Item

Abstract

When only a single image of a face is available, can we generate new images of the face across changes in viewpoint or illumination? The approach presented in this paper acquires its knowledge about possible image changes from other faces and transfers this prior knowledge to a novel face image. In previous work we introduced the concept of linear object classes (Vetter and Poggio, 1997; Vetter, 1997): In an image based approach, a flexible image model of faces was used to synthesize new images of a face when only a single 2D image of that face is available.

In this paper we describe a new general flexible face model which is now “learned” from examples of individual 3D-face data (Cyberware-scans). In an analysis-by-synthesis loop the flexible 3D model is matched to the novel face image. Variation of the model parameters, similar to multidimensional morphing, allows for generating new images of the face where viewpoint, illumination or even the expression is changed.

The key problem for generating a flexible face model is the computation of dense correspondence between all given example faces. A new correspondence algorithm is described, which is a generalization of existing algorithms for optic flow computation to 3D-face data.