Creating Face Models from Vague Mental Images

We present a novel approach to create plausible 3D face models from vague recollections or incomplete descriptions, such as those given by eyewitnesses in police investigations. Our algorithm for navigating face space is based on a 3D morphable model. It exploits correlation between different facial features learned from a database, and uses a set of attribute constraints that restrict the face to a residual subspace. Faces are manipulated by in tuitive parameters or by importing facial elements from a database. To avoid exposure to confusingly different faces, each face in the database is mapped to the residual subspace defined by the constraints.