CVPR 2022 论文列表

The "Roaring 20s" of visual recognition began with the introduction of Vision Transformers (ViTs), which quickly superseded ConvNets as the state-of-the-art image classification model. A vanilla ViT, on the other hand, faces difficulties when applied to general computer vision tasks such as object detection and semantic segmentation. It is the hierarchical Transformers (e.g., Swin Transformers) that reintroduced several ConvNet priors, making Transformers practically viable as a generic vision backbone and demonstrating remarkable performance on a wide variety of vision tasks. However, the effectiveness of such hybrid approaches is still largely credited to the intrinsic superiority of Transformers, rather than the inherent inductive biases of convolutions. In this work, we reexamine the design spaces and test the limits of what a pure ConvNet can achieve. We gradually "modernize" a standard ResNet toward the design of a vision Transformer, and discover several key components that contribute to the performance difference along the way. The outcome of this exploration is a family of pure ConvNet models dubbed ConvNeXt. Constructed entirely from standard ConvNet modules, ConvNeXts compete favorably with Transformers in terms of accuracy and scalability, achieving 87.8% ImageNet top-1 accuracy and outperforming Swin Transformers on COCO detection and ADE20K segmentation, while maintaining the simplicity and efficiency of standard ConvNets.

Although convolutional neural networks (CNNs) achieve state-of-the-art in image classification, recent works address their unreliable predictions due to their excessive dependence on biased training data. Existing unbiased modeling postulates that the bias in the dataset is obvious to know, but it is actually unsuited for image datasets including countless sensory attributes. To mitigate this issue, we present a new scenario that does not necessitate a predefined bias. Under the observation that CNNs do have multi-variant and unbiased representations in the model, we propose a conservative framework that employs this internal information for unbiased learning. Specifically, this mechanism is implemented via hierarchical features captured along the multiple layers and orthogonal regularization. Extensive evaluations on public benchmarks demonstrate our method is effective for unbiased learning.

Modern GANs excel at generating high-quality and diverse images. However, when transferring the pretrained GANs on small target data (e.g., 10-shot), the generator tends to replicate the training samples. Several methods have been proposed to address this few-shot image generation task, but there is a lack of effort to analyze them under a unified framework. As our first contribution, we propose a framework to analyze existing methods during the adaptation. Our analysis discovers that while some methods have a disproportionate focus on diversity preserving which impedes quality improvement, all methods achieve similar quality after convergence. Therefore, the better methods are those that can slow down diversity degradation. Furthermore, our analysis reveals that there is still plenty of room to further slow down diversity degradation. Informed by our analysis and to slow down diversity degradation of the target generator during adaptation, our second contribution proposes to apply mutual information (MI) maximization to retain the source domain's rich multi-level diversity information in the target domain generator. We propose to perform MI maximization by contrastive loss (CL), leverage the generator and discriminator as two feature encoders to extract different multi-level features for computing CL. We refer to our method as Dual ContrastiveLearning (DCL). Extensive experiments on several public datasets show that, while leading to a slower diversity-degrading generator during adaptation, our proposed DCL brings visually pleasant quality and state-of-the-art quantitative performance.

When coming up with phrases of movement, choreographers all have their habits as they are used to their skilled dance genres. Therefore, they tend to return certain patterns of the dance genres that they are familiar with. What if artificial intelligence could be used to help choreographers blend dance genres by suggesting various dances, and one that matches their choreographic style? Numerous task-specific variants of autoregressive networks have been developed for dance generation. Yet, a serious limitation remains that all existing algorithms can return repeated patterns for a given initial pose sequence, which may be inferior. To mitigate this issue, we propose MNET, a novel and scalable approach that can perform music-conditioned pluralistic dance generation synthesized by multiple dance genres using only a single model. Here, we learn a dance-genre aware latent representation by training a conditional generative adversarial network leveraging Transformer architecture. We conduct extensive experiments on AIST++ along with user studies. Compared to the state-of-the-art methods, our method synthesizes plausible and diverse outputs according to multiple dance genres as well as generates outperforming dance sequences qualitatively and quantitatively.

Recent advances in learning 3D shapes using neural implicit functions have achieved impressive results by breaking the previous barrier of resolution and diversity for varying topologies. However, most of such approaches are limited to closed surfaces as they require the space to be divided into inside and outside. More recent works based on unsigned distance function have been proposed to handle complex geometry containing both the open and closed surfaces. Nonetheless, as their direct outputs are point clouds, robustly obtaining high-quality meshing results from discrete points remains an open question. We present a novel learnable implicit representation, called the three-pole signed distance function (3PSDF), that can represent non-watertight 3D shapes with arbitrary topologies while supporting easy field-to-mesh conversion using the classic Marching Cubes algorithm. The key to our method is the introduction of a new sign, the NULL sign, in addition to the conventional in and out labels. The existence of the null sign could stop the formation of a closed isosurface derived from the bisector of the in/out regions. Further, we propose a dedicated learning framework to effectively learn 3PSDF without worrying about the vanishing gradient due to the null labels. Experimental results show that our approach outperforms the previous state-of-the-art methods in a wide range of benchmarks both quantitatively and qualitatively.

