CVPR 2020 Papers

Pruning convolutional filters has demonstrated its effectiveness in compressing ConvNets. Prior art in filter pruning requires users to specify a target model complexity (e.g., model size or FLOP count) for the resulting architecture. However, determining a target model complexity can be difficult for optimizing various embodied AI applications such as autonomous robots, drones, and user-facing applications. First, both the accuracy and the speed of ConvNets can affect the performance of the application. Second, the performance of the application can be hard to assess without evaluating ConvNets during inference. As a consequence, finding a sweet-spot between the accuracy and speed via filter pruning, which needs to be done in a trial-and-error fashion, can be time-consuming. This work takes a first step toward making this process more efficient by altering the goal of model compression to producing a set of ConvNets with various accuracy and latency trade-offs instead of producing one ConvNet targeting some pre-defined latency constraint. To this end, we propose to learn a global ranking of the filters across different layers of the ConvNet, which is used to obtain a set of ConvNet architectures that have different accuracy/latency trade-offs by pruning the bottom-ranked filters. Our proposed algorithm, LeGR, is shown to be 2x to 3x faster than prior work while having comparable or better performance when targeting seven pruned ResNet-56 with different accuracy/FLOPs profiles on the CIFAR-100 dataset. Additionally, we have evaluated LeGR on ImageNet and Bird-200 with ResNet-50 and Mo- bileNetV2 to demonstrate its effectiveness. Code available at https://github.com/cmu-enyac/LeGR.

The learning of the deep networks largely relies on the data with human-annotated labels. In some label insufficient situations, the performance degrades on the decision boundary with high data density. A common solution is to directly minimize the Shannon Entropy, but the side effect caused by entropy minimization, i.e., reduction of the prediction diversity, is mostly ignored. To address this issue, we reinvestigate the structure of classification output matrix of a randomly selected data batch. We find by theoretical analysis that the prediction discriminability and diversity could be separately measured by the Frobenius-norm and rank of the batch output matrix. Besides, the nuclear-norm is an upperbound of the Frobenius-norm, and a convex approximation of the matrix rank. Accordingly, to improve both discriminability and diversity, we propose Batch Nuclear-norm Maximization (BNM) on the output matrix. BNM could boost the learning under typical label insufficient learning scenarios, such as semi-supervised learning, domain adaptation and open domain recognition. On these tasks, extensive experimental results show that BNM outperforms competitors and works well with existing well-known methods. The code is available at https://github.com/cuishuhao/BNM

We propose a way to learn visual features that are compatible with previously computed ones even when they have different dimensions and are learned via different neural network architectures and loss functions. Compatible means that, if such features are used to compare images, then "new" features can be compared directly to "old" features, so they can be used interchangeably. This enables visual search systems to bypass computing new features for all previously seen images when updating the embedding models, a process known as backfilling. Backward compatibility is critical to quickly deploy new embedding models that leverage ever-growing large-scale training datasets and improvements in deep learning architectures and training methods. We propose a framework to train embedding models, called backward-compatible training (BCT), as a first step towards backward compatible representation learning. In experiments on learning embeddings for face recognition, models trained with BCT successfully achieve backward compatibility without sacrificing accuracy, thus enabling backfill-free model updates of visual embeddings.

Semantic reconstruction of indoor scenes refers to both scene understanding and object reconstruction. Existing works either address one part of this problem or focus on independent objects. In this paper, we bridge the gap between understanding and reconstruction, and propose an end-to-end solution to jointly reconstruct room layout, object bounding boxes and meshes from a single image. Instead of separately resolving scene understanding and object reconstruction, our method builds upon a holistic scene context and proposes a coarse-to-fine hierarchy with three components: 1. room layout with camera pose; 2. 3D object bounding boxes; 3. object meshes. We argue that understanding the context of each component can assist the task of parsing the others, which enables joint understanding and reconstruction. The experiments on the SUN RGB-D and Pix3D datasets demonstrate that our method consistently outperforms existing methods in indoor layout estimation, 3D object detection and mesh reconstruction.

Diverse inverse problems in imaging can be cast as variational problems composed of a task-specific data fidelity term and a regularization term. In this paper, we propose a novel learnable general-purpose regularizer exploiting recent architectural design patterns from deep learning. We cast the learning problem as a discrete sampled optimal control problem, for which we derive the adjoint state equations and an optimality condition. By exploiting the variational structure of our approach, we perform a sensitivity analysis with respect to the learned parameters obtained from different training datasets. Moreover, we carry out a nonlinear eigenfunction analysis, which reveals interesting properties of the learned regularizer. We show state-of-the-art performance for classical image restoration and medical image reconstruction problems.

