Skip to content
/ Domain-generalization Public

This is a repository for organizing papers ,codes, and etc related to Domain Generalization

License

MIT license
22 stars 0 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

Ziwei-Niu/Domain-generalization

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

15 Commits
LICENSE
LICENSE
 
 
README.md
README.md
 
 

Repository files navigation

Domain-generalization

This is a repository for organizing papers ,codes, and etc related to Domain Generalization.

Table of Contents

Papers (ongoing)

Survey papers

2021

Theory & Analysis

We list the papers that either provide inspiring theoretical analyses or conduct extensive empirical studies for domain generalization.

2022
2021
2020
2019

  • 研究了一个基于核的学习算法，并建立了一个泛化误差边界对DG多分类进行理论分析：
    Explainable Deep Classification Models for Domain Generalization
    Author：Aniket Anand Deshmukh, Yunwen Lei, Srinagesh Sharma, Urun Dogan, James W. Cutler, Clayton Scott
    arXiv preprint arXiv:1905.10392 (2019)

  • 介绍了不变风险最小化(IRM)，一种估计跨越多个训练分布的不变相关的学习范式：
    Invariant Risk Minimization
    Author：Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, David Lopez-Paz
    arXiv preprint arXiv:1907.02893 (2019)

Multiple Domain Generalization

Multiple Domain Generalization aims to learn a model from multiple source domains that will generalize well on unseen target domains.

Data manipulation

Data Augmentation-Based Methods

Data augmentation-based methods augment original data to enhance the generalization performance of the model, typical augmentation operations include flipping, rotation, scaling, cropping, adding noise, and so on.

2022
2021

  • (DecAug) ：提出了一种新颖的分解特征表示和语义增强的方法，用于OoD泛化
    DecAug: Out-of-Distribution Generalization via Decomposed Feature Representation and Semantic Augmentation
    Author： Haoyue Bai, Rui Sun, Lanqing Hong, Fengwei Zhou, Nanyang Ye, Han-Jia Ye, S.-H. Gary Chan, Zhenguo Li
    Association for the Advancement of Artificial Intelligence (AAAI) (2021)

  • (MixStyle) 本文提出了一种基于概率地混合源域中训练样本的实例级特征统计的方法。混合训练实例的风格导致新的领域被隐含地合成，这增加了源领域的多样性，从而增加了训练模型的泛化能力：
    Domain Generalization with MixStyle
    Author：Kaiyang Zhou, Yongxin Yang, Yu Qiao, Tao Xiang
    International Conference on Learning Representations (ICLR) (2021)
    [Code]

  • (FSDR) 提出了频率空间域随机化（FSDR），通过保留域不变的频率分量（DIFs）和只随机化域可变的频率分量（DVFs），在频率空间中随机化图像：
    FSDR: Frequency Space Domain Randomization for Domain Generalization
    Author：Huang, Jiaxing, Dayan Guan, Aoran Xiao, and Shijian Lu
    Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

  • (ATSRL) 多视角学习，提出对抗性师生表征学习框架，将表征学习和数据增广相结合，前者逐步更新教师网络以得出域通用的表征，而后者则合成数据的外源但合理的分布：
    Adversarial Teacher-Student Representation Learning for Domain Generalization
    Author：Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, Yu-Chiang Frank Wang
    Advances in Neural Information Processing Systems 34 (NeurIPS) (2021)

  • (MBDG) 提出了一种具有收敛保证的新型域泛化算法：
    Model-Based Domain Generalization
    Author：Alexander Robey, George J. Pappas, Hamed Hassani
    Advances in Neural Information Processing Systems 34 (NeurIPS) (2021)
    [Code]

  • (FACT) 开发了一种新颖的基于傅里叶的数据增强策略，并引入了一种称为co-teacher regularization的双重形式的一致性损失来学习域不变表征：
    A Fourier-Based Framework for Domain Generalization
    Author：Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, Qi Tian
    Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

  • (SBL) 开发了一种多视图正则化的元学习算法，在更新模型时采用多个任务来产生合适的优化方向。在测试阶段，利用多个增强的图像来产生多视图预测，通过融合测试图像的不同视图的结果来显著提高模型的可靠性：
    More is Better: A Novel Multi-view Framework for Domain Generalization
    Author：Jian Zhang, Lei Qi, Yinghuan Shi, Yang Gao
    arXiv preprint arXiv:2112.12329 (2021)

  • 这项工作奠定了领域泛化的学习理论基础，提出了两个正式的数据生成模型，相应的风险概念，以及无分布泛化误差分析：
    Domain Generalization by Marginal Transfer Learning
    Author：Blanchard, Gilles, Aniket Anand Deshmukh, Urun Dogan, Gyemin Lee, and Clayton Scott
    Journal of Machine Learning Research (JMLR) (2021)

