This paper presents an improved method to construct dense correspondences for 3D facial analysis, which are capable of providing a full 3D description of a surface and extending the conventional landmark-based approaches. Based on the technique of elastic deformation, the dense correspondences are established by mapping a generic model onto the 3D surface of an individual. The method used here is accurate in driving the conformation of the generic model and efficient in dealing with outliers, which can appear during local deformation. The experiments indicate for open surfaces as face with this method, more than 95 percent of triangles on the deformed generic mesh are within the range of 1 mm to the original model and the average landmark errors are only 2 mm.