Two-stage structural information enhancement for source-free domain adaptation

Source-free domain adaptation (SFDA) uses models trained from source domains to solve similar tasks in unlabeled domains, without accessing source domain data. Existing SFDA methods have not been able to learn spatial and semantic structural information of target domains simultaneously, making them insufficient and inefficient for target domain exploration. To fully explore the target domain structural information, we propose a novel representation learning framework, called structural information enhancement (SIE). SIE has a two-stage approach that, in the first stage, clusters local neighbors and pushes away global non-neighbors in the feature space to obtain spatial structural information. In the second stage, SIE fine-tunes the clustered model using a semantic structure consistency strategy that exploits semantic structural information by mutual learning interpolated sample pairs. Our extensive experiments demonstrate the superiority of our method, and our method can serve as a strong baseline for future SFDA research.