2020

  • (M-ADA) 提出了一种名为对抗性领域增强的新方法来来创建 "虚构 "而又 "具有挑战性 "的样本，进而解决分布外（OOD）的泛化问题：
    Learning to Learn Single Domain Generalization
    Author：Fengchun Qiao, Long Zhao, Xi Peng
    Conference on Computer Vision and Pattern Recognition (CVPR) (2020)
    [Code]

  • (EISNet) 提出了一个新的领域泛化框架（称为EISNet），利用多任务学习范式，从多源领域的图像的外在关系监督和内在自我监督中同时学习如何跨领域泛化：
    Learning from Extrinsic and Intrinsic Supervisions for Domain Generalization
    Author：Shujun Wang, Lequan Yu, Caizi Li, Chi-Wing Fu, Pheng-Ann Heng
    Proceedings of the European Conference on Computer Vision (ECCV) (2020)
    [code]

2019
2018

  • (CROSSGRAD) 保留并利用了domain信息，利用domain信息辅助样本扩增，丰富了域内的数据样本：
    Generalizing Across Domains via Cross-Gradient Training
    Author：Shankar, Shiv, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, Sunita Sarawagi
    International Conference on Learning Representations (ICLR) (2018)
    [Code]

  • 提出了一个迭代程序，用来自一个虚构的目标域的例子来扩充源域数据：
    Generalizing to Unseen Domains via Adversarial Data Augmentation
    Author：Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C. Duchi, Vittorio Murino, Silvio Savarese
    Advances in Neural Information Processing Systems 31 (NeurIPS) (2018)
    [Code]

Data Generation-Based Methods

Data generation-based methods aim to generate new domain samples using some generative models such as Generative Adversarial Networks (GAN), Variational Auto-encoder (VAE).

2022
  • (RICE) 数据生成与因果学习结合，基于修改非因果特征但不改变因果部分的转换，在不明确恢复因果特征的情况下解决OOD问题：
    Out-of-Distribution Generalization With Causal Invariant Transformations
    Author：Ruoyu Wang, Mingyang Yi, Zhitang Chen, Shengyu Zhu
    Conference on Computer Vision and Pattern Recognition (CVPR) (2022)
2021
2020
2018

Representation learning

Domain-Invariant Representation-Based Methods

Domain-invariant representation-based methods aim to reduce the representation discrepancy between multiple source domains in a specific feature space to be domain invariant so that the learned model can have a generalizable capability to the unseen domain.

2022

  • (LADG) 提出了具有空间紧凑性维护的局部对抗式域泛化 (LADG)，解决了以往对抗式域泛化的限制：
    Localized Adversarial Domain Generalization
    Author：Wei Zhu, Le Lu, Jing Xiao, Mei Han, Jiebo Luo, Adam P. Harrison
    Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

  • (BatchFormer) 引入了一个BatchFormer模块，将其应用于每个mini-batch的批处理维度，在训练期间隐含地探索样本关系：
    BatchFormer: Learning To Explore Sample Relationships for Robust Representation Learning
    Author：Zhi Hou, Baosheng Yu, Dacheng Tao
    Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

  • (XDED) 提出了一种跨域集合蒸馏方法（cross-domain ensemble distillation），通过学习域不变表征并鼓励模型达到flat minima：
    Cross-Domain Ensemble Distillation for Domain Generalization
    Author：Kyungmoon Lee, Sungyeon Kim, Suha Kwak
    European Conference on Computer Vision (ECCV) (2022)

2021

  • (SFA) 提出了一种基于特征增强的增强方法，即在训练过程中用高斯噪声扰动特征嵌入对源域数据进行增广：
    A Simple Feature Augmentation for Domain Generalization
    Author：Pan Li, Da Li, Wei Li, Shaogang Gong, Yanwei Fu, Timothy M. Hospedales
    International Conference on Computer Vision (ICCV) (2021)

  • (DFDG) ：在不需要源域标签的情况下，通过类别条件的软标签来协调样本的类别关系，以学习领域不变的类区分特征：
    Robust Domain-Free Domain Generalization with Class-Aware Alignment
    Author：Wenyu Zhang; Mohamed Ragab; Ramon Sagarna
    International Conference on Acoustics, Speech, and Signal Processing (ICASSP CCF-B) (2021)

