Face alignment via joint-AAM

Abstract

In this paper, a joint active appearance model (joint-AAM) framework is proposed for face alignment. The object function consists of more than one active appearance model and some constraint items. It can be optimized through the efficient project-out inverse compositional (POIC) fitting algorithm. By transferring the low dimensional parameter space to the high one, the facial shape can be converged to the acceptable solution easier by joint-AAM comparing to single AAM, especially if the initial solutions locate on each side of the optimal solution. In multi-view case, different AAMs are jointed if the true view is far from the initial views. In single view case, different initial solutions of one AAM can be jointed to handle poor initialization or exaggerative expressions. Alternatively, 3D shape model is employed to impose stronger shape constraints on joint-AAM. A geometrical explanation is given to describe the reason of the robustness of the joint-AAM. The experiments demonstrate its accuracy, robustness and efficiency. The acronyms in this paper are listed in Tab. 1.