Learning a Directional Soft Lane Affordance Model for Road Scenes Using Self-Supervision.