Comparing normal estimation methods for the rendering of unorganized point clouds

Abstract: Surface normals are fundamental in computer graphics applications such as computer vision, object recognition, and lighting calculations. When working with unorganized point clouds of surfaces, there exists a need for fast and accurate normal estimation methods. This thesis presents the investigation and implementation of two different methods of normal estimation on fixed-size local neighborhoods in unorganized pointclouds. Two main categories of tests were conducted. The first type was visual inspection and the second consisted of numeric analysis of the normal estimation process and results. Point cloud data used in the study included numerically exact representations of spheres, cubes, cones, as well as both uniformly sampled or laser-scanned real-world point clouds with millions of points. Complete triangle averaging was found to be the method of choice on small neighborhoods, justified by faster running-time while still estimating high-quality normals. When larger neighborhood sizes were needed, a size breakpoint was found above which principal component analysis should be used instead, which estimates normals of similar quality as the complete triangle averaging but with the added benefit of near-constant running-time independent of neighborhood size.