Energy Clustering for Unsupervised Person Re-identification

Due to the high cost of data annotation in supervised learning for person re-identification (Re-ID) methods, unsupervised learning becomes more attractive in the real world. The Bottom-up Clustering (BUC) approach based on hierarchical clustering serves as one promising unsupervised clustering method. One key factor of BUC is the distance measurement strategy. Ideally, the distance measurement should consider both inter-cluster and intra-cluster distance of all samples. However, BUC uses the minimum distance, only considers a pair of the nearest sample between two clusters and ignores the diversity of other samples in clusters. To solve this problem, we propose to use the energy distance to evaluate both the inter-cluster and intra-cluster distance in hierarchical clustering(E-cluster), and use the sum of squares of deviations(SSD) as a regularization term to further balance the diversity and similarity of energy distance evaluation. We evaluate our method on large scale re-ID datasets, including Market-1501, DukeMTMC-reID and MARS. Extensive experiments show that our method obtains significant improvements over the state-of-the-art unsupervised methods, and even better than some transfer learning methods.