Identity Retaining and Redundancy Reducing Gan for Person Re-Identification

Abstract

Person re-identification (ReID) models trained on one domain suffer performance degradation when tested on other domains. The existing works address this problem by domain translation with identity information preserving. However, these methods focused on adding pixel constraints to preserve identity, which also preserves a lot of redundant information. Therefore, this paper propose an identity retaining and redundancy reducing generative adversarial network (IRGAN), a domain translation method for person ReID. IRGAN is implemented by an unequal-cycle strategy, which imposes both foreground and feature constraints to domain translation. By imposing part-level feature constraints, the redundant information generated by pixel constraints can be reduced. Thus the performance of the domain translation is significantly improved. Experimental results indicate that our method is effective.