Employing region ensembles in a statistical learning framework for robust 3D facial recognition

Recently, region approaches have been applied to 3D face recognition to mitigate performance degradation caused by variations in face expressions. In this work, we investigate the benefits of region ensembles for statistical learning based 3D face recognition. This investigation is conducted using 3D Fisherfaces and demonstrates that region ensembles improve the ability of the Fisherface approach to create discriminating features, even for untrained subject samples. In fact, comparable performance can be achieved using significantly fewer training subjects. In addition, gallery score normalization is integrated into the region ensemble framework and is shown to improve performance over a single normalization of the ensemble match score. Performance improvement is particularly evident at low false alarm rates. An All versus All comparison of the FRGC 2.0 database obtains a 97.1% True Accept Rate (TAR) at 0.1% False Accept Rate (FAR) when the approach is trained with 100 subjects and improves to 98.1% with integrated gallery score normalization.