  • (ATSRL) 多视角学习，提出对抗性师生表征学习框架，将表征学习和数据增广相结合，前者逐步更新教师网络以得出域通用的表征，而后者则合成数据的外源但合理的分布：
    Adversarial Teacher-Student Representation Learning for Domain Generalization
    Author：Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, Yu-Chiang Frank Wang
    Advances in Neural Information Processing Systems 34 (NeurIPS) (2021)

2020

  • (MMLD) 介绍了使用多个潜在域的混合用于领域泛化，作为一种新的和更现实的场景，其试图在不使用域标签的情况下训练一个域泛化的模型：
    Domain Generalization Using a Mixture of Multiple Latent Domains
    Author：Toshihiko Matsuura, Tatsuya Harada
    Association for the Advancement of Artificial Intelligence (AAAI) (2020)
    [Code]

  • (BNE) 依靠特定领域的归一化层来分解每个训练领域的独立表征，然后使用这种隐式嵌入来定位来自未知域的未见过的样本：
    Batch Normalization Embeddings for Deep Domain Generalization
    Author： Mattia Segu, Alessio Tonioni, Federico Tombari
    arXiv preprint arXiv:2011.12672 (2020)

  • (ZSDG) 将DG扩展到一个更具挑战性的环境中，即未见过的领域的标签空间也可能发生变化：
    Zero Shot Domain Generalization
    Author： Udit Maniyar, Joseph K J, Aniket Anand Deshmukh, Urun Dogan, Vineeth N Balasubramanian
    British Machine Vision Conference (BMVC) (2020)

2019

  • (G2DM) 采用了多个一比一的领域判别器，从而在训练时估计并最小化源分布之间的配对分歧：
    Generalizing to unseen domains via distribution matching
    Author：Isabela Albuquerque, João Monteiro, Mohammad Darvishi, Tiago H. Falk, Ioannis Mitliagkas
    arXiv preprint arXiv:1911.00804 (2019)

  • (MDA) 提出了多域判别分析 (MDA) 学习一个领域不变的特征转换：
    Domain Generalization via Multidomain Discriminant Analysis
    Author：Hu, Shoubo, Kun Zhang, Zhitang Chen, Laiwan Chan
    Conference on Uncertainty in Artificial Intelligence (PMLR-UAI) 2019
    [code]

  • (G2DM) 采用了多个一比一的领域判别器，从而在训练时估计并最小化源分布之间的配对分歧：
    Generalizing to unseen domains via distribution matching
    Author：Isabela Albuquerque, João Monteiro, Mohammad Darvishi, Tiago H. Falk, Ioannis Mitliagkas
    arXiv preprint arXiv:1911.00804 (2019)

2018

  • (CIDDG) 提出了一个端到端的条件不变的深度域泛化方法，利用深度神经网络进行领域不变的表征学习：
    Deep Domain Generalization via Conditional Invariant Adversarial Networks
    Author：Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, Dacheng Tao
    European Conference on Computer Vision (ECCV) (2018)

  • (MMD-AAE) 提出了一个基于对抗性自动编码器的新框架，使不同领域的分布一致以学习跨领域的广义潜在特征表示：
    Domain Generalization With Adversarial Feature Learning
    Author：Haoliang Li, Sinno Jialin Pan, Shiqi Wang, Alex C. Kot
    Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

  • (CIDG) 提出了一种条件不变的域泛化方法，考虑到P(X)和P(Y|X)都会跨域变化的情况：
    Domain Generalization via Conditional Invariant Representations
    Author：Ya Li, Mingming Gong, Xinmei Tian, Tongliang Liu, Dacheng Tao
    Association for the Advancement of Artificial Intelligence (AAAI) (2018)

2017
2016

  • (KDICA) 为多源领域泛化来开发一种新的面向属性的特征表示，以方便应用现成的分类器来获得高质量的属性检测器：
    Learning Attributes Equals Multi-Source Domain Generalization
    Author：Chuang Gan, Tianbao Yang, Boqing Gong
    Conference on Computer Vision and Pattern Recognition (CVPR) (2016)

  • (ESRand) 通过减少多领域学习中的分布偏差学习域不变表征，提升模型泛化能力：
    Robust domain generalisation by enforcing distribution invariance
    Author：Erfani, Sarah, Baktashmotlagh, Mahsa, Moshtaghi, Masud, Nguyen, Xuan, Leckie, Christopher, Bailey, James, Kotagiri, Rao
    International Joint Conference on Artificial Intelligence 25 (IJCAI) (2016)

2015
2013

Disentangled Representation Learning-Based Methods

Disentangled representation learning-based methods aim to disentangle domain-specific and domain-invariant parts from source data, and then adopt the domain-invariant one for inference on the target domains.

