ICCV 2021 论文列表

We present EgoRenderer, a system for rendering full-body neural avatars of a person captured by a wearable, egocentric fisheye camera that is mounted on a cap or a VR headset. Our system renders photorealistic novel views of the actor and her motion from arbitrary virtual camera locations. Rendering full-body avatars from such egocentric images come with unique challenges due to the top-down view and large distortions. We tackle these challenges by decomposing the rendering process into several steps, including texture synthesis, pose construction, and neural image translation. For texture synthesis, we propose Ego-DPNet, a neural network that infers dense correspondences between the input fisheye images and an underlying parametric body model, and to extract textures from egocentric inputs. In addition, to encode dynamic appearances, our approach also learns an implicit texture stack that captures detailed appearance variation across poses and viewpoints. For correct pose generation, we first estimate body pose from the egocentric view using a parametric model. We then synthesize an external free-viewpoint pose image by projecting the parametric model to the user-specified target viewpoint. We next combine the target pose image and the textures into a combined feature image, which is transformed into the output color image using a neural image translation network. Experimental evaluations show that EgoRenderer is capable of generating realistic free-viewpoint avatars of a person wearing an egocentric camera. Comparisons to several baselines demonstrate the advantages of our approach.

Differentiable rendering has received increasing interest in the solution of image-based inverse problems. It can benefit traditional optimization-based solutions to inverse problems, but also allows for self-supervision of learning-based approaches for which training data with ground truth annotation is hard to obtain. However, existing differentiable renderers either do not correctly model complex visibility responsible for shadows in the images, or are too slow for being used to train deep architectures over thousands of iterations. To this end, we propose an accurate yet efficient approach for differentiable visibility and soft shadow computation. Our approach is based on the spherical harmonics approximation of the scene illumination and visibility, where the occluding surface is approximated with spheres. This allows for significantly more efficient visibility computation compared to methods based on path tracing without sacrificing quality of generated images. As our formulation is differentiable, it can be used to solve various image-based inverse problems such as texture, lighting, geometry recovery from images using analysis-by-synthesis optimization.

Self-supervised pretraining has been shown to yield powerful representations for transfer learning. These performance gains come at a large computational cost however, with state-of-the-art methods requiring an order of magnitude more computation than supervised pretraining. We tackle this computational bottleneck by introducing a new self-supervised objective, contrastive detection, which tasks representations with identifying object-level features across augmentations. This objective extracts a rich learning signal per image, leading to state-of-the-art transfer accuracy on a variety of downstream tasks, while requiring up to 10x less pretraining. In particular, our strongest ImageNet-pretrained model performs on par with SEER, one of the largest self-supervised systems to date, which uses 1000x more pretraining data. Finally, our objective seamlessly handles pretraining on more complex images such as those in COCO, closing the gap with supervised transfer learning from COCO to PASCAL.

Video compression is a critical component of Internet video delivery. Recent work has shown that deep learning techniques can rival or outperform human-designed algorithms, but these methods are significantly less compute and power-efficient than existing codecs. This paper presents a new approach that augments existing codecs with a small, content-adaptive super-resolution model that significantly boosts video quality. Our method, SRVC, encodes video into two bitstreams: (i) a content stream, produced by compressing downsampled low-resolution video with the existing codec, (ii) a model stream, which encodes periodic updates to a lightweight super-resolution neural network customized for short segments of the video. SRVC decodes the video by passing the decompressed low-resolution video frames through the (time-varying) super-resolution model to reconstruct high-resolution video frames. Our results show that to achieve the same PSNR, SRVC requires 20% of the bits-per-pixel of H.265 in slow mode, and 3% of the bits-per-pixel of DVC, a recent deep learning-based video compression scheme. SRVC runs at 90 frames per second on an NVIDIA V100 GPU.

Outlier rejection and equivalently inlier set optimization is a key ingredient in numerous applications in computer vision such as filtering point-matches in camera pose estimation or plane and normal estimation in point clouds. Several approaches exist, yet at large scale we face a combinatorial explosion of possible solutions and state-of-the-art methods like RANSAC, Hough transform or Branch&Bound require a minimum inlier ratio or prior knowledge to remain practical. In fact, for problems such as camera posing in very large scenes these approaches become useless as they have exponential runtime growth if these conditions aren't met. To approach the problem we present a efficient and general algorithm for outlier rejection based on "intersecting" k-dimensional surfaces in Rd . We provide a recipe for casting a variety of geometric problems as finding a point in Rd which maximizes the number of nearby surfaces (and thus inliers). The resulting algorithm has linear worst-case complexity with a better runtime dependency in the approximation factor than competing algorithms while not requiring domain specific bounds. This is achieved by introducing a space decomposition scheme that bounds the number of computations by successively rounding and grouping samples. Our recipe (and open-source code) enables anybody to derive such fast approaches to new problems across a wide range of domains. We demonstrate the versatility of the approach on several camera posing problems with a high number of matches at low inlier ratio achieving state-of-the-art results at significantly lower processing times.

