Re-ranking for person re-identification

Abstract

Person Re-Identification problem aims at matching people across a network of non-overlapping cameras. When multiple probe people appear concurrently, human could compare them together to give a more accurate matching. However, existing approaches treat each probe person independently, skipping the concurrent information. In this paper, we propose a re-ranking method which utilize that kind of information to refine ranked lists produced by any person re-identification method to create more precise ranked lists. The experimental results on VIPeR dataset show the improved performance when our method is applied.