Unsupervised adversarial domain adaptation with similarity diffusion for person re-identification

Due to the scarcity of human identities labels, unsupervised person re-identification (re-ID) draws much attention recently, which attempts to learn discriminative person representation without labels. Domain adaptation based methods utilize the labeled sample in source domain and transfer the knowledge to the unlabeled target domain. However, for person re-ID, no identities overlapping between source and target domain leads to the difficulty during adaptation. To address this problem, we propose an unsupervised adversarial domain adaptation method for person re-ID which exploits the pixel level and feature level alignments. Specifically, we utilize CycleGAN to transfer the target domain style to the source domain for the pixel level, and we propose an adversarial metric adaptation method which aligns between the source domain and target domain for the feature level. We further explore the feature similarity laying on manifold structure revealed by the features through the similarity diffusion. To verify the efficacy of our proposed method, we conduct extensive experiments on three benchmark datasets: Market1501, Duke-MTMC, and MSMT17. Comparing with state-of-the-art unsupervised domain adaptation approaches, we have comparable performance on ranking metric and significant improvement on mAP metric, which validates the efficacy of the proposed technique for person re-identification tasks.