Regularized Integrated Metric for Person Re-Identification

Abstract

Integrated metric combining both difference and commonness of image pairs has shown to achieve superior performance over difference-only based metrics in similarity learning. The integrated metric can be learned quickly by computing the log-likelihood ratio between the probability distribution functions of similar and dissimilar image pairs. Under pair constrained Gaussian assumption, the learning involves the computation of inverse of covariance matrices using maximum likelihood criterion which may lead to a degraded solution if proper regularization is not done. In this paper, we study the influence of regularization in integrated metric learning. Person re-identification is chosen as the target application to demonstrate the effectiveness of the regularized metric. Moreover, comparison with recent methods on challenging benchmark datasets in the domain of person reidentification like VIPeR and PRID450S shows that our method achieves better or comparable re-identification rates than other methods.