The official codes of paper: Adversarial Style Augmentation for Domain Generalization
One Sentence Summary: AdvStyle explores a broader style space over MixStyle, DSU, and EFDMix by searching for the most challenging domains via adversarial training.
 |
|---|
| Fig.1: MixStyle vs. DSU vs. AdvStyle |
To reproduce our results on cross-domain image classification, and cross-domain person re-identification,
please find the code in ./imcls, and ./reid, respectively.
This work was initially finished in Mar. 2022 and submitted to ECCV2022 and AAAI2023. The corresponding review can be found at: ECCV2022_review and AAAI2023_review. Considering that a similar idea to AdvStyle has been published in NIPS2022, we just remain this paper as a Technique Report for the reference of the community.
To cite AdvStyle in your publications, please use the following bibtex entry:
@article{zhang2023adversarial,
title={Adversarial Style Augmentation for Domain Generalization},
author={Zhang, Yabin and Deng, Bin and Li, Ruihuang and Jia, Kui and Zhang, Lei},
journal={arXiv preprint arXiv:2301.12643},
year={2023}
}