Discriminative figure-centric models for joint action localization and recognition