The Surface-Approximation Polynomials (SAP) descriptor has been shown to be an appropriate global surface descriptor for object categorization tasks in robotic applications [1]. Nevertheless, in the original formulation the SAP descriptor is not invariant against rotations around the camera axis. This paper explains and evaluates two methods which pre-process the input data to yield repeatably well-aligned point clouds for the computation of the SAP descriptor. We show that the SAP descriptor can be rendered robust against rotations while retaining almost the full performance of the original approach which is superior to GFPFH, GRSD and VFH.