Distilling consistent relations for multi-source domain adaptive person re-identification

Abstract

Although unsupervised person re-identification has made great progress by employing domain adaptation, existing methods rarely make use of multiple source datasets. These works neglect a practical scenario where annotated data from multiple domains are available. To address this issue, we propose a novel method to exploit the diversity and consistency of multiple sources to improve domain adaptation for person re-identification. In the proposed method, multiple expert models pre-trained in different source domains are adapted to the target domain by self-training with expert-specific clustering-based pseudo labels. To improve the consistency of multiple experts, we transfer knowledge between experts through dual similarity distillation between experts. Finally, to further encourage experts to learn complementary features, we maintain the feature diversity of multiple experts by representation decorrelation, preventing them from being too similar. Extensive experiments conducted on three benchmarks, Market-1501, DukeMTMC-reID, and MSMT17, demonstrate the effectiveness of the proposed method.