Attribute Memory Transfer Network for Unsupervised Cross-Domain Person Re-Identification

Person re-identification (re-ID) presents various applications in surveillance system, but most existing models are proposed under supervised framework. These methods require large amounts of annotated pedestrian data, which limits their scalability and flexibility in a new application scenario. Aiming to relax this limitation, this article exploits the attribute-invariant characteristics and domain correlations into cross-domain person re-ID, while most unsupervised methods only consider the identity features and ignore the different importance of each source image to the target domain. Specifically, this article proposes an Attribute Memory Transfer Network (AMTNet) with two major contributions of domain-balanced memory and attribute-invariant memory modules. The first domain balanced-memory integrates a domain correlation learning method to evaluate the importance of each source image to the target domain, which is involved into the transfer learning; The second attribute-invariant memory can transfer the source attribute knowledge into the target domain with preserving the identity information to conduct the re-ID process. Extensive evaluated experiments elaborate the superiority of AMTNet on two large datasets of Market-1501 and DukeMTMC-reID, compared with hand-crafted and deep learning feature-based methods.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.