A vast amount of time was spent on model rigging. Allison and I both worked on it, and we found that creating a tolerably realistic looking face was extraordinarily difficult. We are so familiar with what a face is supposed to look like that any deviation is immediately obvious. The resulting rigged model sacrificed some flexibility for reliability. Many constraints were placed on the input in order to prevent unwanted noise and movement.
For future work a more general rig would be useful. We had about 20 markers, and each feature had a particular control setup. To get a wider level of expression many more markers are needed. For example, one of the commercial systems I found used 90 markers. With that much data it would be more useful to do some kind of free form deformation than to setup each feature individually.