Global–local Bi-alignment for purer unsupervised domain adaptation

Unsupervised domain adaptation (UDA) aims to extract domain-invariant features. Existing UDA methods mainly utilize a convolutional neural network (CNN) or vision transformer (ViT) as feature extractor, which align in the latent space that characterizes single view of the object and may lead to matching error—the distributions aligned in the CNN space may still be confused in the ViT space. To address this, we introduce global–local bi-alignment (GLBA) based on a hybrid structure Conformer, which enforces simultaneous alignment in both spaces, following the space-independent assumption: If two domains have the same distribution, their distributions in any latent space are aligned. The framework can be easily combined with previous UDA methods, essentially adding only an alignment loss without the need for elaborate structures or large numbers of parameters. Experiments demonstrate the effectiveness of GLBA and its state-of-the-art (SoTA) performance achieved with comparable parameter complexity. Code is available at https://github.com/JSJ515-Group/GLBA.