Attribute Auxiliary Clustering for Person Re-Identification

Abstract

The main objective of the person re-identification task is to retrieve the specific identity under multiple non-overlapping camera scenarios. Though unsupervised person re-ID has already achieved great performance and even surpasses some classic supervised re-ID methods, the existing methods pay much attention to training the neural networks with the memory-based idea which ignore the quality of the generated pseudo label. The quality of the clustering process does not only depend on the intra-cluster similarity but also on the number of clusters. In this paper, our approach employs an attribute auxiliary clustering method for person re-ID task. The proposed method could divide the generated cluster by the leveraged attribute label. Employed the attribute auxiliary clustering, the task changed from unsupervised case to weakly supervised case. The method is compared with state-of-the-art and analyzes the effectiveness caused by the variation of the cluster number. The proposed approach achieves great performance on the public Market-1501 datasets.