Domain-Camera Adaptation for Unsupervised Person Re-Identification

Abstract

Although supervised person re-identification (Re-ID) performance has been significantly improved in recent years, it is still a challenge for unsupervised person Re-Iddue to its absence of labels across disjoint camera views. On the other hand, Re-Idmodels trained on source domain usually offer poor performance when they are tested on target domain due to inter-domain bias e.g. different classes and intra-domain difference e.g camera variance. To overcome this problem, given a labeled source training domain and an unlabeled target training domain, we propose an unsupervised transfer method, Domain-Camera Adaptation model, to generate a pseudo target domain by bridging inter-domain bias and intra-domain difference. The idea is to fill the absence of labels in target domain by transferring labeled images of source domain to target domain across cameras. Then we propose a cross-domain classification loss to extract discriminative representation across domains. The intuition is to think of unsupervised learning as semi-supervised learning in target domain. We evaluate our deep model on Market-1501 and DukeMTMC-reID and the results show our model outperforms the state-of-art unsupervised Re-ID methods by large margins.