When recognizing a fixed object from a fixed viewpoint, the dominant source of variation in image intensity is lighting changes. We propose a lowdimensional model for human faces that can both synthesize a face image when given lighting conditions and can estimate lighting conditions when given a face tmage. The model can handle non-Lambertian and selfshadowing surfaces such as faces because it does not m a k e any assumptions about either the surface's geometry or bidirectional reflecfance function. The model can be adapted to handle any arbitrary lighting condition,, and is easily extendable to any other viewpoint or to any other object.

No abstract