Multi-View and Multi-Information Clustering for Semi-Supervised Person Re-Identification

Abstract

Deep learning based methods for person re-identification (re-id) have aroused extensive attention in recent years. However, most works adopt fully-supervised learning, which heavily rely on a large amount of labeled training data. And collecting labeled samples is quite time consuming. To address this problem, we present a semi-supervised framework for person re-id. The key point in this work is to estimate the label of unlabeled data, thus a multi-view and multi-information clustering (MVMIC) method is proposed. First, multi-view feature representation is obtained by two Convolutional Neural Networks, then KNN graphs can be constructed by the feature representation. Finally, multi-information is collected from the KNN graphs to select positive pairs and clustering will be achieving. Experimental results on two large-scale datasets demonstrate the superiority of the proposed method.