ICCV 2017 Papers

Previous approaches on 3D shape segmentation mostly rely on heuristic processing and hand-tuned geometric descriptors. In this paper, we propose a novel 3D shape representation learning approach, Directionally Convolutional Network (DCN), to solve the shape segmentation problem. DCN extends convolution operations from images to the surface mesh of 3D shapes. With DCN, we learn effective shape representations from raw geometric features, i.e., face normals and distances, to achieve robust segmentation. More specifically, a two-stream segmentation framework is proposed: one stream is made up by the proposed DCN with the face normals as the input, and the other stream is implemented by a neural network with the face distance histogram as the input. The learned shape representations from the two streams are fused by an element-wise product. Finally, Conditional Random Field (CRF) is applied to optimize the segmentation. Through extensive experiments conducted on benchmark datasets, we demonstrate that our approach outperforms the current state-of-the-arts (both classic and deep learning-based) on a large variety of 3D shapes.

Deep Convolutional Neural Networks (CNNs) achieve substantial improvements in face detection in the wild. Classical CNN-based face detection methods simply stack successive layers of filters where an input sample should pass through all layers before reaching a face/non-face decision. Inspired by the fact that for face detection, filters in deeper layers can discriminate between difficult face/non-face samples while those in shallower layers can efficiently reject simple non-face samples, we propose Inside Cascaded Structure that introduces face/non-face classifiers at different layers within the same CNN. In the training phase, we propose data routing mechanism which enables different layers to be trained by different types of samples, and thus deeper layers can focus on handling more difficult samples compared with traditional architecture. In addition, we introduce a two-stream contextual CNN architecture that leverages body part information adaptively to enhance face detection. Extensive experiments on the challenging FDDB and WIDER FACE benchmarks demonstrate that our method achieves competitive accuracy to the state-of-the-art techniques while keeps real time performance.

Recent approaches for high accuracy detection and tracking of object categories in video consist of complex multistage solutions that become more cumbersome each year. In this paper we propose a ConvNet architecture that jointly performs detection and tracking, solving the task in a simple and effective way. Our contributions are threefold: (i) we set up a ConvNet architecture for simultaneous detection and tracking, using a multi-task objective for frame-based object detection and across-frame track regression; (ii) we introduce correlation features that represent object co-occurrences across time to aid the ConvNet during tracking; and (iii) we link the frame level detections based on our across-frame tracklets to produce high accuracy detections at the video level. Our ConvNet architecture for spatiotemporal object detection is evaluated on the large-scale ImageNet VID dataset where it achieves state-of-the-art results. Our approach provides better single model performance than the winning method of the last ImageNet challenge while being conceptually much simpler. Finally, we show that by increasing the temporal stride we can dramatically increase the tracker speed.

In this paper we present a novel approach for depth map enhancement from an RGB-D video sequence. The basic idea is to exploit the photometric information in the color sequence. Instead of making any assumption about surface albedo or controlled object motion and lighting, we use the lighting variations introduced by casual object movement. We are effectively calculating photometric stereo from a moving object under natural illuminations. The key technical challenge is to establish correspondences over the entire image set. We therefore develop a lighting insensitive robust pixel matching technique that out-performs optical flow method in presence of lighting variations. In addition we present an expectation-maximization framework to recover the surface normal and albedo simultaneously, without any regularization term. We have validated our method on both synthetic and real datasets to show its superior performance on both surface details recovery and intrinsic decomposition.

Previous CNN-based video super-resolution approaches need to align multiple frames to the reference. In this paper, we show that proper frame alignment and motion compensation is crucial for achieving high quality results. We accordingly propose a 'sub-pixel motion compensation' (SPMC) layer in a CNN framework. Analysis and experiments show the suitability of this layer in video SR. The final end-to-end, scalable CNN framework effectively incorporates the SPMC layer and fuses multiple frames to reveal image details. Our implementation can generate visually and quantitatively high-quality results, superior to current state-of-the-arts, without the need of parameter tuning.

