Reproduce “Exploiting Both Domain-specific and Invariant Knowledge via a Win-win Transformer for Unsupervised Domain Adaptation” on OpenResearch

About this paper

Wenxuan Ma, Jinming Zhang, Shuang Li, Chi Harold Liu, Yulin Wang, Wei Li

Researchers from Beijing Institute of Technology, Tsinghua University, and Inceptio Tech developed WinTR, a Transformer-based framework for unsupervised domain adaptation that simultaneously leverages domain-specific and domain-invariant knowledge. This method achieved state-of-the-art results on VisDA-2017 (90.1%), Office-Home (77.2%), and DomainNet (46.8%) benchmarks, outperforming previous Transformer and CNN-based approaches.

Key insights

• Vision Transformers exhibit superior inherent transferability compared to CNNs, making them a strong backbone for UDA tasks.
• Explicitly disentangling and exploiting both domain-specific and domain-invariant knowledge via dedicated tokens and classifiers within a Transformer architecture leads to a synergistic "win-win" outcome in UDA.
• Source-oriented target features, supervised by clean source labels, retain finer structural relationships that are beneficial for refining noisy target pseudo-labels.

Summary by alphaXiv.