We present 3MASSIV, a multilingual, multimodal and multi-aspect, expertly-annotated dataset of diverse short videos extracted from a social media platform. 3MASSIV comprises of 50k short videos (20 seconds average duration) and 100K unlabeled videos in 11 different languages and captures popular short video trends like pranks, fails, romance, comedy expressed via unique audio-visual formats like self-shot videos, reaction videos, lip-synching, self-sung songs, etc. 3MASSIV presents an opportunity for multimodal and multilingual semantic understanding on these unique videos by annotating them for concepts, affective states, media types, and audio language. We present a thorough analysis of 3MASSIV and highlight the variety and unique aspects of our dataset compared to other contemporary popular datasets with strong baselines. We also show how the social media content in 3MASSIV is dynamic and temporal in nature which can be used for various semantic understanding tasks and cross-lingual analysis.

3D computer vision models are commonly used in security-critical applications such as autonomous driving and surgical robotics. Emerging concerns over the robustness of these models against real-world deformations must be addressed practically and reliably. In this work, we propose 3DeformRS, a method to certify the robustness of point cloud Deep Neural Networks (DNNs) against real-world deformations. We developed 3DeformRS by building upon recent work that generalized Randomized Smoothing (RS) from pixel-intensity perturbations to vector-field deformations. In particular, we specialized RS to certify DNNs against parameterized deformations (e.g. rotation, twisting), while enjoying practical computational costs. We leverage the virtues of 3DeformRS to conduct a comprehensive empirical study on the certified robustness of four representative point cloud DNNs on two datasets and against seven different deformations. Compared to previous approaches for certifying point cloud DNNs, 3DeformRS is fast, scales well with point cloud size, and provides comparable-to-better certificates. For instance, when certifying a plain PointNet against a 3deg z-rotation on 1024-point clouds, 3DeformRS grants a certificate 3x larger and 20x faster than previous work.

We study the problem of attribute compression for large-scale unstructured 3D point clouds. Through an in-depth exploration of the relationships between different encoding steps and different attribute channels, we introduce a deep compression network, termed 3DAC, to explicitly compress the attributes of 3D point clouds and reduce storage usage in this paper. Specifically, the point cloud attributes such as color and reflectance are firstly converted to transform coefficients. We then propose a deep entropy model to model the probabilities of these coefficients by considering information hidden in attribute transforms and previous encoded attributes. Finally, the estimated probabilities are used to further compress these transform coefficients to a final attributes bitstream. Extensive experiments conducted on both indoor and outdoor large-scale open point cloud datasets, including ScanNet and SemanticKITTI, demonstrated the superior compression rates and reconstruction quality of the proposed 3DAC.

As 3D object detection on point clouds relies on the geometrical relationships between the points, non-standard object shapes can hinder a method's detection capability. However, in safety-critical settings, robustness to out-of-domain and long-tail samples is fundamental to circumvent dangerous issues, such as the misdetection of damaged or rare cars. In this work, we substantially improve the generalization of 3D object detectors to out-of-domain data by deforming point clouds during training. We achieve this with 3D-VField: a novel data augmentation method that plausibly deforms objects via vector fields learned in an adversarial fashion. Our approach constrains 3D points to slide along their sensor view rays while neither adding nor removing any of them. The obtained vectors are transferable, sample-independent and preserve shape and occlusions. Despite training only on a standard dataset, such as KITTI, augmenting with our vector fields significantly improves the generalization to differently shaped objects and scenes. Towards this end, we propose and share CrashD: a synthetic dataset of realistic damaged and rare cars, with a variety of crash scenarios. Extensive experiments on KITTI, Waymo, our CrashD and SUN RGB-D show the generalizability of our techniques to out-of-domain data, different models and sensors, namely LiDAR and ToF cameras, for both indoor and outdoor scenes. Our CrashD dataset is available at https://crashd-cars.github.io.

Making generative models 3D-aware bridges the 2D image space and the 3D physical world yet remains challenging. Recent attempts equip a Generative Adversarial Network (GAN) with a Neural Radiance Field (NeRF), which maps 3D coordinates to pixel values, as a 3D prior. However, the implicit function in NeRF has a very local receptive field, making the generator hard to become aware of the global structure. Meanwhile, NeRF is built on volume rendering which can be too costly to produce high-resolution results, increasing the optimization difficulty. To alleviate these two problems, we propose a novel framework, termed as VolumeGAN, for high-fidelity 3D-aware image synthesis, through explicitly learning a structural representation and a textural representation. We first learn a feature volume to represent the underlying structure, which is then converted to a feature field using a NeRF-like model. The feature field is further accumulated into a 2D feature map as the textural representation, followed by a neural renderer for appearance synthesis. Such a design enables independent control of the shape and the appearance. Extensive experiments on a wide range of datasets confirm that, our approach achieves sufficiently higher image quality and better 3D control than the previous methods..

