Isomorphic loss function for head pose estimation

Abstract

Accurate head pose estimation is a key step in many practical applications involving face analysis tasks, such as emotion recognition. We address the problem of head pose estimation in still color images acquired with a standard camera with limited resolution details. To achieve the proposed goal, we make use of the recent advances of Deep Convolutional Neural Networks. As head angles with respect on yaw and pitch are continuous, the problem is one of regression. Typical loss function for regression are based on L1 and L2 distances which are notorious for susceptibility to outliers. To address this aspect we introduce an isomorphic transformation which maps the initially infinite space into a closed space compressed at the ends and thus significantly down–weighting the significance of outliers. We have thoroughly evaluated the proposed approach on multiple publicly head pose databases.