TransAdapter: Vision Transformer for Feature-Centric Unsupervised Domain Adaptation

Abdullah Enes Doruk, Erhan Oztop, Hasan Fehmi Ates

Keywords: Unsupervised Domain Adaptation, Transformer, Domain-Invariant Feature Representations

Abstract: Unsupervised Domain Adaptation (UDA) aims to leverage labeled data from a source domain to address tasks in a related but unlabeled target domain. This problem is particularly challenging when there is a significant gap between the source and target domains. Traditional methods have largely focused on minimizing this domain gap by learning domain-invariant feature representations using convolutional neural networks (CNNs). However, recent advances in vision transformers, such as the Swin Transformer, have demonstrated superior performance in various vision tasks. In this work, we propose a novel UDA approach based on the Swin Transformer, introducing three key modules to improve domain adaptation. First, we develop a Graph Domain Discriminator that plays a crucial role in domain alignment by capturing pixel-wise correlations through a graph convolutional layer, operating on both shallow and deep features in the transformer. This module also calculates the entropy for the key attention features of the attention block to better distinguish between the source and target domains. Second, we present an Adaptive Double Attention module that simultaneously processes Windows and Shifted Windows attention to increase long-range dependency features. An attention reweighting mechanism is employed to dynamically adjust the contributions of the attention values, thereby improving feature alignment between domains. Finally, we introduce Cross-Feature Transform, where random Swin Transformer blocks are selectively transformed using our proposed transform module, enhancing the model’s ability to generalize across domains by transferring the source to the target features. Extensive experiments demonstrate that our method improves the state-of-the-art on several challenging UDA benchmarks, confirming the effectiveness of our approach. In particular, our model does not include a task-specific domain alignment module, making it more versatile for various applications.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7483