Mise à profit des blendshapes pour la simulation physique temps-réel d'animation faciale
Vincent Barrielle · theses.fr (ABES) · 2017
Generating synthetic facial animation is a crucial step in the creation of content for a wide variety of digital media such as movies and video games. However, producing convincing results is challenging, since humans are experts in analyzing facial expressions and will hence detect any artifact. The dominant paradigm for the production of high-quality facial animation is the blendshapes paradigm, where facial expressions are decomposed as a linear combination of more basic expressions. However, this technique requires large amounts of work to reach the desired quality, which reserves high-quality animation to large budget movies. Producing high-quality facial animation is possible using physical simulation, but this requires the costly acquisition of medical imaging data.We propose to merge the blendshapes and physical simulation paradigms, to build upon the ubiquity of blendshapes while benefiting from physical simulation for complex effects. We therefore introduce blendforces, a paradigm where blendshapes are interpreted as a basis for approximating the forces emanating from the facial muscles. We show that, combined with an appropriate face physical system, these blendforces can be used to produce convincing facial animation, with natural skin dynamics, handling of lips contacts, sticky lips, inertial effects and handling of gravity. We encompass this framework within a practical realtime performance capture setup, where we produce realtime facial animation with physical effects from a simple RGB camera feed. To the best of our knowledge, this constitutes the first instance of realtime physical simulation applied to the challenging task of facial animation.