Face Detection and 3D Face Reconstruction

A fully automated 3D face reconstruction system for faces in 2D images is presented. It is implemented incorporating a method for facial feature detection into the initialization of the 3D face reconstruction process provided by Blanz et al.[2][3]. The initialization of that method takes alignments of previously selected facial feature points with their according vertices in the 3D face model. Several methods for face detection are presented and discussed. In order to reliably detect facial features independently of rotation and size of a face, support vector classifiers are applied. Their data are represented by the gradient of the images used for training. The training images are generated by using the 3D Morphable Face Model[2]. They are selected in a database of face images with rotations around the vertical axis that was previously fitted to the Morphable Model. That enables to select facial features within these images without user interaction. In order to protect against detection outliers (which would disturb the reconstruction) the feature points are validated by each other on base of their detected geometric position. On base of the detection results a heuristic is provided filling up discarded feature points.