The Treachery of Images: Bayesian Scene Keypoints for Deep Policy Learning in Robotic Manipulation