Recovering the 3D shape of transparent objects using a small number of unconstrained natural images is an ill-posed problem. Complex light paths induced by refraction and reflection have prevented both traditional and deep multiview stereo from solving this challenge. We propose a physically-based network to recover 3D shape of transparent objects using a few images acquired with a mobile phone camera, under a known but arbitrary environment map. Our novel contributions include a normal representation that enables the network to model complex light transport through local computation, a rendering layer that models refractions and reflections, a cost volume specifically designed for normal refinement of transparent shapes and a feature mapping based on predicted normals for 3D point cloud reconstruction. We render a synthetic dataset to encourage the model to learn refractive light transport across different views. Our experiments show successful recovery of high-quality 3D geometry for complex transparent shapes using as few as 5-12 natural images. Code and data will be publicly released.Recovering the 3D shape of transparent objects using a small number of unconstrained natural images is an ill-posed problem. Complex light paths induced by refraction and reflection have prevented both traditional and deep multiview stereo from solving this challenge. We propose a physically-based network to recover 3D shape of transparent objects using a few images acquired with a mobile phone camera, under a known but arbitrary environment map. Our novel contributions include a normal representation that enables the network to model complex light transport through local computation, a rendering layer that models refractions and reflections, a cost volume specifically designed for normal refinement of transparent shapes and a feature mapping based on predicted normals for 3D point cloud reconstruction. We render a synthetic dataset to encourage the model to learn refractive light transport across different views. Our experiments show successful recovery of high-quality 3D geometry for complex transparent shapes using as few as 5-12 natural images. Code and data will be publicly released.

This paper studies model-inversion attacks, in which the access to a model is abused to infer information about the training data. Since its first introduction by [??], such attacks have raised serious concerns given that training data usually contain privacy sensitive information. Thus far, successful model-inversion attacks have only been demonstrated on simple models, such as linear regression and logistic regression. Previous attempts to invert neural networks, even the ones with simple architectures, have failed to produce convincing results. Here we present a novel attack method, termed the generative model-inversion attack, which can invert deep neural networks with high success rates. Rather than reconstructing private training data from scratch, we leverage partial public information, which can be very generic, to learn a distributional prior via generative adversarial networks (GANs) and use it to guide the inversion process. Moreover, we theoretically prove that a model's predictive power and its vulnerability to inversion attacks are indeed two sides of the same coin---highly predictive models are able to establish a strong correlation between features and labels, which coincides exactly with what an adversary exploits to mount the attacks. Our extensive experiments demonstrate that the proposed attack improves identification accuracy over the existing work by about 75% for reconstructing face images from a state-of-the-art face recognition classifier. We also show that differential privacy, in its canonical form, is of little avail to defend against our attacks.

Real-time RGB-D scanning technique has become widely used to progressively scan objects with a hand-held sensor. Existing online methods restore color information per voxel, and thus their quality is often limited by the tradeoff between spatial resolution and time performance. Also, such methods often suffer from blurred artifacts in the captured texture. Traditional offline texture mapping methods with non-rigid warping assume that the reconstructed geometry and all input views are obtained in advance, and the optimization takes a long time to compute mesh parameterization and warp parameters, which prevents them from being used in real-time applications. In this work, we propose a progressive texture-fusion method specially designed for real-time RGB-D scanning. To this end, we first devise a novel texture-tile voxel grid, where texture tiles are embedded in the voxel grid of the signed distance function, allowing for high-resolution texture mapping on the low-resolution geometry volume. Instead of using expensive mesh parameterization, we associate vertices of implicit geometry directly with texture coordinates. Second, we introduce real-time texture warping that applies a spatially-varying perspective mapping to input images so that texture warping efficiently mitigates the mismatch between the intermediate geometry and the current input view. It allows us to enhance the quality of texture over time while updating the geometry in real-time. The results demonstrate that the quality of our real-time texture mapping is highly competitive to that of exhaustive offline texture warping methods. Our method is also capable of being integrated into existing RGB-D scanning frameworks.

Self-supervised audio-visual learning aims to capture useful representations of video by leveraging correspondences between visual and audio inputs. Existing approaches have focused primarily on matching semantic information between the sensory streams. We propose a novel self-supervised task to leverage an orthogonal principle: matching spatial information in the audio stream to the positions of sound sources in the visual stream. Our approach is simple yet effective. We train a model to determine whether the left and right audio channels have been flipped, forcing it to reason about spatial localization across the visual and audio streams. To train and evaluate our method, we introduce a large-scale video dataset, YouTube-ASMR-300K, with spatial audio comprising over 900 hours of footage. We demonstrate that understanding spatial correspondence enables models to perform better on three audio-visual tasks, achieving quantitative gains over supervised and self-supervised baselines that do not leverage spatial audio cues. We also show how to extend our self-supervised approach to 360 degree videos with ambisonic audio.

