Semantic Pose using Deep Networks Trained on Synthetic RGB-D