SIMPLE: SIngle-network with Mimicking and Point Learning for Bottom-up Human Pose Estimation.