In this paper, we focus on weakly supervised learning with noisy training data for both classification and regression problems. We assume that the training outputs are collected from a mixture of a target and correlated noise distributions. Our proposed method simultaneously estimates the target distribution and the quality of each data which is defined as the correlation between the target and data generating distributions. The cornerstone of the proposed method is a Cholesky Block that enables modeling dependencies among mixture distributions in a differentiable manner where we maintain the distribution over the network weights. We first provide illustrative examples in both regression and classification tasks to show the effectiveness of the proposed method. Then, the proposed method is extensively evaluated in a number of experiments where we show that it constantly shows comparable or superior performances compared to existing baseline methods in the handling of noisy data.

In this paper, we present TailorNet, a neural model which predicts clothing deformation in 3D as a function of three factors: pose, shape and style (garment geometry), while retaining wrinkle detail. This goes beyond prior models, which are either specific to one style and shape, or generalize to different shapes producing smooth results, despite being style specific. Our hypothesis is that (even non-linear) combinations of examples smoothes out high frequency components such as fine-wrinkles, which makes learning the three factors jointly hard. At the heart of our technique is a decomposition of deformation into a high frequency and a low frequency component. While the low-frequency component is predicted from pose, shape and style parameters with an MLP, the high-frequency component is predicted with a mixture of shape-style specific pose models. The weights of the mixture are computed with a narrow bandwidth kernel to guarantee that only predictions with similar high-frequency patterns are combined. The style variation is obtained by computing, in a canonical pose, a subspace of deformation, which satisfies physical constraints such as inter-penetration, and draping on the body. TailorNet delivers 3D garments which retain the wrinkles from the physics based simulations (PBS) it is learned from, while running more than 1000 times faster. In contrast to classical PBS, TailorNet is easy to use and fully differentiable, which is crucial for computer vision and learning algorithms. Several experiments demonstrate TailorNet produces more realistic results than prior work, and even generates temporally coherent deformations on sequences of the AMASS dataset, despite being trained on static poses from a different dataset. To stimulate further research in this direction, we will make a dataset consisting of 55800 frames, as well as our model publicly available at https://virtualhumans.mpi-inf.mpg.de/tailornet/.

We consider the problem of predicting the future trajectory of scene agents from egocentric views obtained from a moving platform. This problem is important in a variety of domains, particularly for autonomous systems making reactive or strategic decisions in navigation. In an attempt to address this problem, we introduce TITAN (Trajectory Inference using Targeted Action priors Network), a new model that incorporates prior positions, actions, and context to forecast future trajectory of agents and future ego-motion. In the absence of an appropriate dataset for this task, we created the TITAN dataset that consists of 700 labeled video-clips (with odometry) captured from a moving vehicle on highly interactive urban traffic scenes in Tokyo. Our dataset includes 50 labels including vehicle states and actions, pedestrian age groups, and targeted pedestrian action attributes that are organized hierarchically corresponding to atomic, simple/complex-contextual, transportive, and communicative actions. To evaluate our model, we conducted extensive experiments on the TITAN dataset, revealing significant performance improvement against baselines and state-of-the-art algorithms. We also report promising results from our Agent Importance Mechanism (AIM), a module which provides insight into assessment of perceived risk by calculating the relative influence of each agent on the future ego-trajectory. The dataset is available at https://usa.honda-ri.com/titan

There are two main challenges in Visual Question Answering (VQA). The first one is that each model obtains its strengths and shortcomings when applied to several questions; what is more, the "ceiling effect" for specific questions is difficult to overcome with simple consecutive training. The second challenge is that even the state-of-the-art dataset is of large scale, questions targeted at a single image are off in format and lack diversity in content. We introduce our self-questioning model with multi-agent training: TA-student VQA. This framework differs from standard VQA algorithms by involving question-generating mechanisms and collaborative learning questions between question-answering agents. Thus, TA-student VQA overcomes the limitation of the content diversity and format variation of questions and improves the overall performance of multiple question-answering agents. We evaluate our model on VQA-v2, which outperforms algorithms without such mechanisms. In addition, TA-student VQA achieves a greater model capacity, allowing it to answer more generated questions in addition to those in the annotated datasets.

View synthesis allows for the generation of new views of a scene given one or more images. This is challenging; it requires comprehensively understanding the 3D scene from images. As a result, current methods typically use multiple images, train on ground-truth depth, or are limited to synthetic data. We propose a novel end-to-end model for this task using a single image at test time; it is trained on real images without any ground-truth 3D information. To this end, we introduce a novel differentiable point cloud renderer that is used to transform a latent 3D point cloud of features into the target view. The projected features are decoded by our refinement network to inpaint missing regions and generate a realistic output image. The 3D component inside of our generative model allows for interpretable manipulation of the latent feature space at test time, e.g. we can animate trajectories from a single image. Additionally, we can generate high resolution images and generalise to other input resolutions. We outperform baselines and prior work on the Matterport, Replica, and RealEstate10K datasets.

