The RNN encoder decoder model for object appearance, motion modeling as well as ego-motion modeling. The decoder propagate over time to predict future locations Future object localization and interaction modeling