Towards Optimal Strategies for Training Self-Driving Perception Models in Simulation