Abstract: In order to automatically find facial landmarks of a 3D facial shape, such as nose-tip and inner eye corners, this paper presents an effective approach to automatically detect landmarks of a 3D face in real time.A local surface descriptor based on partition and histogram, named sphere partition histogram descriptor, is designed to characterize local surface at a point on 3D face shape.Then we exploit the SPIDER descriptor to localize facial landmarks.The proposed algorithm is suitable for acceleration by GPU parallel computing.The experiments on FRGC v1.0 and BU-3DFE show that our algorithm outperforms other existing methods on 3D facial landmarks localization.