Gravar-mail: Bayesian learning of visual chunks by human observers