CDTrans: Cross-domain Transformer for Unsupervised Domain Adaptation

2022년 04월 01일 (금) | 발표자: 김승우

This paper proposes a pure transformer-based model CDTrans with two-way center-aware pseudo labeling. The authors mainly claim that the cross-attention module in Transformer is robust to noisy inputs and focus on aligning pseudo labels from different domains. CDTrans achieved better performances than previous SOTA models.