Token Calibration for Transformer-Based Domain Adaptation

Unsupervised Domain Adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain by learning domain-invariant representations. Motivated by the recent success of Vision Transformers (ViTs), several UDA approaches have adopted ViT architectures to exploit fine-grained patch-level representations, which are unified as Transformer-based $D$ omain $A$ daptation (TransDA) independent of CNN-based. However, we have a key observation in TransDA: due to inherent domain shifts, patches (tokens) from different semantic categories across domains may exhibit abnormally high similarities, which can mislead the self-attention mechanism and degrade adaptation performance. To solve that, we propose a novel $P$ atch- $A$ daptation Transformer (PATrans), which first identifies similarity-anomalous patches and then adaptively suppresses their negative impact to domain alignment, i.e. token calibration. Specifically, we introduce a $P$ atch- $A$ daptation $A$ ttention (PAA) mechanism to replace the standard self-attention mechanism, which consists of a weight-shared triple-branch mixed attention mechanism and a patch-level domain discriminator. The mixed attention integrates self-attention and cross-attention to enhance intra-domain feature modeling and inter-domain similarity estimation. Meanwhile, the patch-level domain discriminator quantifies the anomaly probability of each patch, enabling dynamic reweighting to mitigate the impact of unreliable patch correspondences. Furthermore, we introduce a contrastive attention regularization strategy, which leverages category-level information in a contrastive learning framework to promote class-consistent attention distributions. Extensive experiments on four benchmark datasets demonstrate that PATrans attains significant improvements over existing state-of-the-art UDA methods (e.g., 89.2% on the VisDA-2017). Code is available at: https://github.com/YSY145/PATrans.