Confidence Score Weighting Adaptation for Source-Free Unsupervised Domain Adaptation

Unsupervised domain adaptation (UDA) aims to achieve high performance within the unlabeled target domain by leveraging the labeled source domain. Source-free UDA, which is a more challenging UDA task, can access the pre-trained model within the source domain. The pre-trained model, however, provides a noisy pseudo-label; thus, source-free UDA requires robust training. In this study, we propose a Confidence score Weighting Adaptation (CoWA), which is a simple yet effective source-free UDA method. CoWA utilizes the Joint Model-Data Structure (JMDS) confidence score designed for source-free UDA as a sample-wise weight. As components of CoWA, we introduce Suppressed Cross Entropy (SCE) loss and a weight mixup to robustly leverage the low-confidence samples. Experiment results show that CoWA achieves a superior performance compared to other source-free UDA methods on various UDA benchmarks including open-set and partial-set domain adaptation. Furthermore, on several benchmarks, CoWA surpasses state-of-the-art conventional UDA methods that use labeled source domain data.