Combining Generative and Discriminative Models in a Framework for Articulated Pose Estimation