Image-to-Image Translation for Autonomous Driving from Coarsely-Aligned Image Pairs