Regularized Bayesian Metric Learning for Person Re-Identification

Abstract

Person re-identification across disjoint cameras has attracted increasing interest in computer vision due to its wide potential applications in visual surveillance. In this paper, we propose a new regularized Bayesian metric learning (RBML) method for person re-identification. While numerous metric learning methods have been proposed for person re-identification in recent years, most of them suffer from the small sample size (SSS) problem because there are not enough training samples in most practical person re-identification systems, so that the within-class and between-class variations can be well estimated to learn the distance metric. To address this, we propose a RBML method to model and regulate the eigen-spectrums of these two covariance matrices in a parametric manner, so that discriminative information can be better exploited. Experimental results on three widely used datasets demonstrate the advantage of our proposed RBML over the state-of-the-art person re-identification methods.