CVPR 2018 Papers

Finding views with good photo composition is a challenging task for machine learning methods. A key difficulty is the lack of well annotated large scale datasets. Most existing datasets only provide a limited number of annotations for good views, while ignoring the comparative nature of view selection. In this work, we present the first large scale Comparative Photo Composition dataset, which contains over one million comparative view pairs annotated using a cost-effective crowdsourcing workflow. We show that these comparative view annotations are essential for training a robust neural network model for composition. In addition, we propose a novel knowledge transfer framework to train a fast view proposal network, which runs at 75+ FPS and achieves state-of-the-art performance in image cropping and thumbnail generation tasks on three benchmark datasets. The superiority of our method is also demonstrated in a user study on a challenging experiment, where our method significantly outperforms the baseline methods in producing diversified well-composed views.

We present a novel method to incorporate the recent advent in static saliency models to predict the saliency in videos. Our model augments the static saliency models with the Attentional Push effect of the photographer and the scene actors in a shared attention setting. We demonstrate that not only it is imperative to use static Attentional Push cues, noticeable performance improvement is achievable by learning the time-varying nature of Attentional Push. We propose a multi-stream Convolutional Long Short-Term Memory network (ConvLSTM) structure which augments state-of-the-art in static saliency models with dynamic Attentional Push. Our network contains four pathways, a saliency pathway and three Attentional Push pathways. The multi-pathway structure is followed by an augmenting convnet that learns to combine the complementary and time-varying outputs of the ConvLSTMs by minimizing the relative entropy between the augmented saliency and viewers fixation patterns on videos. We evaluate our model by comparing the performance of several augmented static saliency models with state-of-the-art in spatiotemporal saliency on three largest dynamic eye tracking datasets, HOLLYWOOD2, UCF-Sport and DIEM. Experimental results illustrates that solid performance gain is achievable using the proposed methodology.

In this work, we describe man-made structures via an appropriate structure assumption, called Atlanta world, which contains a vertical direction (typically the gravity direction) and a set of horizontal directions orthogonal to the vertical direction. Contrary to the commonly used Manhattan world assumption, the horizontal directions in Atlanta world are not necessarily orthogonal to each other. While Atlanta world permits to encompass a wider range of scenes, this makes the solution space larger and the problem more challenging. Given a set of inputs, such as lines in a calibrated image or surface normals, we propose the first globally optimal method of inlier set maximization for Atlanta direction estimation. We define a novel search space for Atlanta world, as well as its parameterization, and solve this challenging problem by a branch-and-bound framework. Experimental results with synthetic and real-world datasets have successfully confirmed the validity of our approach.

In this paper, we present a novel localized Generative Adversarial Net (GAN) to learn on the manifold of real data. Compared with the classic GAN that {em globally} parameterizes a manifold, the Localized GAN (LGAN) uses local coordinate charts to parameterize distinct local geometry of how data points can transform at different locations on the manifold. Specifically, around each point there exists a {em local} generator that can produce data following diverse patterns of transformations on the manifold. The locality nature of LGAN enables local generators to adapt to and directly access the local geometry without need to invert the generator in a global GAN. Furthermore, it can prevent the manifold from being locally collapsed to a dimensionally deficient tangent subspace by imposing an orthonormality prior between tangents. This provides a geometric approach to alleviating mode collapse at least locally on the manifold by imposing independence between data transformations in different tangent directions. We will also demonstrate the LGAN can be applied to train a robust classifier that prefers locally consistent classification decisions on the manifold, and the resultant regularizer is closely related with the Laplace-Beltrami operator. Our experiments show that the proposed LGANs can not only produce diverse image transformations, but also deliver superior classification performances.

