Faces with Exchanged Texture Average

Human faces diier in shape and texture. This paper describes a representation of grey-level images of human faces based on an automated separation of two-dimensional shape and texture. The separations were done using the point correspondence between the diierent images, which was established through algorithms known from optical ow computation. A linear description of the separated texture and shape spaces allows a smooth modeling of human faces. Images of faces along the principal axes of a small data set of 50 faces are shown. We also reconstruct images of faces using the 49 remaining faces in our data set. These reconstructions are the projections of an image into the space spanned by the textures and shapes of the other faces.