Compact 3D Model Generation based on 2D Views of Human Faces: Application to Face Recognition

Abstract

A face recognition and synthetic­natural hybrid coding tool involving frontal and profile views is presented in this paper. The tool utilizes the above mentioned 2D images ­upon which certain protuberant points are automatically detected­ and adapts a generic 3D head model (polygon mesh) according to the predetermined information gained by the available views. This mesh provides shape information which ­combined with texture information­ is important for the robustness of a recognition system. The main advantage of the proposed approach is its ability to overcome constraints raised from arbitrary variations in scale, rotation and orientation, which are dominant deterioration factors in the identification task. Besides the face recognition aspect, applications which involve geometry transmis­ sion, such as teleconferencing, can take advantage of the proposed algorithm's ability to parametrically describe a complex organic model, such as a human head. During the texture map creation process, issues related to the luminance differences and rotation vari­ ance between the available views, are successfully dealt with. Regarding the adaptation of the polygon topology of the 3D model, we apply a set of localized transformations, in order to preserve the continuity of the human head surface. Besides this, the problem of the minimum organic model representation required is addressed. The aspect under which this issue is verged upon, is that of the solution of a trade­off problem between low computational complexity and high approximation quality.