Design of 3D point cloud dataset of indoor spaces for feature extraction using Autoencoder with PointNet

Abstract

In this study, a novel method was developed to automatically construct a virtual space with a high degree of freedom of expression. The constructed virtual space was designed to reflect the spatial shape of the real space and the arrangement of objects. First, the global shape of the interior space was used to design a dataset for extracting the spatial features of the real space by three-dimensional (3D) scanning of the real space and using a PointNet-based autoencoder. The dataset consisted of the point cloud data of a rectangular 3D object that was a simple imitation of a room in real space and focused on two items, namely the number of input points and the number of data points. The results of the autoencoder restoration indicate that spatial feature extraction can be performed when the number of data is 5000 or more.

Keywords: Three-dimensional (3D) point cloud, Feature extraction, Dataset, PointNet, Virtual space

DOI: 10.54941/ahfe1004672