Fast low-memory streaming MLS reconstruction of point-sampled surfaces

Abstract

We present a simple and efficient method for reconstructing triangulated surfaces from massive oriented point sample datasets. The method combines streaming and parallelization, moving least-squares (MLS) projection, adaptive space subdivision, and regularized isosurface extraction. Besides presenting the overall design and evaluation of the system, our contributions include methods for keeping in-core data structures complexity purely locally output-sensitive and for exploiting both the explicit and implicit data produced by a MLS projector to produce tightly fitting regularized triangulations using a primal isosurface extractor. Our results show that the system is fast, scalable, and accurate. We are able to process models with several hundred million points in about an hour and outperform current fast streaming reconstructors in terms of geometric accuracy.