CVPR 2019 Papers

Compositing a realistic image is a challenging task and usually requires considerable human supervision using professional image editing software. In this work we propose a generative adversarial network (GAN) architecture for automatic image compositing. The proposed model consists of four sub-networks: a transformation network that improves the geometric and color consistency of the composite image, a refinement network that polishes the boundary of the composite image, and a pair of discriminator network and a segmentation network for adversarial learning. Experimental results on both synthesized images and real images show that our model, Geometrically and Color Consistent GANs (GCC-GANs), can automatically generate realistic composite images compared to several state-of-the-art methods, and does not require any manual effort.

While deep convolutional neural networks (CNNs) have achieved impressive success in image denoising with additive white Gaussian noise (AWGN), their performance remains limited on real-world noisy photographs. The main reason is that their learned models are easy to overfit on the simplified AWGN model which deviates severely from the complicated real-world noise model. In order to improve the generalization ability of deep CNN denoisers, we suggest training a convolutional blind denoising network (CBDNet) with more realistic noise model and real-world noisy-clean image pairs. On the one hand, both signal-dependent noise and in-camera signal processing pipeline is considered to synthesize realistic noisy images. On the other hand, real-world noisy photographs and their nearly noise-free counterparts are also included to train our CBDNet. To further provide an interactive strategy to rectify denoising result conveniently, a noise estimation subnetwork with asymmetric learning to suppress under-estimation of noise level is embedded into CBDNet. Extensive experimental results on three datasets of real-world noisy pho- tographs clearly demonstrate the superior performance of CBDNet over state-of-the-arts in terms of quantitative met- rics and visual quality. The code has been made available at https://github.com/GuoShi28/CBDNet.

The reconstruction and analysis of tree-like topological structures in the biomedical images is crucial for biologists and surgeons to understand biomedical conditions and plan surgical procedures. The underlying tree-structure topology reveals how different curvilinear components are anatomically connected to each other. Existing automated topology reconstruction methods have great difficulty in identifying the connectivity when two or more curvilinear components cross or bifurcate, due to their projection ambiguity, imaging noise and low contrast. In this paper, we propose a novel curvilinear structural similarity measure to guide a dominant-set clustering approach to address this indispensable issue. The novel similarity measure takes into account both intensity and geometric properties in representing the curvilinear structure locally and globally, and group curvilinear objects at crossover points into different connected branches by dominant-set clustering. The proposed method is applicable to different imaging modalities, and quantitative and qualitative results on retinal vessel, plant root, and neuronal network datasets show that our methodology is capable of advancing the current state-of-the-art techniques.

3D point cloud generation is of great use for 3D scene modeling and understanding. Real-world 3D object point clouds can be properly described by a collection of low-level and high-level structures such as surfaces, geometric primitives, semantic parts,etc. In fact, there exist many different representations of a 3D object point cloud as a set of point groups. Existing frameworks for point cloud genera-ion either do not consider structure in their proposed solutions, or assume and enforce a specific structure/topology,e.g. a collection of manifolds or surfaces, for the generated point cloud of a 3D object. In this work, we pro-pose a novel decoder that generates a structured point cloud without assuming any specific structure or topology on the underlying point set. Our decoder is softly constrained to generate a point cloud following a hierarchical rooted tree structure. We show that given enough capacity and allowing for redundancies, the proposed decoder is very flexible and able to learn any arbitrary grouping of points including any topology on the point set. We evaluate our decoder on the task of point cloud generation for 3D point cloud shape completion. Combined with encoders from existing frameworks, we show that our proposed decoder significantly outperforms state-of-the-art 3D point cloud completion methods on the Shapenet dataset

This paper proposes a method that uses deep convolutional neural networks to achieve automatic and accurate tooth instance segmentation and identification from CBCT (cone beam CT) images for digital dentistry. The core of our method is a two-stage network. In the first stage, an edge map is extracted from the input CBCT image to enhance image contrast along shape boundaries. Then this edge map and the input images are passed to the second stage. In the second stage, we build our network upon the 3D region proposal network (RPN) with a novel learned-similarity matrix to help efficiently remove redundant proposals, speed up training and save GPU memory. To resolve the ambiguity in the identification task, we encode teeth spatial relationships as an additional feature input in the identification task, which helps to remarkably improve the identification accuracy. Our evaluation, comparison and comprehensive ablation studies demonstrate that our method produces accurate instance segmentation and identification results automatically and outperforms the state-of-the-art approaches. To the best of our knowledge, our method is the first to use neural networks to achieve automatic tooth segmentation and identification from CBCT images.

