Noise Robust Surface Reconstruction by Combining PU and Graph-cut

Abstract

We present a novel method of reconstructing surfaces from 3D scattered points by combining Partition of Unity (PU) and a Graph-cut approach. PU is a local approximation technique, meaning that the surfaces obtained have high accuracy but are sensitive to noise. Graph-cut, on the other hand, is a global algorithm that is robust to noise but produces low-accuracy results because it is a discrete binary operation. Our algorithm combines these two methods to achieve robust, high accuracy surface reconstruction. First, a PU implicit function is constructed by covering a space containing a point cloud with spherical supports of linear polynomials. Graph-cut is then performed to separate the covered domain into inside and outside areas of the object to be reconstructed. Finally, we extract the zero-level of PU using the marching tetrahedra approach.