Face Alignment by Supervised Descent Method with Head Pose Estimation
摘要：Face alignment, which aims at locating facial key points automatically, is an important topic in computer vision community. And many works have been done to solve this problem. The most well-known solution is Supervised Decent Method（SDM）. However, SDM has been designed to use mean shape as initial shape, which is vulnerable to large pose variation. In this paper, we present a novel approach for detection of the facial key points, getting initial shape from a special shape according to the head pose of the data. Experiments show that our approach achieves significant improvement. In both 21 points and 68 points detection cases, our method achieves nearly 50% improvement on challenging dataset IBUG, and about 1% improvement in HELEN and LFPW test set.
2019 4th International Conference on Communication, Image and Signal Processing （CCISP 2019）