Reconstructing 3D shape from a single 2D image is a challenging task, which needs to estimate the detailed 3D structures based on the semantic attributes from 2D image. So far, most of the previous methods still struggle to extract semantic attributes for 3D reconstruction task. Since the semantic attributes of a single image are usually implicit and entangled with each other, it is still challenging to reconstruct 3D shape with detailed semantic structures represented by the input image. To address this problem, we propose 3DAttriFlow to disentangle and extract semantic attributes through different semantic levels in the input images. These disentangled semantic attributes will be integrated into the 3D shape reconstruction process, which can provide definite guidance to the reconstruction of specific attribute on 3D shape. As a result, the 3D decoder can explicitly capture high-level semantic features at the bottom of the network, and utilize low-level features at the top of the network, which allows to reconstruct more accurate 3D shapes. Note that the explicit disentangling is learned without extra labels, where the only supervision used in our training is the input image and its corresponding 3D shape. Our comprehensive experiments on ShapeNet dataset demonstrate that 3DAttriFlow outperforms the state-of-the-art shape reconstruction methods, and we also validate its generalization ability on shape completion task. Code is available at https://github.com/junshengzhou/3DAttriFlow.

We propose a novel approach to 3D scene painting using a configurable 3D scene layout. Our approach takes a 3D scene with semantic class labels as input and trains a 3D scene painting network that synthesizes color values for the input 3D scene. We exploit an off-the-shelf 2D semantic image synthesis method to teach the 3D painting network without explicit color supervision. Experiments show that our approach produces images with geometrically correct structures and supports scene manipulation, such as the change of viewpoint, object poses, and painting style. Our approach provides rich controllability to synthesized images in the aspect of 3D geometry.

We introduce 3D Moments, a new computational photography effect. As input we take a pair of near-duplicate photos, i.e., photos of moving subjects from similar viewpoints, common in people's photo collections. As output, we produce a video that smoothly interpolates the scene motion from the first photo to the second, while also producing camera motion with parallax that gives a heightened sense of 3D. To achieve this effect, we represent the scene as a pair of feature-based layered depth images augmented with scene flow. This representation enables motion interpolation along with independent control of the camera viewpoint. Our system produces photorealistic space-time videos with motion parallax and scene dynamics, while plausibly recovering regions occluded in the original views. We conduct extensive experiments demonstrating superior performance over baselines on public datasets and in-the-wild photos. Project page: https://3d-moments.github.io/.

360deg cameras can capture complete environments in a single shot, which makes 360deg imagery alluring in many computer vision tasks. However, monocular depth estimation remains a challenge for 360deg data, particularly for high resolutions like 2K (2048x1024) and beyond that are important for novel-view synthesis and virtual reality applications. Current CNN-based methods do not support such high resolutions due to limited GPU memory. In this work, we propose a flexible framework for monocular depth estimation from high-resolution 360deg images using tangent images. We project the 360deg input image onto a set of tangent planes that produce perspective views, which are suitable for the latest, most accurate state-of-the-art perspective monocular depth estimators. To achieve globally consistent disparity estimates, we recombine the individual depth estimates using deformable multi-scale alignment followed by gradient-domain blending. The result is a dense, high-resolution 360deg depth map with a high level of detail, also for outdoor scenes which are not supported by existing methods. Our source code and data are available at https://manurare.github.io/360monodepth/.

The application of deep neural networks (DNNs) on 360-degree images has achieved remarkable progress in the recent years. However, DNNs have been demonstrated to be vulnerable to well-crafted adversarial examples, which may trigger severe safety problems in the real-world applications based on 360-degree images. In this paper, we propose an adversarial attack targeting spherical images, called 360-attactk, that transfers adversarial perturbations from perspective-view (PV) images to a final adversarial spherical image. Given a target spherical image, we first represent it with a set of planar PV images, and then perform 2D attacks on them to obtain adversarial PV images. Considering the issue of the projective distortion between spherical and PV images, we propose a distortion-aware attack to reduce the negative impact of distortion on attack. Moreover, to reconstruct the final adversarial spherical image with high aggressiveness, we calculate the spherical saliency map with a novel spherical spectrum method and next propose a saliency-aware fusion strategy that merges multiple inverse perspective projections for the same position on the spherical image. Extensive experimental results show that 360-attack is effective for disturbing spherical images in the black-box setting. Our attack also proves the presence of adversarial transferability from Z2 to SO3 groups.

Estimating a semantically segmented bird's-eye-view (BEV) map from a single image has become a popular technique for autonomous control and navigation. However, they show an increase in localization error with distance from the camera. While such an increase in error is entirely expected - localization is harder at distance - much of the drop in performance can be attributed to the cues used by current texture-based models, in particular, they make heavy use of object-ground intersections (such as shadows), which become increasingly sparse and uncertain for distant objects. In this work, we address these shortcomings in BEV-mapping by learning the spatial relationship between objects in a scene. We propose a graph neural network which predicts BEV objects from a monocular image by spatially reasoning about an object within the context of other objects. Our approach sets a new state-of-the-art in BEV estimation from monocular images across three large-scale datasets, including a 50% relative improvement for objects on nuScenes.