Face pose estimation with ensemble multi-scale representations

Abstract

Face pose estimation plays important roles in broad applications such as visual based surveillance, face authentication, human-computer intelligent interactions, etc. However, face pose estimation is also a challenge issue, especially under complicated real application environments. In this paper, we proposed a novel face pose estimation approach with integrating two multi-scale representations. The first one is multi-scale VGG-Face representations, which using VGG-Face CNN as backbone three middle scale layer outputs are extracted and go through additional transfer learning. The second one is multi-scale Curvelet representations. These two sub multi-scale representations are integrated and then several dense layers processing are added to form the entire ensemble system which is used for the prediction of face pose. The experiment results show that the proposed approach achieved mean absolute errors (MAE) of 0.33° and 0.23° for yaw and pitch angle on CAS-PEAL pose database, and achieved mean absolute errors of 3.88° and 1.98° for yaw and pitch angle on Pointing'04 database.