Towards Efficient Deep Learning for Human-Centric Visual Understanding and Generation