Traditional imaging methods and computer vision algorithms are often ineffective when images are acquired in scattering media, such as underwater, fog, and biological tissue. Here, we explore the use of light field imaging and algorithms for image restoration and depth estimation that address the image degradation from the medium. Towards this end, we make the following three contributions. First, we present a new single image restoration algorithm which removes backscatter and attenuation from images better than existing methods, and apply it to each view in the light field. Second, we combine a novel transmission based depth cue with existing correspondence and defocus cues to improve light field depth estimation. In densely scattering media, our transmission depth cue is critical for depth estimation since the images have low signal to noise ratios which significantly degrades the performance of the correspondence and defocus cues. Finally, we propose shearing and refocusing multiple views of the light field to recover a single image of higher quality than what is possible from a single view. We demonstrate the benefits of our method through extensive experimental results in a water tank.

Shape reconstruction techniques using structured light have been widely researched and developed due to their robustness, high precision, and density. Because the techniques are based on decoding a pattern to find correspondences, it implicitly requires that the projected patterns be clearly captured by an image sensor, i.e., to avoid defocus and motion blur of the projected pattern. Although intensive researches have been conducted for solving defocus blur, few researches for motion blur and only solution is to capture with extremely fast shutter speed. In this paper, unlike the previous approaches, we actively utilize motion blur, which we refer to as a light flow, to estimate depth. Analysis reveals that minimum two light flows, which are retrieved from two projected patterns on the object, are required for depth estimation. To retrieve two light flows at the same time, two sets of parallel line patterns are illuminated from two video projectors and the size of motion blur of each line is precisely measured. By analyzing the light flows, i.e. lengths of the blurs, scene depth information is estimated. In the experiments, 3D shapes of fast moving objects, which are inevitably captured with motion blur, are successfully reconstructed by our technique.

Most natural videos contain numerous events. For example, in a video of a "man playing a piano", the video might also contain "another man dancing" or "a crowd clapping". We introduce the task of dense-captioning events, which involves both detecting and describing events in a video. We propose a new model that is able to identify all such events in a single pass of the video while simultaneously describing the detected events with natural language. Our model introduces a variant of an existing proposal module that is designed to capture both short as well as long events that span minutes. To capture the dependencies between the events in a video, our model introduces a new captioning module that uses contextual information from past and future events to jointly describe all events. We also introduce ActivityNet Captions, a large-scale benchmark for dense-captioning events. ActivityNet Captions contains 20k videos amounting to 849 video hours with 100k total descriptions, each with it's unique start and end time. Finally, we report performances of our model for dense-captioning events, video retrieval and localization.

In this work we introduce a structured prediction model that endows the Deep Gaussian Conditional Random Field (G-CRF) with a densely connected graph structure. We keep memory and computational complexity under control by expressing the pairwise interactions as inner products of low-dimensional, learnable embeddings. The G-CRF system matrix is therefore low-rank, allowing us to solve the resulting system in a few milliseconds on the GPU by using conjugate gradients. As in G-CRF, inference is exact, the unary and pairwise terms are jointly trained end-to-end by using analytic expressions for the gradients, while we also develop even faster, Potts-type variants of our embeddings. We show that the learned embeddings capture pixel-to-pixel affinities in a task-specific manner, while our approach achieves state of the art results on three challenging benchmarks, namely semantic segmentation, human part segmentation, and saliency estimation. Our implementation is fully GPU based, built on top of the Caffe library, and is available at https://github.com/siddharthachandra/gcrf-v2.0

Significant progress has been recently made in Non-Rigid Structure-from-Motion (NRSfM). However, existing methods do not handle poorly-textured surfaces that deform non-smoothly. These are nonetheless common occurrence in real-world applications. An important unanswered question is whether shading can be used to robustly handle these cases. Shading is complementary to motion because it constrains reconstruction densely at textureless regions, and has been used in several other reconstruction problems. The challenge we face is to simultaneously and densely estimate non-smooth, non-rigid shape from each image together with non-smooth, spatially-varying surface albedo (which is required to use shading). We tackle this using an energy-based formulation that combines a physical, discontinuity-preserving deformation prior with motion, shading and contour information. This is a largescale, highly non-convex optimization problem, and we propose a cascaded optimization that converges well without an initial estimate. Our approach works on both unorganized and organized small-sized image sets, and has been empirically validated on four real-world datasets for which all state-of-the-art approaches fail.

