Training And Evaluating Visual Recognition Systems With Limited Annotations