Detection of Facial Landmarks in 3D Face Scans Using the Discriminative Generalized Hough Transform (DGHT)

Abstract

This paper presents the Discriminative Generalized Hough Transform (DGHT) as a technique to localize landmarks in 3D face scans. While the DGHT has been successfully used for the detection of landmarks in 2D and 3D images this work extends the framework to be used with triangle meshes for the first time. Instead of edge features and their respective gradient direction, the relative positions and orientations of the mesh faces are utilized to describe the geometric structures which are relevant for the detection of a specific landmark. Implementing a coarse-to-fine strategy at first a decimated version of the mesh is used to locate the global region of the point of interest, followed by more detailed localizations on higher resolution meshes. The utilized shape models are created in an automated, discriminative training process which assigns individual weights to the single model points, aiming at an increased localization rate. The technique has been applied to detect 38 anthropometric facial landmarks on 99 3D face scans. With an average error of 1.9mm, the most accurate detection was performed for the right alare, the average error when considering all landmarks amounts to 5.1 mm.