FeatureNet: Upsampling of Point Cloud and it’s Associated Features

Abstract

In this paper, we address the problem of 3D Point Cloud Upsampling, that is, given a set of points, the objective is to obtain denser point cloud representation. We achieve this by proposing a deep learning architecture that along with consuming point clouds directly, also accepts associated auxiliary information such as Normals and Colors and consequently upsamples them. We design a novel feature loss function to train this model. We demonstrate our work on ModelNet dataset and show consistent improvements over existing methods.