Abstract: We introduce an approach for learning human actions as interactions between persons and objects in realistic videos. Previous work typically represents actions with low-level features such ...