Feature Extraction of 3D Point Clouds of Indoor Space Using Autoencoder with PointNet and Evaluation through Missing Data Restoration

Abstract

In this study, we focused on the shape of the indoor space point cloud and extracted spatial features as the first step to automatically construct a virtual space where the surrounding environment can be confirmed and the degree of freedom of expression is high, using point cloud data and a generative AI. Specifically, we created an autoencoder using PointNet and evaluated it by restoring missing features. Feature extraction was performed using two types of data: the rectangular parallelepiped point cloud data created virtually and indoor space point cloud data in real space. Spatial feature extraction was evaluated by comparing the shape of the output point cloud with and without missing points and by the distance error between the input point cloud and the restored output point cloud. The results suggest that it may have been possible to extract the spatial features of the general shape of the room, but it was not possible to reproduce the detailed shape of the indoor of the room.