Uncertainty-Aware Clustering for Unsupervised Domain Adaptive Object Re-Identification

Unsupervised Domain Adaptive (UDA) object re-identification (Re-ID) aims at adapting a model trained on labeled source domain to an unlabeled target domain. State-of-the-art Re-ID approaches adopt clustering algorithms generate pseudo-labels for the However, inevitable label noise caused by procedure significantly degrades discriminative power of model. To address this problem, we propose uncertainty-aware framework (UCF) UDA tasks. First, novel hierarchical scheme is proposed promote quality. Second, collaborative instance selection method introduced select images with reliable labels training. Combining both techniques effectively reduces impact noisy labels. In addition, introduce strong baseline that features compact contrastive loss. Our UCF consistently achieves state-of-the-art performance in multiple tasks Re-ID, and gap between unsupervised supervised Re-ID. particular, our MSMT17->Market1501 task better than fully setting Market1501. The code available https://github.com/Wang-pengfei/UCF.