Discrimination of Plant Structures in 3D Point Cloud Through Back-Projection of Labels Derived from 2D Semantic Segmentation

抄録

In the decommissioning of the Fukushima Daiichi Nuclear Power Station, radiation dose calculations necessitate a 3D model of the workspace are performed to determine suitable measures for reducing exposure. However, the construction of a 3D model from a 3D point cloud is a costly endeavor. To separate the geometrical shape regions on 3D point cloud, we are developing a structure discrimination method using 3D and 2D deep learning to contribute to the advancement of 3D modeling automation technology. In this paper, we present a method for transferring and fusing labels to handle 2D prediction labels in 3D space. We propose an exhaustive label fusion method designed for plant facilities with intricate structures. Through evaluation on a mock-up plant dataset, we confirmed the method’s effective performance.