Boosted dynamic Active Shape Model

We present a novel optimization scheme in the popular active shape model(ASM) framework, which increases the accuracy and robustness of searching for a hypothesis shape. The determininistic fitting scheme in traditional ASM is substituted by a probabilistic estimation approach in our work. A set of weighted particles is used to represent each salient feature point to form a shape density, the particle densities are evolved by the observation model and used to update the shape parameter. The displacement between iterations provides the particles with dynamic information to locate a new starting shape density for the next iteration. Furthermore, the likelihood function of the observation model is trained with a Gentleboost regression algorithm and results in the proposal distribution. The proposed algorithm is much more robust in nonlinear and noisy environments and not affected by initialisation conditions, as are current ASM optimization algorithms. Finally, the developed approach provides higher accuracy and increases the convergence range in face segmentation on the BioID public test data set.