ABSTRACT
As we remarked in the previous chapter, when recognizing a fixed object from a fixed viewpoint, the dominant source of variation in image intensity is lighting changes. In this chapter, we propose a low-dimensional model for human faces that can both synthesize a realistic face image when given lighting conditions and estimate lighting conditions when given a face image. The methods and results given here were sketched in Hallinan (94). The most important feature of this model is that it does not make any assumptions about either the surface geometry or the bidirectional reflectance function of the object being modeled. Thus nonLambertian, specular and self-shadowing non-convex surfaces such as faces can be modeled. Other characteristics of the model are that it can be updated to handle any arbitrary lighting condition, it is easily extended to any other viewpoint or to any other object and it is designed for use by recognition or scene analysis algorithms that employ sets of two dimensional viewpoint specific models instead of three dimensional models.
