Integrated Global-Local Metric Learning for Person Re-identification

The task of person re-identification (re-id) is to match images of people observed in different camera views. Recent researches mainly focus on feature representation and metric learning. Many global metric learning approaches have achieved good performance. Since comparing all of the samples with a single global metric is inappropriate to handle heterogeneous data, some local metric learning approaches are proposed. But most of them cannot be used on re-id directly due to some research challenges. Also, they usually need complicated computation to solve the optimization problems with numerous parameters. In order to improve the performance of global metric learning and avoid complex computation, we propose to simultaneously learn local metrics on clusters of samples softly partitioned by Gaussian Mixture Model (GMM) and a global metric on the entire training set. Then the local metrics are combined with the global metric by their posterior probabilities of GMM to obtain an integrated metric for similarity evaluation. Experiments on three challenging datasets (VIPeR, PRID450S and QMUL GRID) verify the effectiveness of the proposed method.