Test-Time Domain Adaptation by Learning Domain-Aware Batch Normalization

AAAI 2024 (Oral)

(Code coming soon)

Background

We focus on the problem of Test-time Domain Adaptation (TTDA) or Few-shot TTDA. When an unseen target domain is encountered at test-time, a few unlabeled images are sampled to update the model towards that domain. The adapted model is then used for testing the data in that domain.

Method overview

⭐ Acknowledgement

Our code is built upon the codebase from MetaDMoE (NeurIPS22)

📋 Citation

If you use this code in your research, please consider citing our paper:

@inproceedings{wu2024test,
  title={Test-Time Domain Adaptation by Learning Domain-Aware Batch Normalization},
  author={Wu, Yanan and Chi, Zhixiang and Wang, Yang and Plataniotis, Konstantinos N and Feng, Songhe},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2024}
}