Denoised Maximum Classi?er Discrepancy for Source-Free Unsupervised Domain Adaptation

Source-Free Unsupervised Domain Adaptation(SFUDA) aims to adapt a pre-trained source model an unlabeled target domain without access the original labeled samples. Many existing SFUDA approaches apply self-training strategy, which involves iteratively selecting confidently predicted samples as pseudo-labeled used train fit domain. However, strategy may also suffer from sample selection bias and be impacted by label noise of In this work, we provide rigorous theoretical analysis on how these two issues affect generalization ability when applying for problem. Based analysis, then propose new Denoised Maximum Classifier Discrepancy (D-MCD) method effectively address issues. particular, first minimize distribution mismatch between selected remaining alleviate bias. Moreover, design strong-weak paradigm denoise samples, where strong network is select while weak helps filter out hard avoid incorrect labels. way, are able ensure both quality pseudo-labels trained We achieve state-of-the-art results three adaptation benchmark datasets, clearly validates effectiveness our proposed approach. Full code available at https://github.com/kkkkkkon/D-MCD.