Disney Research shows a new robust facial capture system
The process takes into account a local deformation subspace, as well as an anatomical bone structure. Through a small set of training shapes it can be constrained to 2D motion data. This method allows for many applications, including high quality motion capture from sparse facial markers or sketch based manipulation from artist input. It can even reconstruct high quality motion capture based on single camera input such as a mobile phone.
The method itself seems to outperform the traditional global blend shape method and is better at tracking the face position correctly.
The video that explains the process can be seen here: