Bridging the Simulation-to-Real Gap of Depth Images for Deep Reinforcement Learning