This paper focuses on the temporal aspect for recognizing human activities in videos; an important visual cue that has long been undervalued. We revisit the conventional definition of activity and restrict it to Complex Action: a set of one-actions with a weak temporal pattern that serves a specific purpose. Related works use spatiotemporal 3D convolutions with fixed kernel size, too rigid to capture the varieties in temporal extents of complex actions, and too short for long-range temporal modeling. In contrast, we use multi-scale temporal convolutions, and we reduce the complexity of 3D convolutions. The outcome is Timeception convolution layers, which reasons about minute-long temporal patterns, a factor of 8 longer than best related works. As a result, Timeception achieves impressive accuracy in recognizing the human activities of Charades, Breakfast Actions and MultiTHUMOS. Further, we demonstrate that Timeception learns long-range temporal dependencies and tolerate temporal extents of complex actions.

The goal of human action anticipation is to predict future actions. Ideally, in real-world applications such as video surveillance and self-driving systems, future actions should not only be predicted with high accuracy but also at arbitrary and variable time-horizons ranging from short- to long-term predictions. Current work mostly focuses on predicting the next action and thus long-term prediction is achieved by recursive prediction of each next action, which is both inefficient and accumulates errors. In this paper, we propose a novel time-conditioned method for efficient and effective long-term action anticipation. There are two key ingredients to our approach. First, by explicitly conditioning our anticipation network on time allows to efficiently anticipate also long-term actions. And second, we propose an attended temporal feature and a time-conditioned skip connection to extract relevant and useful information from observations for effective anticipation. We conduct extensive experiments on the large-scale Epic-Kitchen and the 50Salads Datasets. Experimental results show that the proposed method is capable of anticipating future actions at both short-term and long-term, and achieves state-of-the-art performance.

Evaluation protocols play key role in the developmental progress of text detection methods. There are strict requirements to ensure that the evaluation methods are fair, objective and reasonable. However, existing metrics exhibit some obvious drawbacks: 1) They are not goal-oriented; 2) they cannot recognize the tightness of detection methods; 3) existing one-to-many and many-to-one solutions involve inherent loopholes and deficiencies. Therefore, this paper proposes a novel evaluation protocol called Tightness-aware Intersect-over-Union (TIoU) metric that could quantify completeness of ground truth, compactness of detection, and tightness of matching degree. Specifically, instead of merely using the IoU value, two common detection behaviors are properly considered; meanwhile, directly using the score of TIoU to recognize the tightness. In addition, we further propose a straightforward method to address the annotation granularity issue, which can fairly evaluate word and text-line detections simultaneously. By adopting the detection results from published methods and general object detection frameworks, comprehensive experiments on ICDAR 2013 and ICDAR 2015 datasets are conducted to compare recent metrics and the proposed TIoU metric. The comparison demonstrated some promising new prospects, e.g., determining the methods and frameworks for which the detection is tighter and more beneficial to recognize. Our method is extremely simple; however, the novelty is none other than the proposed metric can utilize simplest but reasonable improvements to lead to many interesting and insightful prospects and solving most the issues of the previous metrics. The code is publicly available at https://github.com/Yuliang-Liu/TIoU-metric.

Current wisdom suggests more labeled image data is always better, and obtaining labels is the bottleneck. Yet curating a pool of sufficiently diverse and informative images is itself a challenge. In particular, training image curation is problematic for fine-grained attributes, where the subtle visual differences of interest may be rare within traditional image sources. We propose an active image generation approach to address this issue. The main idea is to jointly learn the attribute ranking task while also learning to generate novel realistic image samples that will benefit that task. We introduce an end-to-end framework that dynamically "imagines" image pairs that would confuse the current model, presents them to human annotators for labeling, then improves the predictive model with the new examples. On two datasets, we show that by thinking outside the pool of real images, our approach gains generalization accuracy on challenging fine-grained attribute comparisons.

