Reducing Negative Transfer of Random Data in Source-Free Unsupervised Domain Adaptation
Abstract
In domain adaptation, a model trained on one dataset (source domain) is applied to a different but related dataset (target domain). The most cutting-edge method is unsupervised source-free domain adaptation (SFDA), in which source data, source labels, and target labels are not available during adaptation. This thesis explores a realistic scenario where the target dataset includes some images that are unrelated to the adaptation process. This scenario can occur from errors in data collection or processing. We provide experiments and analysis to show that current state-of-the-art (SOTA) SFDA methods suffer significant performance drops under a specific domain adaptation setup when tested on our proposed scenario. To resolve this problem, we propose a new domain adaptation framework that integrates anomaly detection into the adaptation process which removes samples irrelevant to adaptation. Our framework greatly surpasses the performance of current SOTA methods, making it highly applicable in real-world uses of SFDA. In addition, future SFDA research should be benchmarked against our proposed scenario as a better measure of its real-world performance.