Correspondence normal difference: An aligned representation of 3D faces to apply discriminant analysis methods

Since the geometry of the face changes drastically under expression variation, it is one of the greatest challenges in 3D face recognition to design systems that are robust to this variability. In this paper, we introduce a new representation of 3D faces in which all facial features are aligned over different faces. This representation contains highly discriminative features and is particularly useful for the employment of discriminant analysis methods for 3D face recognition. To the best of our knowledge, because of the lack of such alignment, so far, discriminant analysis methods have not been directly applied to 3D faces. Instead, the common approach is to register a probe face to each of the gallery faces, and then calculate the sum of the distances between their points for recognition. This registration also demands extensive computational effort. We demonstrate that the capability of a discriminant method, such as the LDA, to discriminate between the geometric variations resulted from expression changes and the geometric variations resulted from subject difference is beyond the capability of the state-of-art 3D face recognition methods. We achieved a verification rate of 99.5 percent at a false acceptance rate of 0.1 percent on the FRGC v2 database which is, to the best our knowledge, the best performance reported for this database in the literature.