https://github.com/52CV/CVPR-2021-Papers

CVPR2021最新信息及已接收论文/代码(持续更新)

本贴是对 CVPR2021 已接受论文的粗略汇总，后期会有更详细的总结。期待ing......

官网链接：http://cvpr2021.thecvf.com
开会时间：2021年6月19日-6月25日
论文接收公布时间：2021年2月28日

接收论文IDs：

• CVPR 2021 接收论文列表！27%接受率！

🎆🎆🎆更新提示：3月5日新增 23 篇（2点云+1人脸+1自监督+2姿态+1三维+1模型+1医学+1SR+2GAN+1检索+1HOI+1跟踪+1人群计数+1动作检测+1视频字幕+5未分）

🎆🎆🎆更新提示：3月4日新增 33 篇（2目标检测+3点云+1半监督+1医学+5分割+1域泛化+1人脸+1视图合成+16D位姿+2分类+1跟踪+1图像增强+1GAN+1GNN+1图像字幕+1三维+1相机定位+8未分）

目录

🐱	🐶	🐭	🐹	🐯
❌	Workshop征稿	48.人物交互HOI	47.相机定位	46.图像/视频字幕
45.主动学习	44.动作预测	43.表示学习（图像+字幕）	42.超像素	41.视频语言学习
40.模型偏见消除	39.类增量学习	38.持续学习	37.视频插帧	36.动作检测与识别
35.图像聚类	34.图像分类	33.6D位姿估计	32.视图合成	31. 开放集识别
30.新视角合成	29.姿态估计	28.密集预测	27.活体检测	26.视频编解码
25.三维视觉	24.强化学习	23.自动驾驶	22.医学影像	21.Transformer
20.人员重识别/人群计数	19.量化、剪枝、蒸馏、模型压缩与优化	18.航空影像	17.超分辨率	16.视觉问答
15.GAN	14.小/零样本学习，域适应，域泛化	13.图像检索	12.图像增广	11.人脸技术
10.神经架构搜索	9.目标跟踪	8.图像分割	7.目标检测	6.数据增强
5.异常检测	4.自/半/弱监督学习	3.点云	2.图卷积网络GNN	1.未分类

48.人物交互（human-object interaction）

• Learning Asynchronous and Sparse Human-Object Interaction in Videos

47.相机定位(Camera Localization)

• Robust Neural Routing Through Space Partitions for Camera Relocalization in Dynamic Indoor Environments

46.图像/视频字幕

• Scan2Cap: Context-aware Dense Captioning in RGB-D Scans
  ⭐code🏠project📺video
• VX2TEXT: End-to-End Learning of Video-Based Text Generation From Multimodal Inputs
  视频字幕、视频问答和视频对话任务的多模式框架

45.主动学习

• [Vab-AL: Incorporating Class Imbalance and Difficulty with Variational Bayes for Active Learning]

44.动作预测

• Learning the Predictability of the Future
  预测未来
  ⭐code🏠project📺video
  

43.表示学习（图像+字幕）

• VirTex: Learning Visual Representations from Textual Annotations
  ⭐code

42.超像素

• Learning the Superpixel in a Non-iterative and Lifelong Manner
  

41.视频语言学习（video-and-language learning）

• Less is More: CLIPBERT for Video-and-Language Learning via Sparse Sampling
  😮oral⭐code
  

40.模型偏见消除

• Fair Attribute Classification through Latent Space De-biasing
  ⭐code🏠project
  

39.类增量学习（class-incremental learning）

• IIRC: Incremental Implicitly-Refined Classification
  🏠project
  

38.持续学习

• Rainbow Memory: Continual Learning with a Memory of Diverse Samples
  

37.视频插帧

• FLAVR: Flow-Agnostic Video Representations for Fast Frame Interpolation
  ⭐code🏠project
  

36.动作检测与识别

• Coarse-Fine Networks for Temporal Activity Detection in Videos
  
• 3D CNNs with Adaptive Temporal Feature Resolutions
  
• Modeling Multi-Label Action Dependencies for Temporal Action Localization
  😮oral
  提出基于注意力的网络架构来学习视频中的动作依赖性，用于解决多标签时间动作定位任务。

35.图像聚类

• Improving Unsupervised Image Clustering With Robust Learning
  ⭐code
  利用鲁棒学习改进无监督图像聚类技术
  

