USDAP: universal source-free domain adaptation based on prompt learning

Abstract

Universal source-free domain adaptation (USFDA) aims to explore transferring domain-consistent knowledge in the presence of domain shift and category shift, without access to a source domain. Existing works mainly rely on prior domain-invariant knowledge provided by the source model, ignoring the significant discrepancy between the source and target domains. However, directly utilizing the source model will generate noisy pseudo-labels on the target domain, resulting in erroneous decision boundaries. To alleviate the aforementioned issue, we propose a two-stage USFDA approach based on prompt learning, named USDAP. Primarily, to reduce domain differences, during the prompt learning stage, we introduce a learnable prompt designed to align the target domain distribution with the source. Furthermore, for more discriminative decision boundaries, in the feature alignment stage, we propose an adaptive global-local clustering strategy. This strategy utilizes one-versus-all clustering globally to separate different categories and neighbor-to-neighbor clustering locally to prevent incorrect pseudo-label assignments at cluster boundaries. Based on the above two-stage method, target data are adapted to the classification network under the prompt’s guidance, forming more compact category clusters, thus achieving excellent migration performance for the model. We conduct experiments on various datasets with diverse category shift scenarios to illustrate the superiority of our USDAP.