Adaptive pseudo-label threshold for source-free domain adaptation

Source-free domain adaptation (SFDA) endeavors to utilize a source domain model for solving analogous tasks in a novel unlabeled domain, even when the source data remains inaccessible. However, existing SFDA methods ignore the class imbalance and unlabeled data noise, referred to as the class gap, which causes evident differences in the adapting difficulties of classes. To address the above limitations, we propose a novel threshold adaptive strategy (TAS) for SFDA. Specifically, to mitigate the negative effects of class imbalance, we design a Threshold Adapter that automatically modulates domain and class thresholds. Then, to counteract the unlabeled data noise, the Consistency Constraint Learner is employed to enhance the similarity between different augmentations of the same sample. In this way, the class gap can be efficiently spanned by adjusting the pseudo-label threshold and consistency constraints. Extensive experimental evaluations underscore the superiority of our method, positioning it as a robust baseline for future research in SFDA, and our approach extends beyond single-source scenarios to encompass multisource, multi-target, partial-set, and open-set benchmarks.