Inferring the Existence of Geometric Primitives to Represent Non-Discriminable Data

Abstract

In this thesis, we set out to find an algorithm that uses only geometric primitives to represent an input pointcloud. In addition to the problems faced in general primitive fitting, non-discriminable data presents additional data association challenges. We propose to address these challenges by estimating the existence rather than parameters of geometric primitives, and explore various options to do so. We first explore a sampling-based Markov-Chain Monte-Carlo approach together with a ray likelihood model. We then explore a neural network approach and finish by presenting a method to make the Chamfer distance differentiable with respect to primitive existence.