We propose a method for human activity recognition from RGB data that does not rely on any pose information during test time, and does not explicitly calculate pose information internally. Instead, a visual attention module learns to predict glimpse sequences in each frame. These glimpses correspond to interest points in the scene that are relevant to the classified activities. No spatial coherence is forced on the glimpse locations, which gives the attention module liberty to explore different points at each frame and better optimize the process of scrutinizing visual information. Tracking and sequentially integrating this kind of unstructured data is a challenge, which we address by sep- arating the set of glimpses from a set of recurrent tracking/recognition workers. These workers receive glimpses, jointly performing subsequent motion tracking and activity prediction. The glimpses are soft-assigned to the workers, optimizing coherence of the assignments in space, time and feature space using an external memory module. No hard decisions are taken, i.e. each glimpse point is assigned to all existing workers, albeit with different importance. Our methods outperform the state-of-the-art on the largest human activity recognition dataset available to-date, NTU RGB+D, and on the Northwestern-UCLA Multiview Action 3D Dataset.

Perception and being active (having a certain level of motion freedom) are closely tied. Learning active perception and sensorimotor control in the physical world is cumbersome as existing algorithms are too slow to efficiently learn in real-time and robots are fragile and costly. This has given rise to learning in simulation which consequently casts a question on transferring to real-world. In this paper, we investigate learning a real-world perception for active agents, propose Gibson virtual environment for this purpose, and showcase a set of learned complex locomotion abilities. The primary characteristics of the learning environments, which transfer into the trained agents, are I) being from the real-world and reflecting its semantic complexity, II) having a mechanism to ensure no need to further domain adaptation prior to deployment of results in real-world, III) embodiment of the agent and making it subject to constraints of space and physics.

Gestures are a common form of human communication and important for human computer interfaces (HCI). Recent approaches to gesture recognition use deep learning methods, including multi-channel methods. We show that when spatial channels are focused on the hands, gesture recognition improves significantly, particularly when the channels are fused using a sparse network. Using this technique, we improve performance on the ChaLearn IsoGD dataset from a previous best of 67.71% to 82.07%, and on the NVIDIA dataset from 83.8% to 91.28%.

Localizing text in the wild is challenging in the situations of complicated geometric layout of the targets like random orientation and large aspect ratio. In this paper, we propose a geometry-aware modeling approach tailored for scene text representation with an end-to-end learning scheme. In our approach, a novel Instance Transformation Network (ITN) is presented to learn the geometry-aware representation encoding the unique geometric configurations of scene text instances with in-network transformation embedding, resulting in a robust and elegant framework to detect words or text lines at one pass. An end-to-end multi-task learning strategy with transformation regression, text/non-text classification and coordinate regression is adopted in the ITN. Experiments on the benchmark datasets demonstrate the effectiveness of the proposed approach in detecting scene text in various geometric configurations.

We propose a method for predicting the 3D shape of a deformable surface from a single view. By contrast with previous approaches, we do not need a pre-registered template of the surface, and our method is robust to the lack of texture and partial occlusions. At the core of our approach is a geometry-aware deep architecture that tackles the problem as usually done in analytic solutions: first perform 2D detection of the mesh and then estimate a 3D shape that is geometrically consistent with the image. We train this architecture in an end-to-end manner using a large dataset of synthetic renderings of shapes under different levels of deformation, material properties, textures and lighting conditions. We evaluate our approach on a test split of this dataset and available real benchmarks, consistently improving state-of-the-art solutions with a significantly lower computational time.

Maps are a key component in image-based camera localization and visual SLAM systems: they are used to establish geometric constraints between images, correct drift in relative pose estimation, and relocalize cameras after lost tracking. The exact definitions of maps, however, are often application-specific and hand-crafted for different scenarios (e.g. 3D landmarks, lines, planes, bags of visual words). We propose to represent maps as a deep neural net called MapNet, which enables learning a data-driven map representation. Unlike prior work on learning maps, MapNet exploits cheap and ubiquitous sensory inputs like visual odometry and GPS in addition to images and fuses them together for camera localization. Geometric constraints expressed by these inputs, which have traditionally been used in bundle adjustment or pose-graph optimization, are formulated as loss terms in MapNet training and also used during inference. In addition to directly improving localization accuracy, this allows us to update the MapNet (i.e., maps) in a self-supervised manner using additional unlabeled video sequences from the scene. We also propose a novel parameterization for camera rotation which is better suited for deep-learning based camera pose regression. Experimental results on both the indoor 7-Scenes and the outdoor Oxford RobotCar datasets show significant improvement over prior work. The MapNet project webpage is https://goo.gl/mRB3Au.

