FACE DETECTION AND TRACKING WITH 3D PGA CLM

In this paper we describe a system for facial feature detection and tracking using a 3D extension of the Constrained Local Model (CLM) (Cristinacce and Cootes, 2006) algorithm. The use of a 3D shape model allows improved tracking through large head rotations. CLM uses a shape and texture appearance model to generate a set of region template detectors. A search is then performed in the global pose / shape space using these detectors. The proposed extension uses multiple appearance models from different viewpoints and a single 3D shape model built using Principal Geodesic Analysis (PGA) (Fletcher et al., 2004) instead of direct Principal Components Analysis (PCA). During fitting or tracking the current estimate of pose is used to select the appropriate appearance model. We demonstrate our results by fitting the model to image sequences with large head rotations. The results show that the proposed multi-view 3D CLM algorithm using PGA improves the performance of the algorithm using PCA for tracking faces in videos with large out-of-plane head rotations.