RULER: Source-free Domain Adaptive Person Re-identification via Uncertain Label Refinery

Source-free domain adaptive person re-identification (ReID) aims to address the cross-domain person ReID task with a well-trained source model, which solves the limitations of data privacy and transmission costs in real-world scenarios. Existing methods mainly generate pseudo labels for target data, which are unreliable because of distribution shifts. First, the ubiquitous difficult samples may lead to the ambiguity of the model prediction. Second, the source model may have a bias towards certain classes. To alleviate these two problems, we propose a source-free domain adaptive person ReID method via the uncertain label refinery (RULER). RULER consists of uncertainty-aware pseudo-labeling refinery (UPLR) and frequency-weighted contrastive learning (FCL). To reduce the ambiguity of predictions caused by sample label uncertainty, UPLR generates pseudo labels by clustering samples after multiple random dropouts and then integrates the results to obtain high-confidence pseudo labels. Moreover, FCL defines the frequency of each class as the sample weight and introduces a frequency-weighted contrastive loss to alleviate the class biases of the model. RULER improves the quality of pseudo labels and mitigates the source model′s bias towards certain classes. We achieve competitive results compared to state-of-the-art methods on both real-to-real and synthetic-to-real source-free domain adaptation scenarios, validating the effectiveness of RULER.