Physics based modeling of the human face
Ronald P. Fedkiw, Igor Neverov · 2005
Modeling of the human face plays an important role in such applications as computer games, medicine, and special effects for movies. The demand for this area is driven by the ability of the human face to convey emotion and information and by the needs of simulation in the context of facial surgery. Face modeling can be divided into defining the geometrical representation, animating the model, and rendering. It has recently enjoyed considerable progress in the movie industry; in particular the rendering has achieved such level that the face in still images has been made to look indistinguishable from real. However, animation still remains a problem, mainly because of the lack of sound algorithmic theory describing facial motion. Our work addresses this problem, proposing a model and tools for its animation, which both reflect the anatomical control structure and are efficient and robust. We built a highly detailed anatomically accurate model of facial passive tissue, embedded musculature and underlying skeletal structure using volumetric data acquired from a living male subject. The tissues are equipped with a highly nonlinear constitutive model including controllable anisotropic muscle activations based on fiber directions. Building this model required the development of an extensive set of tools that was used to process the Visible Human dataset as well as the MRI and laser scans of a living subject. The model is capable of a fast and robust animation, driven by muscle activations and the kinematic parameters defining the placement of cranium and jaw. To achieve versatile realistic animation requires complex coordinated stimulation of the muscles and control of the kinematic parameters. We propose a solution to this problem by automatically estimating the control parameters from face motion capture marker data. This not only offers an efficient way to drive the model, but also offers a framework for definitive description of facial motion, since our control parameters are anatomically derived, as opposed to phenomenologically learned.