Facial Shape Recovery by Feature Driven Stereo Analysis

This paper describes a hybrid least squares stereo matching and feature detection approach for the accurate metric recovery of 3-D facial shape and shape change. We have incorporated an active shape model to enable a priori image segmentation, facilitate adaptive window stereo matching and to locate seed point correspondences on which full field disparity estimates can be reliably initialised. Results are presented which demonstrate that 3D facial reconstruction detail can be enhanced through thin plate spline disparity map interpolation. The proposed method is sufficiently sensitive to detect localised shape changes induced by simulated surgical correction of abnormal jaw relationships.