Intermediate Views for Face Recognition

Intermediate views are widely used by arbitrary pose face recognition algorithms. Most approaches use real intermediate views or parametric interpolation of the face descriptors. Since the view based methods have been found to yield higher recognition rates, methods for generating synthetic intermediate views are necessary when a dense set of real views is not given. In many important applications only one or two real views are available. When two or more real views are available, it is geometrically possible to reconstruct a 3D model of the subject, which can be used to generate synthetic views at arbitrary poses. When only one real view is available, unique geometric reconstruction of the 3D model of the subject is not possible. This paper presents a technique for synthesizing intermediate views from only one real view by integrating shape from shading methods with a generic 3D model. The new approach is developed within the framework of the standard shape from shading methods. However, the constraints due to the generic model are integrated into the derivation of the new methods. The approach yields three methods for synthesizing intermediate views from one real view. All of the methods are tested on real and synthetic face images using the normalized correlation coefficient, and the reconstructed 3D models are empirically evaluated using the depth information from a laser range finder.