Expression, pose and occlusion resistant 3D facial landmarking

Abstract

This paper contrasts two approaches to facial landmarking in 3D. The first approach is statistical in nature, and is based on modeling the shape of each feature with Gaussian mixtures. The advantage of this approach is the uniform treatment of landmarks. The second approach is a hybrid method to find the nose tip, which does not require learning, and is robust under adverse conditions. We demonstrate the accuracy and cross-database performance of these methods on FRGC and Bosphorus databases.