Stav dette: Effective representations for road scene understanding with scalable learning