Robust label propagation based on prior-guided cross domain data augmentation for few-shot unsupervised domain adaptation

Abstract

Few-shot unsupervised domain adaptation (FS-UDA) aims to leverage knowledge from an imbalanced, labeled source domain and apply it to an unlabeled target domain. The primary difficulties of FS-UDA stem from the disparity in data distributions across source and target domains, coupled with uneven class representation in the source data. Label propagation (LP) is commonly used in domain adaptation scenarios. However, in FS-UDA tasks, LP disproportionately favors the normal classes because the source domain suffers from imbalanced class distribution, which results in insufficient feature representation and a large domain gap for the few-shot classes. To tackle these problems, we introduce a new robust LP approach that leverages prior-guided cross-domain data augmentation for FS-UDA. Unlike conventional approaches that solely utilize source domain visual data for few-shot class augmentation, our proposed method employs contrastive language image pretraining-derived semantic priors to supervise visual feature extractor training and optimize few-shot prototypes. It enhances domain-invariant feature learning while mitigating cross-domain distribution mismatches. We introduce the visual information from the target domain to perform data augmentation via style transfer, obtaining more diverse class-specific information. Subsequently, we capture intradomain and interdomain relationships more accurately by constructing intradomain and interdomain graphs independently for all samples (original and augmented) from both domains, which facilitates more effective LP and makes LP robust to few-shot classes. Furthermore, we introduce an adaptive graph regularization loss to dynamically adjust class weights, enhance intraclass compactness within domains, and reduce intraclass distribution discrepancies between different domains. Comprehensive experiments validate that the proposed method achieves superior performance compared to existing state-of-the-art methods across various FS-UDA tasks. The proposed method achieves 77.3 % and 61.7 % average accuracies for few-shot classes on the Office-31 and Office-Home datasets, respectively.