34.图像分类

• PML: Progressive Margin Loss for Long-tailed Age Classification
  
• Re-labeling ImageNet: from Single to Multi-Labels, from Global to Localized Labels
  ⭐code
  
• Fine-grained Angular Contrastive Learning with Coarse Labels
  😮oral
  使用自监督进行 Coarse Labels（粗标签）的细粒度分类方面的工作。粗标签与细粒度标签相比，更容易和更便宜，因为细粒度标签通常需要域专家。

33.6D位姿估计

• FFB6D: A Full Flow Bidirectional Fusion Network for 6D Pose Estimation
  ⭐code
  

32.视图合成

• ID-Unet: Iterative Soft and Hard Deformation for View Synthesis

31.开放集识别

• Counterfactual Zero-Shot and Open-Set Visual Recognition
  ⭐code
  
• Few-shot Open-set Recognition by Transformation Consistency
  

30.新视角合成

• DeRF: Decomposed Radiance Fields
  🏠project
  
• D-NeRF: Neural Radiance Fields for Dynamic Scenes
  🏠project
  

29.姿态估计

• PCLs: Geometry-aware Neural Reconstruction of 3D Pose with Perspective Crop Layers
  📺video
  通过消除 location-dependent 透视效果来改进3D人体姿势估计技术工作。
  
• CanonPose: Self-supervised Monocular 3D Human Pose Estimation in the Wild
• Camera-Space Hand Mesh Recovery via Semantic Aggregation and Adaptive 2D-1D Registration
  ⭐code
  
• Monocular Real-time Full Body Capture with Inter-part Correlations
  📺video

28.密集预测

• Densely connected multidilated convolutional networks for dense prediction tasks
  提出的D3Net在语义分割&音乐源分离任务上的表现优于SOTA网络
  

27.活体检测

• Cross Modal Focal Loss for RGBD Face Anti-Spoofing
  

26.视频编解码

• MetaSCI: Scalable and Adaptive Reconstruction for Video Compressive Sensing
  ⭐code

25.三维视觉

• A Deep Emulator for Secondary Motion of 3D Characters
  
• Neural Deformation Graphs for Globally-consistent Non-rigid Reconstruction
  😮oral🏠project📺video
  
• Deep Implicit Templates for 3D Shape Representation
  😮oral⭐code🏠project📺video
  CVPR 2021 Oral，清华学者提出Deep Implicit Templates，极大扩展DIF能力
  

24.强化学习

• Hierarchical and Partially Observable Goal-driven Policy Learning with Goals Relational Graph
  

23.自动驾驶

• Patch-NetVLAD: Multi-Scale Fusion of Locally-Global Descriptors for Place Recognition
  ⭐github
  ECCV 2020 Facebook Mapillary Visual Place Recognition Challenge 冠军方案

22.医学影像

• 3D Graph Anatomy Geometry-Integrated Network for Pancreatic Mass Segmentation, Diagnosis, and Quantitative Patient Management
  用纯多模态 CT 影像可替代目前 JHMI 的需要做肿瘤化学检测和 DNA 测序+医学影像的综合多模态诊断流程，从诊断准确度上有可比较性，定量诊断精度更优
  
• Deep Lesion Tracker: Monitoring Lesions in 4D Longitudinal Imaging Studies
  肿瘤影像里面智能 PACS 辅助医生读片的重要功能
  
• Automatic Vertebra Localization and Identification in CT by Spine Rectification and Anatomically-constrained Optimization
  基于CT 影像的骨折/骨质疏松系统
  
• Multi-institutional Collaborations for Improving Deep Learning-based Magnetic Resonance Image Reconstruction Using Federated Learning
  ⭐code
  多机构合作，利用联合学习改进基于深度学习的磁共振图像重建技术
  
• DeepTag: An Unsupervised Deep Learning Method for Motion Tracking on Cardiac Tagging Magnetic Resonance Images
  😮oral⭐code
  DeepTag: 一种无监督的深度学习方法，用于心脏标记磁共振图像的运动跟踪
  