2022
  • (DDG) 将OOD泛化问题形式化为约束性优化问题，称为Disentanglement-constrained Domain Generalization (DDG)：
    Towards Principled Disentanglement for Domain Generalization
    Author：Hanlin Zhang, Yi-Fan Zhang, Weiyang Liu, Adrian Weller, Bernhard Schölkopf, Eric P. Xing
    Conference on Computer Vision and Pattern Recognition (CVPR) (2022)
    [code]
2021
2020
2019
  • (DADA) 提出了一种新的深度对抗性分解自动编码器 (DADA)来分解潜在空间中的域不变特征：
    DIVA: Domain Invariant Variational Autoencoders
    Author：Xingchao Peng, Zijun Huang, Ximeng Sun, Kate Saenko
    International Conference on Machine Learning (PMLR-ICML) (2019)
    [code]
2018
2017

  • 为端到端DG学习开发了一个低秩参数化的CNN模型，其次提出了一个新的DG数据集——PACS，具有更大的域偏移：
    Deeper, broader and artier domain generalization
    Author：Zhengming Ding, Yun Fu
    Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2017)
    [code]

  • (SLRC) 开发了一个具有结构化低秩约束的深度域泛化框架，通过捕捉多个相关源领域的一致知识来促进未见过的目标域评估：
    Deep Domain Generalization With Structured Low-Rank Constraint
    Author：Li, Da, Yongxin Yang, Yi-Zhe Song, Timothy M. Hospedales
    IEEE Transactions on Image Processing (TIP CCF-A) (2017)

2012
  • (Undo-Bias) 提出了一个鉴别性的框架，在训练中直接利用数据集的偏差：
    Undoing the damage of dataset bias
    Author：Khosla, Aditya, Tinghui Zhou, Tomasz Malisiewicz, Alexei A. Efros, Antonio Torralba
    European Conference on Computer Vision (ECCV) (2012)
    [code]

Learning strategy

Some methods also use some machine learning paradigms to solve DG tasks.

Ensemble Learning-Based Methods

Ensemble learning usually combines multiple models, such as classifiers or experts, to enhance the power of models to make accurate prediction.

2022

  • (DDG) 通过将语义和变化表征分离到不同的子空间，同时强制执行不变性约束，以学习语义概念的内在表征：
    Towards Unsupervised Domain Generalization
    Author：Xingxuan Zhang, Linjun Zhou, Renzhe Xu, Peng Cui, Zheyan Shen, Haoxin Liu
    Conference on Computer Vision and Pattern Recognition (CVPR) (2022)
    [code]

  • (XDED) 提出了一种跨域集合蒸馏方法（cross-domain ensemble distillation），通过学习域不变表征并鼓励模型达到flat minima：
    Cross-Domain Ensemble Distillation for Domain Generalization
    Author：Kyungmoon Lee, Sungyeon Kim, Suha Kwak
    European Conference on Computer Vision (ECCV) (2022）

2021
2020
2018
2015
  • (MVDG) 使用具有多种类型特征（即多视角特征）的源域样本来学习具有泛化能力的分类器：
    Multi-view domain generalization for visual recognition
    Author：Niu, Li, Wen Li, and Dong Xu
    Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2015)
2015

Meta-Learning-Based Methods

Meta-learning-based methods aim to divide the data form multi-source domains into meta-train and meta-test sets to simulate domain shift.

2022

  • (MetaCNN) 提出元卷积神经网络，通过将图像的卷积特征分解为元特征，并作为 "视觉词汇"：
    Meta Convolutional Neural Networks for Single Domain Generalization
    Author：Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, Xian-Sheng Hua
    Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

  • (DDG) 探索一种无需训练的机制来调整模型以适应不可知的目标领域。将网络参数解耦为静态和动态部分，以区分域共享和域特定的特征，其中后者由元调整器针对不同域的新样本进行动态调整：
    Dynamic Domain Generalization
    Author：Zhishu Sun, Zhifeng Shen, Luojun Lin, Yuanlong Yu, Zhifeng Yang, Shicai Yang, Weijie Chen
    International Joint Conference on Artificial Intelligence (IJCAI) (2022)
    [Code]

2021
2020

  • (M-ADA) 提出了一种名为对抗性领域增强的新方法来来创建 "虚构 "而又 "具有挑战性 "的样本，进而解决分布外（OOD）的泛化问题：
    Learning to Learn Single Domain Generalization
    Author：Fengchun Qiao, Long Zhao, Xi Peng
    Conference on Computer Vision and Pattern Recognition (CVPR) (2020)
    [Code]

