Biased metric learning for person-independent head pose estimation

Abstract

In this paper, a new biased metric learning (BML) method is proposed for human head pose estimation problem. Traditional approaches focus on modeling a smooth low-dimensional manifold embedded in the high dimensional feature space. Such manifold-embedding methods, linear or nonlinear, suffer from one common drawback, that all neighbors are identified based on the Euclidean distance in the original feature space. However, the nature local structure of data is always corrupted by various factors in this original feature space. The proposed BML method aims at obtaining a global optimal linear transformation from the input feature space into a new semantic space which is characterized by pose angles. The metric is trained with the goal that local semantic structure of data with same label is preserved while the biased distance of differently labeled data is maximized. The learning process also reduces to a convex optimization by formulating it as a semidefinite problem (SDP). Numerous experiments demonstrate the superiority of our BML method over several current states of art approaches on publicly available dataset.