21.Transformer

• Transformer Interpretability Beyond Attention Visualization
  ⭐code
  
• UP-DETR: Unsupervised Pre-training for Object Detection with Transformers
  
  😮oral⭐code
• Pre-Trained Image Processing Transformer
  
• 3D Vision Transformers for Action Recognition
  用于动作识别的3D视觉Transformer
• MIST: Multiple Instance Spatial Transformer Network
  试图从热图中进行可微的top-K选择(MIST)（目前在自然图像上也有了一些结果；) 用它可以在没有任何定位监督的情况下进行检测和分类（并不是它唯一能做的事情!）

20.人员重识别

• Meta Batch-Instance Normalization for Generalizable Person Re-Identification
  
• Cross-Modal Collaborative Representation Learning and a Large-Scale RGBT Benchmark for Crowd Counting

19.量化、剪枝、蒸馏、模型压缩与优化

• Learning Student Networks in the Wild
  
• ReXNet: Diminishing Representational Bottleneck on Convolutional Neural Network
  ⭐code
  
• Manifold Regularized Dynamic Network Pruning
  
• RepVGG: Making VGG-style ConvNets Great Again
  ⭐code
  
• Coordinate Attention for Efficient Mobile Network Design

18.航空影像

• Dogfight: Detecting Drones from Drone Videos
  
• PointFlow: Flowing Semantics Through Points for Aerial Image Segmentation
  

17.超分辨率

• Data-Free Knowledge Distillation For Image Super-Resolution
  
• AdderSR: Towards Energy Efficient Image Super-Resolution
  ⭐code
  
• Cross-MPI: Cross-scale Stereo for Image Super-Resolution using Multiplane Images
  🏠project📺video
  CVPR 2021，Cross-MPI以底层场景结构为线索的端到端网络，在大分辨率（x8）差距下也可完成高保真的超分辨率

16.视觉问答

• Weakly-supervised Grounded Visual Question Answering using Capsules
  

15.GAN

• Exploiting Spatial Dimensions of Latent in GAN for Real-time Image Editing
  
• Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation
  ⭐code🏠project
  
• Hijack-GAN: Unintended-Use of Pretrained, Black-Box GANs
  
• Image-to-image Translation via Hierarchical Style Disentanglement
  ⭐code
• Efficient Conditional GAN Transfer with Knowledge Propagation across Classes
  ⭐code
• Anycost GANs for Interactive Image Synthesis and Editing
  ⭐code🏠project📺video
  Anycost GAN，可适应广泛的硬件和延迟要求，以及实现交互式图像编辑
• TediGAN: Text-Guided Diverse Image Generation and Manipulation
  ⭐code🏠project📺video
• Generative Hierarchical Features from Synthesizing Images
  😮oral⭐code🏠project
  作者称预训练 GAN 生成器可以当作是一种学习的多尺度损失。用它进行训练可以带来高度竞争的层次化和分离的视觉特征，称之为生成层次化特征（GH-Feat）。并进一步表明，GH-Feat不仅有利于生成性任务，更重要的是有利于分辨性任务，包括人脸验证、关键点检测、layout prediction、迁移学习、style mixing、图像编辑等。

14.小/零样本学习，域适应，域泛化

• Exploring Complementary Strengths of Invariant and Equivariant Representations for Few-Shot Learning
  
• Exploring Complementary Strengths of Invariant and Equivariant Representations for Few-Shot Learning
  
• FSDR: Frequency Space Domain Randomization for Domain Generalization
  受 JPEG 将空间图像转换为多个频率分量(FCs)的启发，提出频率空间域随机化(FSDR)，通过保留域变量FCs(DIFs)和只随机化域变量FCs(DVFs)来随机化频率空间的图像。
• Domain Generalization via Inference-time Label-Preserving Target Projections
  😮 Oral

13.图像检索

• Probabilistic Embeddings for Cross-Modal Retrieval
  
• QAIR: Practical Query-efficient Black-Box Attacks for Image Retrieval

12.图像增强

• Multi-Stage Progressive Image Restoration
  ⭐code
  
• Auto-Exposure Fusion for Single-Image Shadow Removal
  ⭐code
  
• DeFMO: Deblurring and Shape Recovery of Fast Moving Objects
  ⭐code📺video
  

11. 人脸技术

• A 3D GAN for Improved Large-pose Facial Recognition
  
• When Age-Invariant Face Recognition Meets Face Age Synthesis: A Multi-Task Learning Framework
  ⭐github
  
