Learning Road Scene-level Representations via Semantic Region Prediction