Efficient action recognition has become crucial to extend the success of action recognition to many real-world applications. Contrary to most existing methods, which mainly focus on selecting salient frames to reduce the computation cost, we focus more on making the most of the selected frames. To this end, we employ two networks of different capabilities that operate in tandem to efficiently recognize actions. Given a video, the lighter network processes more frames while the heavier one only processes a few. In order to enable the effective interaction between the two, we propose dynamic knowledge propagation based on a cross-attention mechanism. This is the main component of our framework that is essentially a student-teacher architecture, but as the teacher model continues to interact with the student model during inference, we call it a dynamic student-teacher framework. Through extensive experiments, we demonstrate the effectiveness of each component of our framework. Our method outperforms competing state-of-the-art methods on two video datasets: ActivityNet-v1.3 and Mini-Kinetics.

Measuring similarity between two images often requires performing complex reasoning along different axes (e.g., color, texture, or shape). Insights into what might be important for measuring similarity can can be provided by annotated attributes, but prior work tends to view these annotations as complete, resulting in them using a simplistic approach of predicting attributes on single images, which are, in turn, used to measure similarity. However, it is impractical for a dataset to fully annotate every attribute that may be important. Thus, only representing images based on these incomplete annotations may miss out on key information. To address this issue, we propose the Pairwise Attribute-informed similarity Network (PAN), which breaks similarity learning into capturing similarity conditions and relevance scores from a joint representation of two images. This enables our model to identify that two images contain the same attribute, but can have it deemed irrelevant (e.g., due to fine-grained differences between them) and ignored for measuring similarity between the two images. Notably, while prior methods of using attribute annotations are often unable to outperform prior art, PAN obtains a 4-9% improvement on compatibility prediction between clothing items on Polyvore Outfits, a 5% gain on few shot classification of images using Caltech-UCSD Birds (CUB), and over 1% boost to Recall@1 on In-Shop Clothes Retrieval.

A neural radiance field (NeRF) is a scene model supporting high-quality view synthesis, optimized per scene. In this paper, we explore enabling user editing of a category-level NeRF trained on a shape category. Specifically, we propose a method for propagating coarse 2D user scribbles to the 3D space, to modify the color or shape of a local region. First, we propose a conditional radiance field that incorporates new modular network components, including a branch that is shared across object instances in the category. Observing multiple instances of the same category, our model learns underlying part semantics without any supervision, thereby allowing the propagation of coarse 2D user scribbles to the entire 3D region (e.g., chair seat) in a consistent fashion. Next, we investigate for the editing tasks which components of our network require updating. We propose a hybrid network update strategy that targets the later network components, which balances efficiency and accuracy. During user interaction, we formulate an optimization problem that both satisfies the user's constraints and preserves the original object structure. We demonstrate our approach on a variety of editing tasks over three shape datasets and show that it outperforms prior neural editing approaches. Finally, we edit the appearance and shape of a real photograph and show that the edit propagates to extrapolated novel views.

In this paper, we proposed EPP-MVSNet, a novel deep learning network for 3D reconstruction from multi-view stereo (MVS). EPP-MVSNet can accurately aggregate features at high resolution to a limited cost volume with an optimal depth range, thus, leads to effective and efficient 3D construction. Distinct from existing works which measure feature cost at discrete positions which affects the 3D reconstruction accuracy, EPP-MVSNet introduces an epipolar assembling-based kernel that operates on adaptive intervals along epipolar lines for making full use of the image resolution. Further, we introduce an entropy-based refining strategy where the cost volume describes the space geometry with the little redundancy. Moreover, we design a light-weighted network with Pseudo-3D convolutions integrated to achieve high accuracy and efficiency. We have conducted extensive experiments on challenging datasets Tanks & Temples(TNT), ETH3D and DTU. As a result, we achieve promising results on all datasets and the highest F-Score on the online TNT intermediate benchmark. Code is available at https://gitee.com/mindspore/mindspore/tree/master/model_zoo/research/cv/eppmvsnet.

