Illuminating the Face

Abstract

Shape from shading for general objects is a difficult problem, because it is ill-conditioned. However, when the domain is restricted to a specific class of objects, prior knowledge of the class may be used to provide useful constraints, thereby making the problem tractable. This paper presents a novel method for solving the shape from shading problem in the restricted domain of human faces. We show how knowledge of face shapes in general may be incorporated as a statistical model and used to recover the surface normals from a single input image. The model also captures shadow and specularity information without explicitly recovering the 3D shape. New images of the face under novel illumination can then be realistically rendered.