Gallery based k-reciprocal-like re-ranking for heavy cross-camera discrepancy in person re-identification

Abstract

Recently more and more attention is paid to the re-ranking step for person re-identification in computer vision community, especially those fully automatic, unsupervised solutions. Among them, k-reciprocal re-ranking (KR) method achieved big success. However, when the heavy cross-camera discrepancy exists between query and gallery datasets, it may degrade the performance. To alleviate the heavy cross-camera discrepancy between query and gallery datasets, we propose a gallery based k-reciprocal-like re-ranking (GKR) method. GKR adopts graph matching to construct the matching correspondence between query and gallery datasets. Then the proposed k-reciprocal-like neighbors are computed only on gallery dataset instead of on the union of query and gallery datasets like KR does. Moreover, to perform unsupervised video-based person re-identification, we incorporate our proposed GKR method into the dynamic label graph matching (DGM) framework, which can improve the cross-camera labels estimating in training step but also can improve the re-identification accuracy by re-ranking in testing step. Experimental results by supervised and unsupervised solutions on some benchmarks, demonstrate the effectiveness of our GKR method to handle the cross-camera discrepancy problem for person re-identification.