Towards Efficient Deep Learning For Human-Centric Visual Understanding And Generation