Korean Journal of Computational Design and Engineering

This paper performs experiments to analyze effects of geometric information based feature for deep learning networks to detect hole boundary points. When point cloud is acquired by conventional camera, it may include holes due to occlusion, blur or hardware spec. Therefore, deep learning based hole detection method is proposed using geometric features and deep learning networks. First, two geometric feature values are computed, which are the biggest angle gap and the distance between a query point and centroid of its neighborhood. Then, geometric feature is added to input points and trained. In experiments, geometric feature is analyzed whether these are still effective for Modelnet40 dataset, which contains 40 categories of objects. As a result, performance results show that geometric feature indeed helps the neural networks to correctly detect hole boundary points of the input pointset, which improves the performance of point classification.

키워드 Hole detection, PointNet, Point cloud, Point classification