Asymmetric Distance Learning for Unsupervised Video Person Re-Identification with Tracklet Neighborhood Re-Ranking

Abstract

The gruelling human-annotation and lack of sufficient labeled data make unsupervised person re-identification (re-ID) an important component in research. This paper proposes a re-ID system for unsupervised video-based re-ID, which mainly contains an asymmetric distance learning approach and a re-ranking meth-od. Specifically, using the sequence information provided by video, asymmetric learning makes a distinctive projection for features in each view, while label estimation makes this procedure efficient and effective. To further refine the results of the ranking list, an unsupervised re-ranking technique based on the already computed distance is introduced to the system. We show that both of our asymmetric distance learning and re-ranking method have achieved state-of-the-art performance on PRID-2011, iLIDS-VID and MARS datasets, meanwhile restrains the computational costs. The experiments show that our asymmetric learning method is suitable for video-based re-ID with multiple cameras, and the proposed re-ranking method is a good solution to refine the ranking list for video-based re-ID.