This paper tackles the problem of novel view synthesis from a single image. In particular, we target real-world scenes with rich geometric structure, a challenging task due to the large appearance variations of such scenes and the lack of simple 3D models to represent them. Modern, learning-based approaches mostly focus on appearance to synthesize novel views and thus tend to generate predictions that are inconsistent with the underlying scene structure. By contrast, in this paper, we propose to exploit the 3D geometry of the scene to synthesize a novel view. Specifically, we approximate a real-world scene by a fixed number of planes, and learn to predict a set of homographies and their corresponding region masks to transform the input image into a novel view. To this end, we develop a new region-aware geometric transform network that performs these multiple tasks in a common framework. Our results on the outdoor KITTI and the indoor ScanNet datasets demonstrate the effectiveness of our network to generate high-quality synthetic views that respect the scene geometry, thus outperforming the state-of-the-art methods.

It is often laborious and costly to manually annotate videos for training high-quality video recognition models, so there has been some work and interest in exploring alternative, cheap, and yet often noisy and indirect, training signals for learning the video representations. However, these signals are still coarse, supplying supervision at the whole video frame level, and subtle, sometimes enforcing the learning agent to solve problems that are even hard for humans. In this paper, we instead explore geometry, a grand new type of auxiliary supervision for the self-supervised learning of video representations. In particular, we extract pixel-wise geometry information as flow fields and disparity maps from synthetic imagery and real 3D movies. Although the geometry and high-level semantics are seemingly distant topics, surprisingly, we find that the convolutional neural networks pre-trained by the geometry cues can be effectively adapted to semantic video understanding tasks. In addition, we also find that a progressive training strategy can foster a better neural network for the video recognition task than blindly pooling the distinct sources of geometry cues together. Extensive results on video dynamic scene recognition and action recognition tasks show that our geometry guided networks significantly outperform the competing methods that are trained with other types of labeling-free supervision signals.

In this paper, we generalize the Stochastic Gradient Descent (SGD) and RMSProp algorithms to the setting of Riemannian optimization. SGD is a popular method for large scale optimization. In particular, it is widely used to train the weights of Deep Neural Networks. However, gradients computed using standard SGD can have large variance, which is detrimental for the convergence rate of the algorithm. Other methods such as RMSProp and ADAM address this issue. Nevertheless, these methods cannot be directly applied to constrained optimization problems. In this paper, we extend some popular optimization algorithm to the Riemannian (constrained) setting. We substantiate our proposed extensions with a range of relevant problems in machine learning such as incremental Principal Component Analysis, computating the Riemannian centroids of SPD matrices, and Deep Metric Learning. We achieve competitive results against the state of the art for fine-grained object recognition datasets.

Deep convolutional neural networks have been shown to be vulnerable to arbitrary geometric transformations. However, there is no systematic method to measure the invariance properties of deep networks to such transformations. We propose ManiFool as a simple yet scalable algorithm to measure the invariance of deep networks. In particular, our algorithm measures the robustness of deep networks to geometric transformations in a worst-case regime as they can be problematic for sensitive applications. Our extensive experimental results show that ManiFool can be used to measure the invariance of fairly complex networks on high dimensional datasets and these values can be used for analyzing the reasons for it. Furthermore, we build on ManiFool to propose a new adversarial training scheme and we show its effectiveness on improving the invariance properties of deep neural networks.

