Exploiting semantic embedding and visual feature for facial action unit detection (Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR 2021)