Fully immersive experiences in AR/VR depend on reconstructing the full body pose of the user without restricting their motion. In this paper we study the use of body-worn electromagnetic (EM) field-based sensing for the task of 3D human pose reconstruction. To this end, we present a method to estimate SMPL parameters from 6-12 EM sensors. We leverage a customized wearable system consisting of wireless EM sensors measuring time-synchronized 6D poses at 120 Hz. To provide accurate poses even with little user instrumentation, we adopt a recently proposed hybrid framework, learned gradient descent (LGD), to iteratively estimate SMPL pose and shape from our input measurements. This allows us to harness powerful pose priors to cope with the idiosyncrasies of the input data and achieve accurate pose estimates. The proposed method uses AMASS to synthesize virtual EM-sensor data and we show that it generalizes well to a newly captured real dataset consisting of a total of 36 minutes of motion from 5 subjects. We achieve reconstruction errors as low as 31.8 mm and 13.3 degrees, outperforming both pure learning- and pure optimization-based methods. Code and data is available under https://ait.ethz.ch/projects/2021/em-pose.

We present the novel Efficient Line Segment Detector and Descriptor (ELSD) to simultaneously detect line segments and extract their descriptors in an image. Unlike the traditional pipelines that conduct detection and description separately, ELSD utilizes a shared feature extractor for both detection and description, to provide the essential line features to the higher-level tasks like SLAM and image matching in real time. First, we design a one-stage compact model, and propose to use the mid-point, angle and length as the minimal representation of line segment, which also guarantees the center-symmetry. The non-centerness suppression is proposed to filter out the fragmented line segments caused by lines' intersections. The fine offset prediction is designed to refine the mid-point localization. Second, the line descriptor branch is integrated with the detector branch, and the two branches are jointly trained in an end-to-end manner. In the experiments, the proposed ELSD achieves the state-of-the-art performance on the Wireframe dataset and YorkUrban dataset, in both accuracy and efficiency. The line description ability of ELSD also outperforms the previous works on the line matching task.

Autonomous systems need to understand the semantics and geometry of their surroundings in order to comprehend and safely execute object-level task specifications. This paper proposes an expressive yet compact model for joint object pose and shape optimization, and an associated optimization algorithm to infer an object-level map from multi-view RGB-D camera observations. The model is expressive because it captures the identities, positions, orientations, and shapes of objects in the environment. It is compact because it relies on a low-dimensional latent representation of implicit object shape, allowing onboard storage of large multi-category object maps. Different from other works that rely on a single object representation format, our approach has a bi-level object model that captures both the coarse level scale as well as the fine level shape details. Our approach is evaluated on the large-scale real-world ScanNet dataset and compared against state-of-the-art methods.

While learned video codecs have demonstrated great promise, they have yet to achieve sufficient efficiency for practical deployment. In this work, we propose several ideas for learned video compression which allow for improved performance for the low-latency mode (I- and P-frames only) along with a considerable increase in computational efficiency. In this setting, for natural videos our approach compares favorably across the entire R-D curve under metrics PSNR, MS-SSIM and VMAF against all mainstream video standards (H.264, H.265, AV1) and all ML codecs. At the same time, our approach runs at least 5x faster and has fewer parameters than all ML codecs which report these figures. Our contributions include a flexible-rate framework allowing a single model to cover a large and dense range of bitrates, at a negligible increase in computation and parameter count; an efficient backbone optimized for ML-based codecs; and a novel in-loop flow prediction scheme which leverages prior information towards more efficient compression. We benchmark our method, which we call ELF-VC (Efficient, Learned and Flexible Video Coding) on popular video test sets UVG and MCL-JCV under metrics PSNR, MS-SSIM and VMAF. For example, on UVG under PSNR, it reduces the BD-rate by 44% against H.264, 26% against H.265, 15% against AV1, 35% against the current best ML codec. At the same time, on an NVIDIA Titan V GPU our approach encodes/decodes VGA at 49/91 FPS, HD 720 at 19/35 FPS, and HD 1080 at 10/18 FPS.

Image-level weakly supervised semantic segmentation is a challenging task. As classification networks tend to capture notable object features and are insensitive to overactivation, class activation map (CAM) is too sparse and rough to guide segmentation network training. Inspired by the fact that erasing distinguishing features force networks to collect new ones from non-discriminative object regions, we using relationships between CAMs to propose a novel weakly supervised method. In this work, we apply these features, learned from erased images, as segmentation supervision, driving network to study robust representation. In specifically, object regions obtained by CAM techniques are erased on images firstly. To provide other regions with segmentation supervision, Erased CAM Supervision Net (ECSNet) generates pixel-level labels by predicting segmentation results of those processed images. We also design the rule of suppressing noise to select reliable labels. Our experiments on PASCAL VOC 2012 dataset show that without data annotations except for ground truth image-level labels, our ECS-Net achieves 67.6% mIoU on test set and 66.6% mIoU on val set, outperforming previous state-of-the-art methods.

