2D or Not 2D? That Is the Question: What Can We Learn From Computational Models Operating on Two-Dimensional Representations of Faces?
An additional difﬁculty comes from the fact that the perceptual appearance of a face changes dramatically with changes in expression, orientation, or lighting (see Fig. 11.1). Hence, a second constraint imposed on a face representation is that it should be ﬂexible enough to accommodate these transformations. Early models of face processing attempted to solve this problem by using a geometrical coding of faces. Key features were localized in the faces, and various measurements were taken between these key features. This type of coding has the advantages of (a) capturing information useful for discriminating among faces, and (b) being relatively insensitive to transformations. Its main drawback is that it discards texture information that might be useful for tasks such as sex or race categorization or identifying the expression or orientation of a face.