and visual odometry. We then identify regions inconsistent with the estimated camera motion and compute per-pixel optical flow only at these regions. This flow proposal is fused with the camera motion-based flow proposal using fusion moves to obtain the final optical flow and motion segmentation. This unified framework benefits all four tasks -- stereo, optical flow, visual odometry and motion segmentation leading to overall higher accuracy and efficiency. Our method is currently ranked third on the KITTI 2015 scene flow benchmark. Furthermore, our CPU implementation runs in 2-3 seconds per frame which is 1-3 orders of magnitude faster than the top six methods. We also report a thorough evaluation on challenging Sintel sequences with fast camera and object motion, where our method consistently outperforms OSF [Menze2015], which is currently ranked second on the KITTI benchmark.

In this paper, we address a haze removal problem from a single nighttime image, even in the presence of varicolored and non-uniform illumination. The core idea lies in a novel maximum reflectance prior. We first introduce the nighttime hazy imaging model, which includes a local ambient illumination item in both direct attenuation term and scattering term. Then, we propose a simple but effective image prior, maximum reflectance prior, to estimate the varying ambient illumination. The maximum reflectance prior is based on a key observation: for most daytime haze-free image patches, each color channel has very high intensity at some pixels. For the nighttime haze image, the local maximum intensities at each color channel are mainly contributed by the ambient illumination. Therefore, we can directly estimate the ambient illumination and transmission map, and consequently restore a high quality haze-free image. Experimental results on various nighttime hazy images demonstrate the effectiveness of the proposed approach. In particular, our approach has the advantage of computational efficiency, which is 10-100 times faster than state-of-the-art methods.

We present Fast Fourier Color Constancy (FFCC), a color constancy algorithm which solves illuminant estimation by reducing it to a spatial localization task on a torus. By operating in the frequency domain, FFCC produces lower error rates than the previous state-of-the-art by 13-20% while being 250-3000 times faster. This unconventional approach introduces challenges regarding aliasing, directional statistics, and preconditioning, which we address. By producing a complete posterior distribution over illuminants instead of a single illuminant estimate, FFCC enables better training techniques, an effective temporal smoothing technique, and richer methods for error analysis. Our implementation of FFCC runs at 700 frames per second on a mobile device, allowing it to be used as an accurate, real-time, temporally-coherent automatic white balance algorithm.

In this paper we propose a novel boosting-based sliding window solution for object detection which can keep up with the precision of the state-of-the art deep learning approaches, while being 10 to 100 times faster. The solution takes advantage of multisensorial perception and exploits information from color, motion and depth. We introduce multimodal multiresolution filtering of signal intensity, gradient magnitude and orientation channels, in order to capture structure at multiple scales and orientations. To achieve scale invariant classification features, we analyze the effect of scale change on features for different filter types and propose a correction scheme. To improve recognition we incorporate 2D and 3D context by generating spatial, geometric and symmetrical channels. Finally, we evaluate the proposed solution on multiple benchmarks for the detection of pedestrians, cars and bicyclists. We achieve competitive results at over 25 frames per second.

We present a method for the fast 3D face reconstruction of people wearing glasses. Our method explicitly and robustly models the case in which a face to be reconstructed is partially occluded by glasses. We propose a simple and generic model for glasses that copes with a wide variety of different shapes, colors and styles, without the need for any database or learning. Our algorithm is simple, fast and requires only small amounts of both memory and runtime resources, allowing for a fast interactive 3D reconstruction on commodity mobile phones. The thorough evaluation of our approach on synthetic and real data demonstrates superior reconstruction results due to the explicit modeling of glasses.

Matrix and tensor factorization methods are often used for finding underlying low-dimensional patterns from noisy data. In this paper, we study non-linear tensor factoriza- tion methods based on deep variational autoencoders. Our approach is well-suited for settings where the relationship between the latent representation to be learned and the raw data representation is highly complex. We apply our ap- proach to a large dataset of facial expressions of movie- watching audiences (over 16 million faces). Our experi- ments show that compared to conventional linear factoriza- tion methods, our method achieves better reconstruction of the data, and further discovers interpretable latent factors.

In this work we pursue a data-driven approach to the problem of estimating surface normals from a single intensity image, focusing in particular on human faces. We introduce new methods to exploit the currently available facial databases for dataset construction and tailor a deep convolutional neural network to the task of estimating facial surface normals `in-the-wild'. We train a fully convolutional network that can accurately recover facial normals from images including a challenging variety of expressions and facial poses. We compare against state-of-the-art face Shape-from-Shading and 3D reconstruction techniques and show that the proposed network can recover substantially more accurate and realistic normals. Furthermore, in contrast to other existing face-specific surface recovery methods, we do not require the solving of an explicit alignment step due to the fully convolutional nature of our network.