  • (Meta-CVAE) 提出了元条件变异自动编码器（Meta-CVAE），一个新的元概率潜变量框架，用于领域泛化：
    Meta conditional variational auto-encoder for domain generalization
    Author：Zhiqiang Ge, Zhihuan Song, Xin Li, Lei Zhang
    Computer Vision and Image Understanding (2020)

2019
2018

  • (MetaReg) 用一个新的正则化函数来编码域泛化的概念，并提出了在 "学会学习"（或）元学习框架中寻找这样一个正则化函数的问题
    MetaReg: Towards Domain Generalization using Meta-Regularization
    Author：Balaji, Yogesh, Swami Sankaranarayanan, and Rama Chellappa
    Advances in Neural Information Processing Systems (NeurIPS) (2018)

  • (MLDG) 首次提出用于DG的元学习方法，通过在每个小批次中合成虚拟测试域来模拟训练期间的训练/测试域偏移：
    Learning to generalize: Meta-learning for domain generalization
    Author：Li, Da, Yongxin Yang, Yi-Zhe Song, Timothy M. Hospedales
    AAAI Conference on Artificial Intelligence (AAAI) 2018
    [code]

Gradient Operation-Based Methods

Gradient operation-based methods mainly consider using gradient information to force the network learn generalized representations.

2021
  • (DGvGS) 描述了在领域偏移情况下出现的冲突梯度会降低泛化性能，并设计了基于梯度手术的新型梯度协议策略来减轻其影响：
    Domain Generalization via Gradient Surgery
    Author：Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, Enzo Ferrante
    International Conference on Computer Vision (ICCV) (2021)
    [Code]
2019

Regularization-Based Methods

2021

  • (SBL) 开发了一种多视图正则化的元学习算法，在更新模型时采用多个任务来产生合适的优化方向。在测试阶段，利用多个增强的图像来产生多视图预测，通过融合测试图像的不同视图的结果来显著提高模型的可靠性：
    More is Better: A Novel Multi-view Framework for Domain Generalization
    Author：Jian Zhang, Lei Qi, Yinghuan Shi, Yang Gao
    arXiv preprint arXiv:2112.12329 (2021)

  • (MBDG) 提出了一种具有收敛保证的新型域泛化算法：
    Model-Based Domain Generalization
    Author：Alexander Robey, George J. Pappas, Hamed Hassani
    Advances in Neural Information Processing Systems 34 (NeurIPS) (2021)
    [Code]

  • (SelfReg) 提出了一种新的基于自监督对比学习的领域泛化的正则化方法，其只使用正面的数据对，解决了由负面数据对采样引起的各种问题：
    SelfReg: Self-Supervised Contrastive Regularization for Domain Generalization
    Author：Daehee Kim, Youngjun Yoo, Seunghyun Park, Jinkyu Kim, Jaekoo Lee
    Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

2020

  • (LDDG) 通过变分编码器学习一个具有代表性的特征空间，并用一个新的线性依赖正则化项来捕捉从不同领域收集的医学数据中的可共享信息，以提升医学图像分类模型泛化能力：
    Domain Generalization for Medical Imaging Classification with Linear-Dependency Regularization
    Author：Haoliang Li, Yufei Wang, Renjie Wan, Shiqi Wang, Tie-Qiang Li, Alex Kot
    Neural Information Processing Systems 33 (NeurIPS) (2020)
    [Code]

  • (DG_via_ER) 提出了一个衡量所学特征与类标签之间依赖性的熵正则化项，保证在所有源领域中学习条件不变的特征，从而可以学习具有更好泛化能力的分类器：
    Domain Generalization via Entropy Regularization
    Author：Shanshan Zhao, Mingming Gong, Tongliang Liu, Huan Fu, Dacheng Tao
    Neural Information Processing Systems 33 (NeurIPS) (2020)
    [Code]

  • (RSC) 引入一种简单的训练启发式方法，以提高跨域泛化能力。这种方法抛弃了与每个周期的高梯度相关的表征，并迫使模型用剩余的信息进行预测：
    Self-Challenging Improves Cross-Domain Generalization
    Author：Zeyi Huang, Haohan Wang, Eric P. Xing, Dong Huang
    European Conference on Computer Vision (ECCV) (2020)

2019
  • (IRM) 奠基之作，跳出经验风险最小化--不变风险最小化：
    Invariant Risk Minimization
    Author：Arjovsky, Martin and Bottou, Leon and Gulrajani, Ishaan and Lopez-Paz, David
    arXiv preprint arXiv:1907.02893 (2019)
    [code]
2018
  • (MetaReg) 用一个新的正则化函数来编码域泛化的概念，并提出了在 "学会学习"（或）元学习框架中寻找这样一个正则化函数的问题：
    MetaReg: Towards Domain Generalization using Meta-Regularization
    Author：Balaji, Yogesh, Swami Sankaranarayanan, and Rama Chellappa
    Advances in Neural Information Processing Systems (NeurIPS) (2018)

Normalization-Based Methods

Normalization-based methods calibrate data from different domains by normalizing them with their statistic.

