Common-covariance based person re-identification model

Abstract

Abstract This paper develops a new covariance-matrix based person re-identification model. It takes each person as a Gaussian distribution with a common covariance matrix. The learning process is to find an optimal covariance matrix among all the distributions. This paper has two major contributions: (1) It proves the statistical meaning of Mahalanobis matrix as a common covariance matrix among all related classes; (2) The intra-class distance is measured using the determinant of the common covariance matrix. The proposed model largely reduces the computational cost. Moreover, it automatically avoids imbalanced computations between the distances for inter-class images and the distances for intra-class images. Experimental results demonstrate that the proposed model is more efficient than many state-of-the-art methods in both accuracy and time cost.