Object re-identification with distribution corrected ranking list

Abstract

Object re-identification is a retrieval task to identify the same identity, usually utilizing the ranking list to sort the features. However, most existing ranking lists only calculate the distance from the probe sample to its k-nearest neighbor samples in the gallery set but ignore the distribution of the gallery set. Meanwhile, we find that the gallery set tends to converge into different groups. When the probe lies near the boundary, its k-nearest neighbors often contain samples of different identities from the probe sample. This paper presents a Distribution Corrected Ranking List (DCRL) to correct the ranking by considering the distance between the probe sample and its nearest distribution (i.e., the cluster center). Specifically, given a probe sample, first, we cluster the gallery samples with the agglomerative hierarchical clustering method to obtain the cluster centers. Second, we compute the distance between the probe and the cluster centers. Third, the final ranking list is considered as the weighted combination of the original ranking list (i.e., the distance between the probe sample and its k-nearest neighbor in the gallery) and the distribution ranking list (i.e., the distance between the probe sample and its nearest cluster center). The proposed distribution corrected ranking list is suitable for large datasets because it does not require human intervention or labeled data. Experimental results on several benchmark re-identification datasets, such as Market-1501, DukeMTMC-reID, MSMT17, VeRi-776, and VeRi-Wild, show that the proposed method achieves state-of-the-art performance.