Experimental investigation of multiclass 3D point cloud completion

Abstract

Point clouds captured in real world environments face challenges which cause them to be incomplete or ambiguous to varying degrees. While numerous solutions have been proposed to tackle these challenges, there have been fewer attempts at taking a multiclass approach for completing point clouds. A multiclass approach may generalize better to ambiguity of the partial point clouds by completing them according to a range of specified or plausible class outcomes. In this project, experiments were conducted using various methodologies for integrating multiclass shape completion capabilities into an existing shape completion neural network with and without the use of a conditional Generative Adversarial Network (GAN). The resulting multiclass pretrained models and their completed point clouds are also evaluated, with the feasibility of their implementation and performance of their results being discussed.