Towards Personalization of 3 D Virtual Humans - A Head Start

This thesis gives an overview of the research of a system that generates a 3D model of a human head from 2D images. The location and size of the eyes, eyebrows, nose and mouth in 2D orthogonal images are detected using skin color detection, template matching, value detection and a anthropometric model. The results of these detections are combined for more robust results. Manual initialization is needed to correct differences in lighting. The use of the anthropometric model yields and average reduction of search area of 91%-97% depending on the feature resulting in a speed up of 11-33. In 90% of the test cases the features are found correctly. The location information of the facial features are used to deform a 3D generic head. Current deformation parameters are: head width, eye height, eye width, eyebrow height, mouth width, mouth height, nose width, nose height and nose length. The application is generic enough to allow for any number of deformations. A generic texture is applied to the deformed 3D head. A user experiment was conducted, where it was assessed whether humans could identify the generated model out of 4 different 3D models. This was true in 40% of the cases.