Novel methods for real-time 3D facial recognition

In this paper we discuss our approach to real-time 3D face recognition. We argue the need for real time operation in a realistic scenario and highlight the required pre- and post-processing operations for effective 3D facial recognition. We focus attention to some operations including face and eye detection, and fast post-processing operations such as hole filling, mesh smoothing and noise removal. We consider strategies for hole filling such as bilinear and polynomial interpolation and Laplace and conclude that bilinear interpolation is preferred. Gaussian and moving average smoothing strategies are compared and it is shown that moving average can have the edge over Gaussian smoothing. The regions around the eyes normally carry a considerable amount of noise and strategies for replacing the eyeball with a spherical surface and the use of an elliptical mask in conjunction with hole filling are compared. Results show that the elliptical mask with hole filling works well on face models and it is simpler to implement. Finally performance issues are considered and the system has demonstrated to be able to perform real-time 3D face recognition in just over 1s 200ms per face model for a small database.