Facial pose from 3D data

The distribution of the apparent 3D shape of human faces across the view-sphere is complex, owing to factors such as variations in identity, facial expression, minor occlusions and noise. In this paper, we use the technique of support vector regression on wavelet sub-bands to learn a model relating facial shape (obtained from 3D scanners) to 3D pose in an identity-invariant manner. The proposed method yields an estimation accuracy of 97–99% within an error of +/− 9° on a large set of data obtained from two different sources. The method could be used for pose estimation in a view-invariant face recognition system.