2022
  • (GpreBN) 重新审视了批量归一化(BN)，并提出了一种新的测试阶段的BN层设计：
    Test-time Batch Normalization
    Author：Tao Yang, Shenglong Zhou, Yuwang Wang, Yan Lu, Nanning Zheng
    arXiv preprint arXiv:2205.10210 (2022)
2021

  • (SNR) 提出了一个简单而有效的风格标准化和重构 (SNR) 模块,通过归一化 (In-stance Normalization，IN) 过滤掉风格的变化：
    Style Normalization and Restitution for Domain Generalization and Adaptation
    Author： Xin Jin, Cuiling Lan, Wenjun Zeng, Zhibo Chen
    IEEE Transactions on Multimedia (TMM CCF-B) (2021)

  • (FACT) 开发了一种新颖的基于傅里叶的数据增强策略，并引入了一种称为co-teacher regularization的双重形式的一致性损失来学习域不变表征：
    A Fourier-Based Framework for Domain Generalization
    Author：Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, Qi Tian
    Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

2020

Causality-Based Methods

Causality-based methods analyze and address the DG problem from a causal perspective.

2022
2021
2019
  • (IRM) 奠基之作，跳出经验风险最小化--不变风险最小化：
    Invariant Risk Minimization
    Author：Arjovsky, Martin and Bottou, Leon and Gulrajani, Ishaan and Lopez-Paz, David
    arXiv preprint arXiv:1907.02893 (2019)
    [code]

Test-Time-Based Methods

Test-time-based methods leverage the test data, which is available at test-time, to improve generalization performance without any further model training.

2022

  • (T3A) 提出了test-time template adjuster(T3A)，利用测试数据将调整与预测同时进行：
    Test-Time Classifier Adjustment Module for Model-Agnostic Domain Generalization
    Author：Yusuke Iwasawa, Yutaka Matsuo
    Neural Information Processing Systems 34 (NeurIPS) (2021)
    [Slides]

  • (GpreBN) 重新审视了批量归一化(BN)，并提出了一种新的测试阶段的BN层设计：
    Test-time Batch Normalization
    Author：Tao Yang, Shenglong Zhou, Yuwang Wang, Yan Lu, Nanning Zheng
    arXiv preprint arXiv:2205.10210 (2022)

  • (TASD) 利用医学分割图像的语义形状先验信息，并设计了具有双一致性的测试阶段适应策略：
    Single-domain Generalization in Medical Image Segmentation via Test-time Adaptation from Shape Dictionary
    Author：Quande Liu, Cheng Chen1, Qi Dou1, Pheng-Ann Heng
    Association for the Advancement of Artificial Intelligence 36 (AAAI) (2022)

  • (DDG) 探索一种无需训练的机制来调整模型以适应不可知的目标领域。将网络参数解耦为静态和动态部分，以区分域共享和域特定的特征，其中后者由元调整器针对不同域的新样本进行动态调整：
    Dynamic Domain Generalization
    Author：Zhishu Sun, Zhifeng Shen, Luojun Lin, Yuanlong Yu, Zhifeng Yang, Shicai Yang, Weijie Chen
    International Joint Conference on Artificial Intelligence (IJCAI) (2022)
    [Code]

  • (TAF-Cal) 通过在测试时用源原型来校准目标风格，减轻了训练期间对目标域数据没有了解的压力：
    Test-time Fourier Style Calibration for Domain Generalization
    Author：Xingchen Zhao, Chang Liu, Anthony Sicilia, Seong Jae Hwang, Yun Fu
    arXiv preprint arXiv:2205.06427 (2022)

Others

2022
  • (DFF) 提出了Deep Frequency Filtering (DFF),能够增强可迁移的频率成分，并抑制潜在空间中无益于泛化的成分：
    Deep Frequency Filtering for Domain Generalization
    Author：Shiqi Lin, Zhizheng Zhang, Zhipeng Huang, Yan Lu, Cuiling Lan, Peng Chu, Quanzeng You, Jiang Wang, Zicheng Liu, Amey Parulkar, Viraj Navkal, Zhibo Chen
    arXiv preprint arXiv:2203.12198 (2022)
2021

Single Domain Generalization

Single domain generalization aims learn a generalized model only use one domain, which is more challenge but more realistic.

