Multi source-free domain adaptation based on pseudo-label knowledge mining

Abstract

MSFDA methods were proposed to train unlabeled target data using a group of source pre-trained models, without directly accessing labeled source domain data. Through transferring knowledge to target domain using pseudo labels obtained by source pre-trained models, existing methods have shown potential for cross-domain classification. However, these models have not directly addressed the negative knowledge transfer caused by incorrect pseudo labels. In this study, we focus on the problem and propose a multi-source-free domain adaptation method based on pseudo-label knowledge mining. Specifically, we first utilize average entropy weighting to compute pseudo labels for target data. Then, we assign a confidence level to each target sample, considering it as either high or low. Finally, we generate mixed augmented target samples and conduct different self-training tasks for those with different confidence to alleviate the negative transfer resulting from inaccurate pseudo labels. Experimental results on three datasets demonstrate the effectiveness of our proposed method.