Active Appearance Models (AAMs) have been used with great success for subject specific non-rigid face alignment/tracking. Unfortunately, AAMs suffer inherent drawbacks when being applied to faces that have not been seen previously. We refer to this as the "Generic AAM" problem.
At present we are pursuing techniques for non-rigid face alignment based on Constrained Local Models (CLMs) that exhibit superior generic performance over conventional AAMs. Specifically, we have been investigating two flavors of CLMs we refer to as Exhaustive Local Search (ELS) and Robust Convex Quadratic Fitting (RCQF).