Sparsity-Regularized Geometric Mean Metric Learning for Kinship Verification

Yunhao Xu, Junlin Hu

Abstract: Kinship verification through face images is a challenging research problem in biometrics. In this paper, we propose a sparsity-regularized geometric mean metric learning (SGMML) method to improve the well-known geometric mean metric learning (GMML) method and apply it to kinship verification task. Unlike GMML method that utilizes a linear map with fixed dimension, our SGMML method is capable of automatically learning the best projection dimension by employing the sparsity constraints on Mahalabios metric matrix. The proposed SGMML can effectively tackle the over-fitting problem and the data mixing up problem in the projected space. We conduct experiments on two benchmark kinship verification datasets, and experimental results demonstrate the effectiveness of our SGMML approach in kinship verification.