2022
2021
2020
  • (M-ADA) 提出了一种名为对抗性领域增强的新方法来来创建 "虚构 "而又 "具有挑战性 "的样本，进而解决分布外（OOD）的泛化问题：
    Learning to Learn Single Domain Generalization
    Author：Fengchun Qiao, Long Zhao, Xi Peng
    Conference on Computer Vision and Pattern Recognition (CVPR) (2020)
    [Code]
2018
  • (GUD) 一种新的对抗性数据增强方法用于解决单一源域泛化问题，该方法可以在未见过的数据分布中学习到更好的泛化性：
    Generalizing to Unseen Domains via Adversarial Data Augmentation
    Author：Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C. Duchi, Vittorio Murino, Silvio Savarese
    Advances in Neural Information Processing Systems 31 (NeurIPS) (2018)
    [Code]

Self-Supervised Domain Generalization

Self-supervised domain generalization methods improve model generalization ability by solving some pretext tasks with data itself.

2021

  • (ATSRL) 多视角学习，提出对抗性师生表征学习框架，将表征学习和数据增广相结合，前者逐步更新教师网络以得出域通用的表征，而后者则合成数据的外源但合理的分布：
    Adversarial Teacher-Student Representation Learning for Domain Generalization
    Author：Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, Yu-Chiang Frank Wang
    Advances in Neural Information Processing Systems 34 (NeurIPS) (2021)

  • (SelfReg) 提出了一种新的基于自监督对比学习的领域泛化的正则化方法，其只使用正面的数据对，解决了由负面数据对采样引起的各种问题：
    SelfReg: Self-Supervised Contrastive Regularization for Domain Generalization
    Author：Daehee Kim, Youngjun Yoo, Seunghyun Park, Jinkyu Kim, Jaekoo Lee
    Conference on Computer Vision and Pattern Recognition (CVPR) (2021)

  • (PDEN) 提出了一个新颖的渐进式域扩展网络 (PDEN) 学习框架，通过逐渐生成模拟目标与数据，提升模型泛化能力：
    Progressive Domain Expansion Network for Single Domain Generalization
    Author：Fengchun Qiao, Xi Peng
    Conference on Computer Vision and Pattern Recognition (CVPR) (2021)
    [Code]

2020
  • (EISNet) 提出了一个新的领域泛化框架（称为EISNet），利用多任务学习范式，从多源领域的图像的外在关系监督和内在自我监督中同时学习如何跨领域泛化：
    Learning from Extrinsic and Intrinsic Supervisions for Domain Generalization
    Author：Shujun Wang, Lequan Yu, Caizi Li, Chi-Wing Fu, Pheng-Ann Heng
    Proceedings of the European Conference on Computer Vision (ECCV) 2020
    [code]
2020
  • (JiGen) 以监督的方式学习语义标签，并通过从自我监督的信号中学习如何解决相同图像上的拼图来提升泛化能力：
    Domain Generalization by Solving Jigsaw Puzzles
    Author：Fabio M. Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, Tatiana Tommasi
    Conference on Computer Vision and Pattern Recognition (CVPR) (2019)
    [Code]

Semi/Weak/Un-Supervised Domain Generalization

Semi/weak-supervised domain generalization assumes that a part of the source data is unlabeled, while unsupervised domain generalization assumes no training supervision.

2022

  • (BrAD) 提出了一种新颖的自监督跨域学习方法，将所有的域在语义上与一个共同的辅助桥域进行对齐：
    Unsupervised Domain Generalization by Learning a Bridge Across Domains
    Author：Sivan Harary, Eli Schwartz, Assaf Arbelle, Peter Staar, Shady Abu-Hussein, Elad Amrani, Roei Herzig, Amit Alfassy, Raja Giryes, Hilde Kuehne, Dina Katabi, Kate Saenko, Rogerio S. Feris, Leonid Karlinsky
    Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

  • (DARLING) 关注模型预训练的过程对DG任务的影响，设计了一个在DG数据集无监督预训练的算法：
    Towards Unsupervised Domain Generalization
    Author：Xingxuan Zhang, Linjun Zhou, Renzhe Xu, Peng Cui, Zheyan Shen, Haoxin Liu
    Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

  • (PCL) 提出了一种基于代理的对比学习方法，用代理对样本的关系取代了原来对比学习中的样本对样本的关系，缓解了正向对齐问题：
    PCL: Proxy-Based Contrastive Learning for Domain Generalization
    Author：Xufeng Yao, Yang Bai, Xinyun Zhang, Yuechen Zhang, Qi Sun, Ran Chen, Ruiyu Li, Bei Yu
    Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

2021

Open/Heterogeneous Domain Generalization

Open/heterogeneous domain generalization assumes the label space of one domain is different from that of another domain.

