Gabor Factor Analysis for 2D+3D Facial Landmark Localization

We propose a coarse-to-fme method for facial landmark localization that relies on unsupervised modeling of landmark features obtained through different Gabor filter channels. The input to the system is a registered near-frontal 2D and 3D face image pair, with background clutter. The system aims at complete automatic detection of seven facial landmarks; the nose tip, eye and mouth corners, respectively. A structural analysis subsystem is employed to detect incorrect landmarks and to correct them. We compare our local features with two widely used Gabor jet based methods, and illustrate their superior performance