Today, there are two major paradigms for vision-based autonomous driving
systems: mediated perception approaches that parse an entire scene to make a
driving decision, and behavior reflex approaches that directly map an input
image to a driving action by a regressor. In this paper, we