Evaluation of the Modelling of Local Areas and Errors of Localization in FRGC' 05

We present an evaluation of a probabilistic, part-based algorithm designed at The Ohio State University. Our algorithm is robust to errors of precision made by the (automatic) face and facial feature detector and to local image changes due to, for example, expression and illumination. Our contributions include the design of a novel face and facial feature detector and the justification of the use of the Mahalanobis cosine distance. We show results on experiments 1 and 4 in the FRGC (Version 2) test/database. Our algorithm includes a new face detector that is used to demonstrate the robustness of our algorithm to small errors of localization.