True video understanding requires making sense of non-lambertian scenes where the color of light arriving at the camera sensor encodes information about not just the last object it collided with, but about multiple mediums -- colored windows, dirty mirrors, smoke or rain. Layered video representations have the potential of accurately modelling realistic scenes but have so far required stringent assumptions on motion, lighting and shape. Here we propose a learning-based approach for multi-layered video representation: we introduce novel uncertainty-capturing 3D convolutional architectures and train them to separate blended videos. We show that these models then generalize to single videos, where they exhibit interesting abilities: color constancy, factoring out shadows and separating reflections. We present quantitative and qualitative results on real world videos.

As deep learning continues to make progress for challenging perception tasks, there is increased interest in combining vision, language, and decision-making. Specifically, the Vision and Language Navigation (VLN) task involves navigating to a goal purely from language instructions and visual information without explicit knowledge of the goal. Recent successful approaches have made in-roads in achieving good success rates for this task but rely on beam search, which thoroughly explores a large number of trajectories and is unrealistic for applications such as robotics. In this paper, inspired by the intuition of viewing the problem as search on a navigation graph, we propose to use a progress monitor developed in prior work as a learnable heuristic for search. We then propose two modules incorporated into an end-to-end architecture: 1) A learned mechanism to perform backtracking, which decides whether to continue moving forward or roll back to a previous state (Regret Module) and 2) A mechanism to help the agent decide which direction to go next by showing directions that are visited and their associated progress estimate (Progress Marker). Combined, the proposed approach significantly outperforms current state-of-the-art methods using greedy action selection, with 5% absolute improvement on the test server in success rates, and more importantly 8% on success rates normalized by the path length.

We present a new model to determine relative skill from long videos, through learnable temporal attention modules. Skill determination is formulated as a ranking problem, making it suitable for common and generic tasks. However, for long videos, parts of the video are irrelevant for assessing skill, and there may be variability in the skill exhibited throughout a video. We therefore propose a method which assesses the relative overall level of skill in a long video by attending to its skill-relevant parts. Our approach trains temporal attention modules, learned with only video-level supervision, using a novel rank-aware loss function. In addition to attending to task-relevant video parts, our proposed loss jointly trains two attention modules to separately attend to video parts which are indicative of higher (pros) and lower (cons) skill. We evaluate our approach on the EPIC-Skills dataset and additionally annotate a larger dataset from YouTube videos for skill determination with five previously unexplored tasks. Our method outperforms previous approaches and classic softmax attention on both datasets by over 4% pairwise accuracy, and as much as 12% on individual tasks. We also demonstrate our model's ability to attend to rank-aware parts of the video.

We propose 3DSmoothNet, a full workflow to match 3D point clouds with a siamese deep learning architecture and fully convolutional layers using a voxelized smoothed density value (SDV) representation. The latter is computed per interest point and aligned to the local reference frame (LRF) to achieve rotation invariance. Our compact, learned, rotation invariant 3D point cloud descriptor achieves 94.9% average recall on the 3DMatch benchmark data set, outperforming the state-of-the-art by more than 20 percent points with only 32 output dimensions. This very low output dimension allows for near realtime correspondence search with 0.1 ms per feature point on a standard PC. Our approach is sensor- and scene-agnostic because of SDV, LRF and learning highly descriptive features with fully convolutional layers. We show that 3DSmoothNet trained only on RGB-D indoor scenes of buildings achieves 79.0% average recall on laser scans of outdoor vegetation, more than double the performance of our closest, learning-based competitors. Code, data and pre-trained models are available online at https://github.com/zgojcic/3DSmoothNet.

We present a novel framework for edge-aware optimization that is an order of magnitude faster than the state of the art while maintaining comparable results. Our key insight is that the optimization can be formulated by leveraging properties of the domain transform, a method for edge-aware filtering that defines a distance-preserving 1D mapping of the input space. This enables our method to improve performance for a wide variety of problems including stereo, depth super-resolution, render from defocus, colorization, and especially high-resolution depth filtering, while keeping the computational complexity linear in the number of pixels. Our method is highly parallelizable and adaptable, and it has demonstrable linear scalability with respect to image resolutions. We provide a comprehensive evaluation of our method w.r.t speed and accuracy for a variety of tasks.

