3D Facial Landmarking under Expression, Pose, and Occlusion Variations

Automatic localization of 3D facial features is important for face recognition, tracking, modeling and expression analysis. Methods developed for 2D images were shown to have problems working across databases acquired with different illumination conditions. Expression variations, pose variations and occlusions also hamper accurate detection of landmarks. In this paper we assess a fully automatic 3D facial landmarking algorithm that relies on accurate statistical modeling of facial features. This algorithm can be employed to model any facial landmark, provided that the facial poses present in the training and test conditions are similar. We test this algorithm on the recently acquired Bosphorus 3D face database, and also inspect cross-database performance by using the FRGC database. Then, a curvature-based method for localizing the nose tip is introduced and shown to perform well under severe conditions.