A System for Advanced Facial Animation

This thesis describes a system for extracting three-dimensional models from two-dimensional images over the restricted domain of frontal images of human faces. A three-dimensional face model template is fit to a three-dimensional digitizations of a set of training subjects. The standard-form models allow different subjects to be treated in a uniform manner. The particular template model chosen is almost identical to one used for head tracking and expression extraction from video, allowing more accurate modelling of the subject. The models are used to synthesize small face images with a series of discrete orientations, and then subjected to Karhunen-Lobve expansion (principal component analysis), yielding one low-dimensional of coupled eigenimages and eigenmodes per orientation. Subsequently, face images within the range of rotations used to generate the eigenspaces can be projected onto the eigenspaces, yielding a coarse estimate of facial orientation, recovering missing data, and producing an approximate facial model. Thesis Supervisor: Alex (Sandy) Pentland Title: Head, Perceptual Computing Section, The Media Laboratory Thesis Supervisor: Irfan A. Essa Title: Research Scientist, Perceptual Computing Section, The Media Laboratory