We propose a novel method for fitting multiple geometric models to multi-structural data via convex relaxation. Unlike greedy methods - which maximise the number of inliers - our approach efficiently searches for a soft assignment of points to geometric models by minimising the energy of the overall assignment. The inherently parallel nature of our approach, as compared to the sequential approach found in state-of-the-art energy minimisation techniques, allows for the elegant treatment of a scaling factor that occurs as the number of features in the data increases. This results in an energy minimisation that, per iteration, is as much as two orders of magnitude faster on comparable architectures thus bringing real-time, robust performance to a wider set of geometric multi-model fitting problems. We demonstrate the versatility of our approach on two canonical problems in estimating structure from images: plane extraction from RGB-D images and homography estimation from pairs of images. Our approach seamlessly adapts to the different metrics brought forth in these distinct problems. In both cases, we report results on publicly available data-sets that in most instances outperform the state-of-the-art while simultaneously presenting run-times that are as much as an order of magnitude faster.

We propose GeoNet, a jointly unsupervised learning framework for monocular depth, optical flow and ego-motion estimation from videos. The three components are coupled by the nature of 3D scene geometry, jointly learned by our framework in an end-to-end manner. Specifically, geometric relationships are extracted over the predictions of individual modules and then combined as an image reconstruction loss, reasoning about static and dynamic scene parts separately. Furthermore, we propose an adaptive geometric consistency loss to increase robustness towards outliers and non-Lambertian regions, which resolves occlusions and texture ambiguities effectively. Experimentation on the KITTI driving dataset reveals that our scheme achieves state-of-the-art results in all of the three tasks, performing better than previously unsupervised methods and comparably with supervised ones.

In this paper, we propose Geometric Neural Network (GeoNet) to jointly predict depth and surface normal maps from a single image. Building on top of two-stream CNNs, our GeoNet incorporates geometric relation between depth and surface normal via the new depth-to-normal and normal- to-depth networks. Depth-to-normal network exploits the least square solution of surface normal from depth and im- proves its quality with a residual module. Normal-to-depth network, contrarily, refines the depth map based on the con- straints from the surface normal through a kernel regression module, which has no parameter to learn. These two net- works enforce the underlying model to efficiently predict depth and surface normal for high consistency and corre- sponding accuracy. Our experiments on NYU v2 dataset verify that our GeoNet is able to predict geometrically con- sistent depth and normal maps. It achieves top performance on surface normal estimation and is on par with state-of-the- art depth estimation methods.

Generative Adversarial Nets (GANs) are very successful at modeling distributions from given samples, even in the high-dimensional case. However, their formulation is also known to be hard to optimize and often not stable. While this is particularly true for early GAN formulations, there has been significant empirically motivated and theoretically founded progress to improve stability, for instance, by using the Wasserstein distance rather than the Jenson-Shannon divergence. Here, we consider an alternative formulation for generative modeling based on random projections which, in its simplest form, results in a single objective rather than a saddle-point formulation. By augmenting this approach with a discriminator we improve its accuracy. We found our ap- proach to be significantly more stable compared to even the improved Wasserstein GAN. Further, unlike the traditional GAN loss, the loss formulated in our method is a good mea- sure of the actual distance between the distributions and, for the first time for GAN training, we are able to show estimates for the same.

Recent deep learning based approaches have shown promising results for the challenging task of inpainting large missing regions in an image. These methods can generate visually plausible image structures and textures, but often create distorted structures or blurry textures inconsistent with surrounding areas. This is mainly due to ineffectiveness of convolutional neural networks in explicitly borrowing or copying information from distant spatial locations. On the other hand, traditional texture and patch synthesis approaches are particularly suitable when it needs to borrow textures from the surrounding regions. Motivated by these observations, we propose a new deep generative model-based approach which can not only synthesize novel image structures but also explicitly utilize surrounding image features as references during network training to make better predictions. The model is a feed-forward, fully convolutional neural network which can process images with multiple holes at arbitrary locations and with variable sizes during the test time. Experiments on multiple datasets including faces (CelebA, CelebA-HQ), textures (DTD) and natural images (ImageNet, Places2) demonstrate that our proposed approach generates higher-quality inpainting results than existing ones. Code, demo and models are available at: https://github.com/JiahuiYu/generative_inpainting.