Rotation-Invariant 3D Reconstruction for Face Recognition.

Lighting variations and geometrical transformations an the image formation process can severely degrade the performance of many face recognition techniques. In previous work, Atick et al. proposed a KL expansion- based technique for 30 facial surface reconstruction. Since a facial surface is intrinsic to the face and independent to lighting conditions, this leads to face recognition algorithms insensitive to lighting variations. Atick et al.’s technique, however, is sensitive to 30 afine transformations. In this paper, we describe a novel technique that makes face surface reconstruction rotation-invariant. Specifically, rotation transformations are described by parameters that are estimated during the reconstruction process, along with the KL expansion coefjicients. Experimental results indicate this technique can significantly improve recognition performance. Finally, our approach extends naturally to all other aJgine transformations, including translation and scaling.