3D head pose computation from 2D images: templates versus features

In this paper we examine the problem of determining 3-D human head pose from a sequence of 2-D gray scale images. Computation of head pose is vital in many areas of human head modeling and applications include model-based video compression, face recognition and 3D head tracking. A significant amount of research has been done in the general area of 3D object alignment for recognition and modeling purposes, with much of the work concentrating on applying techniques to simple polyhedral objects such as cubes or polyhedrons. We take these techniques a step further and apply them to grey scale images of human heads using a set of distinguishing facial features as well as 2D template regions to compute the pose relative to a given 3D model. In the feature point approach, we use a small subset of features and optimize their selection by minimizing the error in mapping the remaining points with the computed pose. In the template method we minimize the difference in energy between the input facial region and the template database. We obtain results for a wide range of head orientations and scales and the methods proved to be quite accurate.