Determining the position and orientation of a calibrated camera from a single image with respect to a 3D model is an essential task for many applications. When 2D-3D correspondences can be obtained reliably, perspective-n-point solvers can be used to recover the camera pose. However, without the pose it is non-trivial to find cross-modality correspondences between 2D images and 3D models, particularly when the latter only contains geometric information. Consequently, the problem becomes one of estimating pose and correspondences jointly. Since outliers and local optima are so prevalent, robust objective functions and global search strategies are desirable. Hence, we cast the problem as a 2D-3D mixture model alignment task and propose the first globally-optimal solution to this formulation under the robust L2 distance between mixture distributions. We derive novel bounds on this objective function and employ branch-and-bound to search the 6D space of camera poses, guaranteeing global optimality without requiring a pose estimate. To accelerate convergence, we integrate local optimization, implement GPU bound computations, and provide an intuitive way to incorporate side information such as semantic labels. The algorithm is evaluated on challenging synthetic and real datasets, outperforming existing approaches and reliably converging to the global optimum.

We present a system for learning full body neural avatars, i.e. deep networks that produce full body renderings of a person for varying body pose and varying camera pose. Our system takes the middle path between the classical graphics pipeline and the recent deep learning approaches that generate images of humans using image-to-image translation. In particular, our system estimates an explicit two-dimensional texture map of the model surface. At the same time, it abstains from explicit shape modeling in 3D. Instead, at test time, the system uses a fully-convolutional network to directly map the configuration of body feature points w.r.t. the camera to the 2D texture coordinates of individual pixels in the image frame. We show that such system is capable of learning to generate realistic renderings while being trained on videos annotated with 3D poses and foreground masks. We also demonstrate that maintaining an explicit texture representation helps our system to achieve better generalization compared to systems that use direct image-to-image translation.

We introduce, TextureNet, a neural network architecture designed to extract features from high-resolution signals associated with 3D surface meshes (e.g., color texture maps). The key idea is to utilize a 4-rotational symmetric(4-RoSy) field to define a domain for convolution on a surface. Though 4-RoSy fields have several properties favor-able for convolution on surfaces (low distortion, few singularities, consistent parameterization, etc.), orientations are ambiguous up to 4-fold rotation at any sample point. So, we introduce a new convolutional operator invariant to the4-RoSy ambiguity and use it in a network to extract features from high-resolution signals on geodesic neighborhoods of a surface. In comparison to alternatives, such as PointNet-based methods which lack a notion of orientation, the coherent structure given by these neighborhoods results in significantly stronger features. As an example application, we demonstrate the benefits of our architecture for 3D semantic segmentation of textured 3D meshes. The results show that our method outperforms all existing methods on the basis of mean IoU by a significant margin in both geometry-only(6.4%) and RGB+Geometry (6.9-8.2%) settings.

This paper addresses the problem of interpolating visual textures. We formulate this problem by requiring (1) by-example controllability and (2) realistic and smooth interpolation among an arbitrary number of texture samples. To solve it we propose a neural network trained simultaneously on a reconstruction task and a generation task, which can project texture examples onto a latent space where they can be linearly interpolated and projected back onto the image domain, thus ensuring both intuitive control and realistic results. We show our method outperforms a number of baselines according to a comprehensive suite of metrics as well as a user study. We further show several applications based on our technique, which include texture brush, texture dissolve, and animal hybridization.

In this paper, we propose Text2Scene, a model that generates various forms of compositional scene representations from natural language descriptions. Unlike recent works, our method does NOT use Generative Adversarial Networks (GANs). Text2Scene instead learns to sequentially generate objects and their attributes (location, size, appearance, etc) at every time step by attending to different parts of the input text and the current status of the generated scene. We show that under minor modifications, the proposed framework can handle the generation of different forms of scene representations, including cartoon-like scenes, object layouts corresponding to real images, and synthetic images. Our method is not only competitive when compared with state-of-the-art GAN-based methods using automatic metrics and superior based on human judgments but also has the advantage of producing interpretable results.