Despite great advances witnessed on facial image alignment in recent years, high accuracy high speed face alignment algorithms still have rooms to improve especially for applications where computation resources are limited. Addressing this issue, we propose a new face landmark localization algorithm by combining global regression and local refinement. In particular, for a given image, our algorithm first estimates its global facial shape through a global regression network (GRegNet) and then using cascaded local refinement networks (LRefNet) to sequentially improve the alignment result. Compared with previous face alignment algorithms, our key innovation is the sharing of low level features in GRegNet with LRefNet. Such feature sharing not only significantly improves the algorithm efficiency, but also allows full exploration of rich locality-sensitive details carried with shallow network layers and consequently boosts the localization accuracy. The advantages of our algorithm is clearly validated in our thorough experiments on four popular face alignment benchmarks, 300-W, AFLW, COFW and WFLW. On all datasets, our algorithm produces state-of-the-art alignment accuracy, while enjoys the smallest computational complexity.