Exploiting semantic embedding and visual feature for facial action unit detection (CVPR 2021)