Recently, 3D face reconstruction and face alignment tasks are graduallycombined into one task: 3D dense face alignment. Its goal is to reconstruct the3D geometric structure of face with pose information. In this paper, we proposea graph convolution network to regress 3D face coordinates. Our method directlyperforms feature learning on the 3D face mesh, where the geometric structureand details are well preserved. Extensive experiments show that our approachgains superior performance over state-of-the-art methods on several challengingdatasets.