This paper studies Semi-Supervised Domain Adaptation (SSDA), a practical yet under-investigated research topic that aims to learn a model of good performance using unlabeled samples and a few labeled samples in the target domain, with the help of labeled samples from a source domain. Several SSDA methods have been proposed recently, which however fail to fully exploit the value of the few labeled target samples. In this paper, we propose Enhanced Categorical Alignment and Consistency Learning (ECACL), a holistic SSDA framework that incorporates multiple mutually complementary domain alignment techniques. ECACL includes two categorical domain alignment techniques that achieve class-level alignment, a strong data augmentation based technique that enhances the model's generalizability and a consistency learning based technique that forces the model to be robust with image perturbations. These techniques are applied on one or multiple of the three inputs (labeled source, unlabeled target, and labeled target) and align the domains from different perspectives. ECACL unifies them together and achieves fairly comprehensive domain alignments that are much better than the existing methods: For example, ECACL raises the state-of-the-art accuracy from 68.4 to 81.1 on VisDA2017 and from 45.5 to 53.4 on DomainNet for the 1-shot setting. Our code is available at https://github.com/kailigo/pacl.

Based on the relaxed search space, differential architecture search (DARTS) is efficient in searching for a high-performance architecture. However, the unbalanced competition among operations that have different trainable parameters causes the model collapse. Besides, the inconsistent structures in the search and retraining stages causes cross-stage evaluation to be unstable. In this paper, we call these issues as an operation gap and a structure gap in DARTS. To shrink these gaps, we propose to induce equalized and consistent optimization in differentiable architecture search (EC-DARTS). EC-DARTS decouples different operations based on their categories to optimize the operation weights so that the operation gap between them is shrinked. Besides, we introduce an induced structural transition to bridge the structure gap between the model structures in the search and retraining stages. Extensive experiments on CIFAR10 and ImageNet demonstrate the effectiveness of our method. Specifically, on CIFAR10, we achieve a test error of 2.39%, while only 0.3 GPU days on NVIDIA TITAN V. On ImageNet, our method achieves a top-1 error of 23.6% under the mobile setting.

Recently, there has been an increasing number of efforts to introduce models capable of generating natural language explanations (NLEs) for their predictions on vision-language (VL) tasks. Such models are appealing, because they can provide human-friendly and comprehensive explanations. However, there is a lack of comparison between existing methods, which is due to a lack of re-usable evaluation frameworks and a scarcity of datasets. In this work, we introduce e-ViL and e-SNLI-VE. e-ViL is a benchmark for explainable vision-language tasks that establishes a unified evaluation framework and provides the first comprehensive comparison of existing approaches that generate NLEs for VL tasks. It spans four models and three datasets and both automatic metrics and human evaluation are used to assess model-generated explanations. e-SNLI-VE is currently the largest existing VL dataset with NLEs (over 430k instances). We also propose a new model that combines UNITER, which learns joint embeddings of images and text, and GPT-2, a pre-trained language model that is well-suited for text generation. It surpasses the previous state of the art by a large margin across all datasets. Code and data are available here: https://github.com/maximek3/e-ViL.

We study the problem of aligning two sets of 3D geometric primitives given known correspondences. Our first contribution is to show that this primitive alignment framework unifies five perception problems including point cloud registration, primitive (mesh) registration, category-level 3D registration, absolution pose estimation (APE), and category-level APE. Our second contribution is to propose DynAMical Pose estimation (DAMP), the first general and practical algorithm to solve primitive alignment problem by simulating rigid body dynamics arising from virtual springs and damping, where the springs span the shortest distances between corresponding primitives. We evaluate DAMP in simulated and real datasets across all five problems, and demonstrate (i) DAMP always converges to the globally optimal solution in the first three problems with 3D-3D correspondences; (ii) although DAMP sometimes converges to suboptimal solutions in the last two problems with 2D-3D correspondences, using a scheme for escaping local minima, DAMP always succeeds. Our third contribution is to demystify the surprising empirical performance of DAMP and formally prove a global convergence result in the case of point cloud registration by charactering local stability of the equilibrium points of the underlying dynamical system.

We present an algorithm for generating novel views at arbitrary viewpoints and any input time step given a monocular video of a dynamic scene. Our work builds upon recent advances in neural implicit representation and uses continuous and differentiable functions for modeling the time-varying structure and the appearance of the scene. We jointly train a time-invariant static NeRF and a time-varying dynamic NeRF, and learn how to blend the results in an unsupervised manner. However, learning this implicit function from a single video is highly ill-posed (with infinitely many solutions that match the input video). To resolve the ambiguity, we introduce regularization losses to encourage a more physically plausible solution. We show extensive quantitative and qualitative results of dynamic view synthesis from casually captured videos.