An Entropy-Based Pseudo-Label Mixup Method for Source-Free Domain Adaptation

Qinghan Chen, Zhiyang Lu, Ming Cheng

Abstract: Domain adaptation (DA) is widely used in pattern recognition and computer vision. It learns a model from the source domain data and applies it in the target domain. Most of the DA techniques require access to source domain data. However, it could be impossible in some cases due to privacy concerns or license agreements. Some recent approaches address this issue through the source-free domain adaptation (SFDA) technique using pseudo labels. However, their performance is greatly affected by the quality of pseudo-label generation. In this work, we propose an entropy-based SFDA method that can generate high-quality pseudo labels. Furthermore, a pseudo-label mixup technique is designed to reduce the gap among similar data while increasing the distances among dissimilar data. We have shown through extensive experiments that our approach can achieve better results than state-of-the-art methods.