• Multi-attentional Deepfake Detection
  
• MagFace: A Universal Representation for Face Recognition and Quality Assessment
  😮oral⭐code
  人脸识别+质量，今年的Oral presentation。 代码待整理

10.神经架构搜索

• AttentiveNAS: Improving Neural Architecture Search via Attentive
  
• HourNAS: Extremely Fast Neural Architecture Search Through an Hourglass Lens
  
• ReNAS: Relativistic Evaluation of Neural Architecture Search
  

9.目标跟踪

• Probabilistic Tracklet Scoring and Inpainting for Multiple Object Tracking
  
• Rotation Equivariant Siamese Networks for Tracking
  
• Track to Detect and Segment: An Online Multi-Object Tracker
  🏠project📺video
  
• Multiple Object Tracking with Correlation Learning

8.图像分割

• 4D Panoptic LiDAR Segmentation
  
• PLOP: Learning without Forgetting for Continual Semantic Segmentation
  
• Cross-View Regularization for Domain Adaptive Panoptic Segmentation
  😮oral
  用于域自适应全景分割的跨视图正则化方法
  
• Information-Theoretic Segmentation by Inpainting Error Maximization
  
• Towards Semantic Segmentation of Urban-Scale 3D Point Clouds: A Dataset, Benchmarks and Challenges
  ⭐dataset📺video
  
• Exploring Data Efficient 3D Scene Understanding with Contrastive Scene Contexts
  😮oral🏠project📺video
• Real-Time High Resolution Background Matting
  😮oral⭐code🏠project📺video
  最新开源抠图技术，实时快速高分辨率，4k(30fps)、现代GPU（60fps）
  解读：单块GPU实现4K分辨率每秒30帧，华盛顿大学实时视频抠图再升级，毛发细节到位
  最新开源抠图技术，实时快速高分辨率，4k(30fps)、现代GPU（60fps）
• Part-aware Panoptic Segmentation
  

7.目标检测

• Multiple Instance Active Learning for Object Detection
  ⭐code
  
• Positive-Unlabeled Data Purification in the Wild for Object Detection
  
• Depth from Camera Motion and Object Detection
  ⭐github📺video
• There is More than Meets the Eye: Self-Supervised Multi-Object Detection and Tracking with Sound by Distilling Multimodal Knowledge
  🏠project
  
• Categorical Depth Distribution Network for Monocular 3D Object Detection
  😮oral
  
• Semantic Relation Reasoning for Shot-Stable Few-Shot Object Detection
  首个研究少样本检测任务的语义关系推理，并证明它可提升强基线的潜力。
• Towards Open World Object Detection
  😮oral⭐code
• General Instance Distillation for Object Detection
  
• Distilling Object Detectors via Decoupled Features
  
• 3DIoUMatch: Leveraging IoU Prediction for Semi-Supervised 3D Object Detection
  😮oral⭐code🏠project📺video
  更多：CVPR 2021|利用IoU预测进行半监督式3D目标检测

6.数据增广

• KeepAugment: A Simple Information-Preserving Data Augmentation
  

5.异常检测

• Multiresolution Knowledge Distillation for Anomaly Detection
  

4.自/半/弱监督学习

• Weakly Supervised Learning of Rigid 3D Scene Flow
  ⭐code🏠project
  
• Adaptive Consistency Regularization for Semi-Supervised Transfer Learning
  ⭐code
  
• Self-supervised Geometric Perception
  😮oral⭐code
  作者称 SGP 是第一个在几何感知中进行特征学习的通用框架，不需要任何来自 ground-truth 几何标签的监督。SGP以EM方式运行，它迭代执行几何模型的鲁棒估计以生成伪标签，并在噪声伪标签的监督下进行特征学习。将 SGP 应用于相机姿势估计和点云配准，并证明在大规模真实数据集中，SGP 的性能等同于甚至优于监督的权威。

3.点云

• PREDATOR: Registration of 3D Point Clouds with Low Overlap
  😮oral⭐code🏠project
  
• Diffusion Probabilistic Models for 3D Point Cloud Generation
  ⭐code
  
• Style-based Point Generator with Adversarial Rendering for Point Cloud Completion
  
• SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration
  ⭐code
• MultiBodySync: Multi-Body Segmentation and Motion Estimation via 3D Scan Synchronization
  😮oral⭐code
• TPCN: Temporal Point Cloud Networks for Motion Forecasting
  用于运动预测的时空点云网络
  
• PointGuard: Provably Robust 3D Point Cloud Classification

2.图卷积网络GNN

• Sequential Graph Convolutional Network for Active Learning
  
• Quantifying Explainers of Graph Neural Networks in Computational Pathology
  

1.未分类

• Inverting the Inherence of Convolution for Visual Recognition
  
• Representative Batch Normalization with Feature Calibration
  
• UC2: Universal Cross-lingual Cross-modal Vision-and-Language Pretraining
  
• Reconsidering Representation Alignment for Multi-view Clustering
  
• Self-supervised Simultaneous Multi-Step Prediction of Road Dynamics and Cost Map
  
• Instance Localization for Self-supervised Detection Pretraining
  ⭐code
  
• Model-Contrastive Federated Learning
  提出模型对比学习来解决联合学习中的非IID数据问题
  
• Neural Geometric Level of Detail:Real-time Rendering with Implicit 3D Surfaces
  😮Oral⭐code🏠project
  
• Data-Free Model Extraction
  ⭐code
  
• Single-Stage Instance Shadow Detection with Bidirectional Relation Learning
  😮oral
  
• Continual Adaptation of Visual Representations via Domain Randomization and Meta-learning
  😮oral
• PatchmatchNet: Learned Multi-View Patchmatch Stereo
  😮oral⭐code
• [Online Bag-of-Visual-Words Generation for Unsupervised Representation Learning]
• [Semantic Palette: Guiding Scene Generation with Class Proportions]
• Function4D: Real-time Human Volumetric Capture from Very Sparse Consumer RGBD Sensors
  😮oral
• POSEFusion:Pose-guided Selective Fusion for Single-view Human Volumetric Capture
  😮oral
• Multi-Objective Interpolation Training for Robustness to Label Noise
  ⭐code
• Right for the Right Concept: Revising Neuro-Symbolic Concepts by Interacting with their Explanations
• Simpler Certified Radius Maximization by Propagating Covariances
  😮oral📺video

Workshop 征稿ing

• Visual Perception for Navigation in Human Environments
  第二届人类环境导航视觉感知征稿 ⚠️4月15截止

• UG 2 + Challenge
  旨在通过应用图像恢复和增强算法提高分析性能，推动对 "difficult"图像的分析。参与者任务是开发新的算法，以改进对在问题条件下拍摄的图像分析。
  👑10K美元奖金
  

  • 低能见度环境下的目标检测
    • 雾霾条件下的(半)监督目标检测
    • (半)低光条件下的人脸检测
  • 黑暗视频中的动作识别
    • 黑暗中进行完全监督动作识别
    • 黑暗中进行半监督动作识别

• Continual Learning in Computer Vision 征稿中
  旨在聚集学术界和工业界的研究人员和工程师，讨论持续学习的最新进展。
  

  • Best paper award: 500 USD + 500 USD worth of Huawei cloud credits (HUAWEI)
  • Overall Challenge winner: 1,000 USD + 500 USD worth of Huawei cloud credits (HUAWEI)
  • Supervised-Learning track winner: 500 USD (HUAWEI)
  • Reinforcement-Learning track winner: 500 USD (ServiceNow)

• 第四届UG2研讨会和竞赛：弥合计算成像与视觉识别之间的鸿沟

• 10万美元奖金！CVPR 2021 重磅赛事，安全AI挑战者计划

  • CVPR 2021大赛， 安全AI 之防御模型的「白盒对抗攻击」解析
  • 还在刷榜ImageNet？找出模型的脆弱之处更有价值！

• Responsible Computer Vision
  ⚠️3月25日截止
  本次研讨会将广泛讨论计算机视觉背景下负责任的人工智能的三个主要方面：公平性；可解释性和透明度；以及隐私。

• Holistic Video Understanding
  目的是建立一个整合所有语义概念联合识别的视频基准，因为每个任务的单一类标签往往不足以描述视频的整体内容。

• ThreeDWorld Transport Challenge
  ⚠️6月1截止
  📺video

扫码CV君微信（注明：CVPR）入微信交流群：