Deep Learning based Geometric Information Analysis of Hole Detection Algorithm for Complex 3D Models

This paper presents the analysis results of deep learning network performances when various geometric information is added for 3D models or point clouds. When a 3D point set or point cloud is given as the input, various geometric information is first computed and used for train- ing deep learning networks to test effects for hole boundary detection. In this paper, two geo- metric information values are computed, which are the biggest angle gap and the distance between a query point and centroid of its neighborhood. This paper trains the deep neural net- works (PointNet and PointNet++) by adding geometric information to input points, and ana- lyzes whether geometric information is effective for point classification. The paper performs the experiments for Modelnet10 dataset, which contains various point clouds to test effects of the geometric information, and shows that both the geometric information helps the neural net- works to correctly detect hole boundary points of the input pointset, which improves the perfor- mance of point classification.