Recent CNN-based methods for image deraining have achieved excellent performance in terms of reconstruction error as well as visual quality. However, these methods are limited in the sense that they can be trained only on fully labeled data. Due to various challenges in obtaining real world fully-labeled image deraining datasets, existing methods are trained only on synthetically generated data and hence, generalize poorly to real-world images. The use of real-world data in training image deraining networks is relatively less explored in the literature. We propose a Gaussian Process-based semi-supervised learning framework which enables the network in learning to derain using synthetic dataset while generalizing better using unlabeled real-world images. Through extensive experiments and ablations on several challenging datasets (such as Rain800, Rain100H and DDN-SIRR), we show that the proposed method, when trained on limited labeled data, achieves on-par performance with fully-labeled training. Additionally, we demonstrate that using unlabeled real-world images in the proposed GP-based framework results in superior performance as compared to existing methods.

Autonomous driving system development is critically dependent on the ability to replay complex and diverse traffic scenarios in simulation. In such scenarios, the ability to accurately simulate the vehicle sensors such as cameras, lidar or radar is hugely helpful. However, current sensor simulators leverage gaming engines such as Unreal or Unity, requiring manual creation of environments, objects, and material properties. Such approaches have limited scalability and fail to produce realistic approximations of camera, lidar, and radar data without significant additional work. In this paper, we present a simple yet effective approach to generate realistic scenario sensor data, based only on a limited amount of lidar and camera data collected by an autonomous vehicle. Our approach uses texture-mapped surfels to efficiently reconstruct the scene from an initial vehicle pass or set of passes, preserving rich information about object 3D geometry and appearance, as well as the scene conditions. We then leverage a SurfelGAN network to reconstruct realistic camera images for novel positions and orientations of the self-driving vehicle and moving objects in the scene. We demonstrate our approach on the Waymo Open Dataset and show that it can synthesize realistic camera data for simulated scenarios. We also create a novel dataset that contains cases in which two self-driving vehicles observe the same scene at the same time. We use this dataset to provide additional evaluation and demonstrate the usefulness of our SurfelGAN model.

This paper introduces SuperGlue, a neural network that matches two sets of local features by jointly finding correspondences and rejecting non-matchable points. Assignments are estimated by solving a differentiable optimal transport problem, whose costs are predicted by a graph neural network. We introduce a flexible context aggregation mechanism based on attention, enabling SuperGlue to reason about the underlying 3D scene and feature assignments jointly. Compared to traditional, hand-designed heuristics, our technique learns priors over geometric transformations and regularities of the 3D world through end-to-end training from image pairs. SuperGlue outperforms other learned approaches and achieves state-of-the-art results on the task of pose estimation in challenging real-world indoor and outdoor environments. The proposed method performs matching in real-time on a modern GPU and can be readily integrated into modern SfM or SLAM systems. The code and trained weights are publicly available at github.com/magicleap/SuperGluePretrainedNetwork.

StyleGAN generates photorealistic portrait images of faces with eyes, teeth, hair and context (neck, shoulders, background), but lacks a rig-like control over semantic face parameters that are interpretable in 3D, such as face pose, expressions, and scene illumination. Three-dimensional morphable face models (3DMMs) on the other hand offer control over the semantic parameters, but lack photorealism when rendered and only model the face interior, not other parts of a portrait image (hair, mouth interior, background). We present the first method to provide a face rig-like control over a pretrained and fixed StyleGAN via a 3DMM. A new rigging network, RigNet is trained between the 3DMM's semantic parameters and StyleGAN's input. The network is trained in a self-supervised manner, without the need for manual annotations. At test time, our method generates portrait images with the photorealism of StyleGAN and provides explicit control over the 3D semantic parameters of the face.

We introduce a novel principle for self-supervised feature learning based on the discrimination of specific transformations of an image. We argue that the generalization capability of learned features depends on what image neighborhood size is sufficient to discriminate different image transformations: The larger the required neighborhood size and the more global the image statistics that the feature can describe. An accurate description of global image statistics allows to better represent the shape and configuration of objects and their context, which ultimately generalizes better to new tasks such as object classification and detection. This suggests a criterion to choose and design image transformations. Based on this criterion, we introduce a novel image transformation that we call limited context inpainting (LCI). This transformation inpaints an image patch conditioned only on a small rectangular pixel boundary (the limited context). Because of the limited boundary information, the inpainter can learn to match local pixel statistics, but is unlikely to match the global statistics of the image. We claim that the same principle can be used to justify the performance of transformations such as image rotations and warping. Indeed, we demonstrate experimentally that learning to discriminate transformations such as LCI, image warping and rotations, yields features with state of the art generalization capabilities on several datasets such as Pascal VOC, STL-10, CelebA, and ImageNet. Remarkably, our trained features achieve a performance on Places on par with features trained through supervised learning with ImageNet labels.