The static and dynamic features of a face model are either created algorithmically, or are derived from human facial data. This section describes the current facial data acquisition techniques and processing algorithms used to derive face model parameters from the human facial data. We avoid the temptation to describe higher level recognition algorithms such as facial expression understanding, visual speech recognition (lipreading), or face recognition. These topics are comprehensively reviewed in the report from the 1992 NSF Workshop on Facial Expression Understanding.
Face model parameters include head position and orientation, eye orientations, eyelid closure, jaw position, facial surface shape and texture (hair, ears, wrinkles), and a variety of oral cavity parameters (lips, teeth, and tongue position). Virtually every available sensing modality has been used to acquire facial data including optical, acoustical, mechanical, electromylogram (EMG), tomography, x-ray and dissection. The sensing techniques cover a complete range of spatial and temporal resolution, and invasiveness. No method for facial data acquisition exists which does not compromise in one or more important dimensions. For example, laser range finders provide good 3 dimensional surface maps but only of static objects.
The following description of data acquisition techniques is followed by an overview of processing techniques for general face parameters. Finally, methods for the derivation of visual speech parameters are described in a separate section due to the use of special techniques in a variety of other contexts.