Registration of three-dimensional face scans with average face models

bstract. The accuracy of a three-dimensional (3-D) face recogniion system depends on a correct registration that aligns the facial urfaces and makes a comparison possible. The best results obained so far use a costly one-to-all registration approach, which equires the registration of each facial surface to all faces in the allery. We explore the approach of registering the new facial surace to an average face model (AFM), which automatically estabishes correspondence to the preregistered gallery faces. We proose a new algorithm for constructing an AFM and show that it orks better than a recent approach. We inspect thin-plate spline nd iterative closest-point-based registration schemes under anual or automatic landmark detection prior to registration. Exending the single-AFM approach, we consider employing categorypecific alternative AFMs for registration and evaluate the effect on ubsequent classification. We perform simulations with multiple FMs that correspond to different clusters in the face shape space nd compare these with genderand morphology-based groupings. e show that the automatic clustering approach separates the aces into gender and morphology groups, consistent with the other ace effect reported in the psychology literature. Last, we describe nd analyze a regular resampling method, that significantly inreases the accuracy of registration. © 2008 SPIE and S&T. DOI: 10.1117/1.2896291