So far we have been using range data obtained from cylindrical geometry CyberwareTM range finders to build person-dependent realistic face models. Such scanners produce a dense range image with itscorresponding cylindrical color texture. However, the dataset cannot be used directly because it is too dense (in average 1.4 million vertices) and sometimes includes some outliers.

To achieve both visual realism and real-time computation, we need a geometric model with a limited number of vertices but with enough details in order to distinguish facial features such as the lips or eyebrows. We have developed a reconstruction system based on deformable simplex meshes to build such models. The final result is a 3D wireframe,  refined specifically a some facial features to offer more precision, while keeping the number of vertices reasonably low. This wireframe is then mapped with a cylindrical texture image (coming from the CYBERWARE^TM scan), via a set of texture coordinates, and looks highly realistic under any point of view.

Currently, part of our research is focused on the development of a low cost 3D modeling method. Our main goal is to obtain realistic 3D head models using very basic equipment. Eventually we will test these models in our system to prove the model-independence of our approach.

 

We also implemented some animation techniques so that our face models are capable of facial expressions.