2021
2020
  • 提出了一种新的异质域泛化方法，即用两种不同的采样策略将多个源域的样本混合起来：
    Heterogeneous Domain Generalization Via Domain Mixup
    Author：Yufei Wang, Haoliang Li, Alex C. Kot
    IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP CCF-B) (2020)
    [Code]
2019
  • 考虑了更具挑战性的异质领域泛化设置，即未见过的领域与已见过的领域不共享标签空间，目标是训练一个对新数据和新类别有用的现成的特征表示：
    Feature-Critic Networks for Heterogeneous Domain Generalization
    Author：Yiying Li, Yongxin Yang, Wei Zhou, Timothy Hospedales
    International Conference on Machine Learning (PMLR-ICML) (2019)
    [Code]

Federated Domain Generalization

Federated domain generalization assumes that source data is non-shared and can not be collected to train a centralized model for data privacy and transmission restrictions.

2022
  • (FedKD) 提出了一种基于知识蒸馏的联邦跨域自适应的解决方法，将多个本地训练的模型对公开数据集的预测的均值蒸馏给中央模型，以实现知识的转移，其思路可以借鉴在DG上：
    Preserving Privacy in Federated Learning with Ensemble Cross-Domain Knowledge Distillation
    Author：Xuan Gong, Abhishek Sharma, Srikrishna Karanam, Ziyan Wu, Terrence Chen, David Doermann, Arun Innanje
    Association for the Advancement of Artificial Intelligence (AAAI) (2022)
    [Code]
2021

Datasets

We list the widely used benchmark datasets for domain generalization including classification and segmentation.

Dataset Task #Domain #Class #Sample Description
PACS Classification 4 7 9,991 Art, Cartoon, Photo, Sketch
VLCS Classification 4 5 10,729 Caltech101, LabelMe, SUN09, VOC2007
Office-Home Classification 4 65 15,588 Art, Clipart, Product, Real
Office-31 Classification 3 31 4,110 Amazon, Webcam, DSLR
Office-Caltech Classification 4 10 2,533 Caltech, Amazon, Webcam, DSLR
Digits-DG Classification 4 10 24,000 MNIST, MNIST-M, SVHN, SYN
Digit-5 Classification 5 10 ~10,000 MNIST, MNIST-M, SVHN, SYN, USPS
Rotated MNIST Classification 6 10 7,000 Rotated degree: {0, 15, 30, 45, 60, 75}
Colored MNIST Classification 3 2 7,000 Colored degree: {0.1, 0.3, 0.9}
CIFAR-10-C Classification -- 4 60,000 The test data are damaged by 15 corruptions (each with 5 intensity levels) drawn from 4 categories (noise, blur, weather, and digital)
CIFAR-100-C Classification -- 4 60,000 The test data are damaged by 15 corruptions (each with 5 intensity levels) drawn from 4 categories (noise, blur, weather, and digital)
DomainNet Classification 6 345 586,575 Clipart, Infograph, Painting, Quick-draw, Real, Sketch
miniDomainNet Classification 4 345 140,006 A smaller and less noisy version of DomainNet including Clipart, Painting, Real, Sketch
VisDA-17 Classification 3 12 280,157 3 domains of synthetic-to-real generalization
Terra Incognita Classification 4 10 24,788 Wild animal images taken at locations L100, L38, L43, L46
Prostate MRI Medical image segmentation 6 -- 116 Contains prostate T2-weighted MRI data from 6 institutions: Site A~F
Fundus OC/OD Medical image segmentation 4 -- 1060 Contains fundus images from 4 institutions: Site A~D
[GTA5-Cityscapes](GTA5 and Cityscapes) Semantic segmentation 2 -- 29,966 2 domains of synthetic-to-real generalization

Libraries

We list the libraries of domain generalization.

Other Resources

  • A collection of domain generalization papers organized by amber0309.
  • A collection of domain generalization papers organized by jindongwang.
  • A collection of papers on domain generalization, domain adaptation, causality, robustness, prompt, optimization, generative model, etc, organized by yfzhang114.
  • A collection of awesome things about domain generalization organized by junkunyuan.

Contact

  • If you would like to add/update the latest publications / datasets / libraries, please directly add them to this README.md.
  • If you would like to correct mistakes/provide advice, please contact us by email ([email protected]).
  • You are welcomed to update anything helpful.

Acknowledgements

About

This is a repository for organizing papers ,codes, and etc related to Domain Generalization

Resources

Readme

License

MIT license
Activity

Stars

22 stars

Watchers

2 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

No packages published