QRKISS: A Two-Stage Metric Learning via QR-Decomposition and KISS for Person Re-Identification

Abstract

Person re-identification is a challenging task in the field of intelligent video surveillance because there are wide variations between pedestrian images. As a classical metric learning method, Keep It Simple and Straightforward (KISS) has shown good performance for person re-identification. However, when the dimension of data is high, the KISS method may perform poorly because of small sample size problem. A common solution to this problem is to apply dimensionality reduction technologies to original data before the KISS metric learning, such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). In this paper, to learn a discriminant and robust metric, we propose a novel two-stage metric learning via QR-Decomposition and KISS, named QRKISS. The first stage of QRKISS is to project original data into a lower dimensional space by QR decomposition. In this lower dimensional space, the trace of the covariance matrix of interpersonal differences can reach maximum. Based on KISS method, the second stage of QRKISS obtains a Mahalanobis matrix in the low-dimension space. We conduct thorough validation experiments on the VIPeR, PRID 450S and CUHK01 datasets, which demonstrate that QRKISS method is better than other KISS-based metric learning methods and achieves state-of-the-art performance.