Subitizing (i.e., instant judgement on the number) and detection of salient objects are human inborn abilities. These two tasks influence each other in the human visual system. In this paper, we delve into the complementarity of these two tasks. We propose a multi-task deep neural network with weight prediction for salient object detection, where the parameters of an adaptive weight layer are dynamically determined by an auxiliary subitizing network. The numerical representation of salient objects is therefore embedded into the spatial representation. The proposed joint network can be trained end-to-end using back-propagation. Experiments show that the proposed multi-task network outperforms existing multi-task architectures, and the auxiliary subitizing network provides strong guidance to salient object detection by reducing false positives and producing coherent saliency maps. Moreover, the proposed method is an unconstrained method able to handle images with/without salient objects. Finally, we show state-of-theart performance on different salient object datasets.

The recent proliferation of high resolution cameras presents an opportunity to achieve unprecedented levels of precision in visual 3D reconstruction. Yet the camera calibration pipeline, developed decades ago using checkerboards, has remained the de facto standard. In this paper, we ask the question: are checkerboards the optimal pattern for high precision calibration? We empirically demonstrate that deltille grids (regular triangular tiling) produce the highest precision calibration of the possible tilings of Euclidean plane. We posit that they should be the new standard for high-precision calibration and present a complete ecosystem for calibration using deltille grids including: (1) a highly precise corner detection algorithm based on polynomial surface fitting; (2) an indexing scheme based on polarities extracted from the fitted surfaces; and (3) a 2D coding system for deltille grids, which we refer to as DelTags, in lieu of conventional matrix barcodes. We demonstrate state-of-the-art performance and apply the full calibration ecosystem through the use of 3D calibration objects for multiview camera calibration.

Convolutional neural networks (CNNs) are inherently limited to model geometric transformations due to the fixed geometric structures in its building modules. In this work, we introduce two new modules to enhance the transformation modeling capacity of CNNs, namely, deformable convolution and deformable RoI pooling. Both are based on the idea of augmenting the spatial sampling locations in the modules with additional offsets and learning the offsets from target tasks, without additional supervision. The new modules can readily replace their plain counterparts in existing CNNs and can be easily trained end-to-end by standard back-propagation, giving rise to deformable convolutional networks. Extensive experiments validate the effectiveness of our approach on sophisticated vision tasks of object detection and semantic segmentation. The code would be released.

In this paper, we address the problem of person re-identification, which refers to associating the persons captured from different cameras. We propose a simple yet effective human part-aligned representation for handling the body part misalignment problem. Our approach decomposes the human body into regions (parts) which are discriminative for person matching, accordingly computes the representations over the regions, and aggregates the similarities computed between the corresponding regions of a pair of probe and gallery images as the overall matching score. Our formulation, inspired by attention models, is a deep neural network modeling the three steps together, which is learnt through minimizing the triplet loss function without requiring body part labeling information. Unlike most existing deep learning algorithms that learn a global or spatial partition-based local representation, our approach performs human body partition, and thus is more robust to pose changes and various human spatial distributions in the person bounding box. Our approach shows state-of-the-art results over standard datasets, Market-1501, CUHK03, CUHK01 and VIPeR.

