Source-free semi-supervised domain adaptation via progressive Mixup

Abstract

Existing domain adaptation methods usually perform explicit representation alignment by simultaneously accessing the source data and target data. However, the source data are not always available due to the privacy preserving consideration or bandwidth limitations. To address this issue, source-free domain adaptation is proposed to perform domain adaptation without accessing the source data. Recently, the adaptation paradigm is attracting increasing attention, and multiple works have been proposed for unsupervised source-free domain adaptation. However, without utilizing any supervised signal and source data at the adaptation stage, the optimization of the target model is unstable and fragile. To alleviate the problem, we focus on utilizing a few labeled target data to guide the adaptation, which forms our method into semi-supervised domain adaptation under a source-free setting. We propose a progressive data interpolation strategy including progressive anchor selection and dynamic interpolation rate to reduce the intra-domain discrepancy and inter-domain representation gap. Extensive experiments on three public datasets demonstrate the effectiveness as well as the better scalability of our method.