Ring-push metric learning for person reidentification

Abstract

Person reidentification (re-id) has been widely studied because of its extensive use in video surveillance and forensics applications. It aims to search a specific person among a nonoverlapping camera network, which is highly challenging due to large variations in the cluttered background, human pose, and camera viewpoint. We present a metric learning algorithm for learning a Mahalanobis distance for re-id. Generally speaking, there exist two forces in the conventional metric learning process, one pulling force that pulls points of the same class closer and the other pushing force that pushes points of different classes as far apart as possible. We argue that, when only a limited number of training data are given, forcing interclass distances to be as large as possible may drive the metric to overfit the uninformative part of the images, such as noises and backgrounds. To alleviate overfitting, we propose the ring-push metric learning algorithm. Different from other metric learning methods that only punish too small interclass distances, in the proposed method, both too small and too large inter-class distances are punished. By introducing the generalized logistic function as the loss, we formulate the ring-push metric learning as a convex optimization problem and utilize the projected gradient descent method to solve it. The experimental results on four public datasets demonstrate the effectiveness of the proposed algorithm.