In this paper, we revisit the LASSO sparse representation problem, which has been studied and used in a variety of different areas, ranging from signal processing and information theory to computer vision and machine learning. In the vision community, it found its way into many important applications, including face recognition, tracking, super resolution, image denoising, to name a few. Despite advances in efficient sparse algorithms, solving large-scale LASSO problems remains a challenge. To circumvent this difficulty, people tend to downsample and subsample the problem (e.g. via dimensionality reduction) to maintain a manageable sized LASSO, which usually comes at the cost of losing solution accuracy. This paper proposes a novel circulant reformulation of the LASSO that lifts the problem to a higher dimension, where ADMM can be efficiently applied to its dual form. Because of this lifting, all optimization variables are updated using only basic element-wise operations, the most computationally expensive of which is a 1D FFT. In this way, there is no need for a linear system solver nor matrix-vector multiplication. Since all operations in our FFTLasso method are element-wise, the subproblems are completely independent and can be trivially parallelized (e.g. on a GPU). The attractive computational properties of FFTLasso are verified by extensive experiments on synthetic and real data and on the face recognition task. They demonstrate that FFTLasso scales much more effectively than a state-of-the-art solver.

We present a descriptor, called fully convolutional self-similarity (FCSS), for dense semantic correspondence. To robustly match points among different instances within the same object class, we formulate FCSS using local self-similarity (LSS) within a fully convolutional network. In contrast to existing CNN-based descriptors, FCSS is inherently insensitive to intra-class appearance variations because of its LSS-based structure, while maintaining the precise localization ability of deep neural networks. The sampling patterns of local structure and the self-similarity measure are jointly learned within the proposed network in an end-to-end and multi-scale manner. As training data for semantic correspondence is rather limited, we propose to leverage object candidate priors provided in existing image datasets and also correspondence consistency between object pairs to enable weakly-supervised learning. Experiments demonstrate that FCSS outperforms conventional handcrafted descriptors and CNN-based descriptors on various benchmarks.

Improvements in color constancy have arisen from the use of convolutional neural networks (CNNs). However, the patch-based CNNs that exist for this problem are faced with the issue of estimation ambiguity, where a patch may contain insufficient information to establish a unique or even a limited possible range of illumination colors. Image patches with estimation ambiguity not only appear with great frequency in photographs, but also significantly degrade the quality of network training and inference. To overcome this problem, we present a fully convolutional network architecture in which patches throughout an image can carry different confidence weights according to the value they provide for color constancy estimation. These confidence weights are learned and applied within a novel pooling layer where the local estimates are merged into a global solution. With this formulation, the network is able to determine "what to learn" and "how to pool" automatically from color constancy datasets without additional supervision. The proposed network also allows for end-to-end training, and achieves higher efficiency and accuracy. On standard benchmarks, our network outperforms the previous state-of-the-art while achieving 120x greater efficiency.

Deep networks have shown impressive performance on many computer vision tasks. Recently, deep convolutional neural networks (CNNs) have been used to learn discriminative texture representations. One of the most successful approaches is Bilinear CNN model that explicitly captures the second order statistics within deep features. However, these networks cut off the first order information flow in the deep network and make gradient back-propagation difficult. We propose an effective fusion architecture - FASON that combines second order information flow and first order information flow. Our method allows gradients to back-propagate through both flows freely and can be trained effectively. We then build a multi-level deep architecture to exploit the first and second order information within different convolutional layers. Experiments show that our method achieves improvements over state-of-the-art methods on several benchmark datasets.

Many man-made objects have intrinsic symmetries and Manhattan structure. By assuming an orthographic projection model, this paper addresses the estimation of 3D structures and camera projection using symmetry and/or Manhattan structure cues, which occur when the input is single- or multiple-image from the same category, e.g., multiple different cars. Specifically, analysis on the single image case implies that Manhattan alone is sufficient to recover the camera projection, and then the 3D structure can be reconstructed uniquely exploiting symmetry. However, Manhattan structure can be difficult to observe from a single image due to occlusion. To this end, we extend to the multiple-image case which can also exploit symmetry but does not require Manhattan axes. We propose a novel rigid structure from motion method, exploiting symmetry and using multiple images from the same category as input. Experimental results on the Pascal3D+ dataset show that our method significantly outperforms baseline methods.

There have been remarkable improvements in the semantic labelling task in the recent years. However, the state of the art methods rely on large-scale pixel-level annotations. This paper studies the problem of training a pixel-wise semantic labeller network from image-level annotations of the present object classes. Recently, it has been shown that high quality seeds indicating discriminative object regions can be obtained from image-level labels. Without additional information, obtaining the full extent of the object is an inherently ill-posed problem due to co-occurrences. We propose using a saliency model as additional information and hereby exploit prior knowledge on the object extent and image statistics. We show how to combine both information sources in order to recover 80% of the fully supervised performance - which is the new state of the art in weakly supervised training for pixel-wise semantic labelling.

This paper presents a novel algorithm that utilizes a 2D floorplan to align panorama RGBD scans. While effective panorama RGBD alignment techniques exist, such a system requires extremely dense RGBD image sampling. Our approach can significantly reduce the number of necessary scans with the aid of a floorplan image. We formulate a novel Markov Random Field inference problem as a scan placement over the floorplan, as opposed to the conventional scan-to-scan alignment. The technical contributions lie in multi-modal image correspondence cues (between scans and schematic floorplan) as well as a novel coverage potential avoiding an inherent stacking bias. The proposed approach has been evaluated on five challenging large indoor spaces. To the best of our knowledge, we present the first effective system that utilizes a 2D floorplan image for building-scale 3D pointcloud alignment. The source code and the data are shared with the community to further enhance indoor mapping research.