Conventional engineering parts usually composed by algebraic surfaces are important investigation objects of reverse engineering. In this paper a step segmentation method for algebraic models is proposed. The mesh normals and curvedness of every vertex are estimated as a shape descriptor. In the first segmentation fourteen directions are chosen initially, and a k-means algorithm according to the normal vectors is used, then the surface is divided to form patches by a region-growing scheme, as well as some sharp edges or flat areas are detected. In order to identify algebraic surface, curvedness of a patch is set as the criterion by which the surface merged into near constant curvedness region. Especially a novel mean shift algorithm is adopted in this method, that a powerful technique for clustering in image process, and is extended to normal filtering while preserving the features to increase robustness of the method. Experimental evaluations using scan data or noise data demonstrate the efficiency of the proposed method.