Bi-directional Re-ranking for Person Re-identification

Abstract

For person re-identification, previous re-ranking methods focus on the unidirectional query-find-gallery ranking list and target to improve the performance of person re-identification. However, the matched images with the same identity may get lower ranks in the query-find-gallery ranking list, which limits the improvement of these re-ranking methods. To solve this problem, we propose the Bi-directional re-ranking method. Different from existing methods, we consider the bi-directional matching including the query-find-gallery ranking list and the gallery-find-query ranking list. In addition, we construct the graph of image relationship based on feature distances and expand the qualified images other than the initial top-k nearest images. By combining the bi-directional re-ranking performance and the k-neighbor similarity score, we re-rank the initial ranking list and get higher improvements. Extensive experiments show that the Bi-directional re-ranking method can facilitate the state-of-the-art person re-identification methods.