Neural MLS Surface Reconstruction of Point Clouds

The ability to accurately perceive and model natural structures is a central consideration of many application, including robotic. In this project we propose a new neural based approach to reconstruction: starting with the moving least squares reconstruction approach, we make use of a multi-layer perception to develop a set of learned basis functions for optimal reconstruction. To enable the development of the novel approach, we first construct a novel 2D point-set dataset we refer to as Shapeset. We then construct the PyPointset library, which implements three different, existing, surface reconstruction methods to act as benchmarks. Finally, we present our experimentation with the novel neural approach, highlighting strengths and current weaknesses, and provide an outlook to future work on this topic.