The problem of domain generalization is to learn from multiple training domains, and extract a domain-agnostic model that can then be applied to an unseen domain. Domain generalization (DG) has a clear motivation in contexts where there are target domains with distinct characteristics, yet sparse data for training. For example recognition in sketch images, which are distinctly more abstract and rarer than photos. Nevertheless, DG methods have primarily been evaluated on photo-only benchmarks focusing on alleviating the dataset bias where both problems of domain distinctiveness and data sparsity can be minimal. We argue that these benchmarks are overly straightforward, and show that simple deep learning baselines perform surprisingly well on them. In this paper, we make two main contributions: Firstly, we build upon the favorable domain shift-robust properties of deep learning methods, and develop a low-rank parameterized CNN model for end-to-end DG learning. Secondly, we develop a DG benchmark dataset covering photo, sketch, cartoon and painting domains. This is both more practically relevant, and harder (bigger domain shift) than existing benchmarks. The results show that our method outperforms existing DG alternatives, and our dataset provides a more significant DG challenge to drive future research.

This paper addresses the task of set prediction using deep learning. This is important because the output of many computer vision tasks, including image tagging and object detection, are naturally expressed as sets of entities rather than vectors. As opposed to a vector, the size of a set is not fixed in advance, and it is invariant to the ordering of entities within it. We define a likelihood for a set distribution and learn its parameters using a deep neural network. We also derive a loss for predicting a discrete distribution corresponding to set cardinality. Set prediction is demonstrated on the problem of multi-class image classification. Moreover, we show that the proposed cardinality loss can also trivially be applied to the tasks of object counting and pedestrian detection. Our approach outperforms existing methods in all three cases on standard datasets.

Creating road maps is essential to the success of many applications such as autonomous driving and city planning. Most approaches in industry focus on leveraging expensive sensors mounted on top of a fleet of cars. This results in very accurate estimates when using techniques that involve a user in the loop. However, these solutions are very expensive and have small coverage. In contrast, in this paper we propose an approach that directly estimates road topology from aerial images. This provides us with an affordable solution which has large coverage. Towards this goal, we take advantage of the latest developments in deep learning to have an initial segmentation of the aerial images. We then propose an algorithm that reasons about missing connections in the extracted road topology as a shortest path problem which can be solved efficiently. We demonstrate the effectiveness of our approach in the challenging TorontoCity dataset and show very significant improvements over the state-of-the-art.

We present a novel deep learning architecture for fusing static multi-exposure images. Current multi-exposure fusion (MEF) approaches use hand-crafted features to fuse input sequence. However, the weak hand-crafted representations are not robust to varying input conditions. Moreover, they perform poorly for extreme exposure image pairs. Thus, it is highly desirable to have a method that is robust to varying input conditions and capable of handling extreme exposure without artifacts. Deep representations have known to be robust to input conditions and have shown phenomenal performance in a supervised setting. However, the stumbling block in using deep learning for MEF was the lack of sufficient training data and an oracle to provide the ground-truth for supervision. To address the above issues, we have gathered a large dataset of multi-exposure image stacks for training and to circumvent the need for ground truth images, we propose an unsupervised deep learning framework for MEF utilizing a no-reference quality metric as loss function. The proposed approach uses a novel CNN architecture trained to learn the fusion operation without reference ground truth image. The model fuses a set of common low level features extracted from each image to generate artifact-free perceptually pleasing results. We perform extensive quantitative and qualitative evaluation and show that the proposed technique outperforms existing state-of-the-art approaches for a variety of natural images.

3D context has been shown to be an extremely important cue for scene understanding, yet very little research has been done on integrating context information with deep models. This paper presents an approach to embed 3D context into the topology of a neural network trained to perform holistic scene understanding. Given a depth image depicting a 3D scene, our network aligns the observed scene with a predefined 3D scene template, and then reasons about the existence and location of each object within the scene template. In doing so, our model recognizes multiple objects in a single forward pass of a 3D convolutional neural network, capturing both global scene and local object information simultaneously. To create training data for this 3D network, we generate partly hallucinated depth images which are rendered by replacing real objects with a repository of CAD models of the same object category. Extensive experiments demonstrate the effectiveness of our algorithm compared to the state of the art.