Fast Simplex Optimization for Active Appearance Model

This paper presents a fast optimization method for active appearance model based on Nelder & Mead simplex in the case of mouth alignment under different expressions. This optimization defines a new constraint space. It uses a Gaussian mixture to initialize and constraint the search of an optimal solution. The Gaussian mixture is applied on the dominant; eigenvectors representing the reduced data given by Principal Component Analysis. The new algorithm constraints avoid calculating errors of solutions that don't represent researched forms and textures. The constraint operator added to simplex verifies in each iteration that the solution belongs to the space of research. The tests performed in the context of generalization (learning and testing datasets are different) on two datasets show that our method presents a better convergence rate and less computation time compared to the AAM classically optimized.