Deep-learning guided structural object detection in large-scale, occluded indoor point cloud datasets

Abstract: Automatic geometry digitisation of existing buildings remains challenging due to the large scale and heavy clutter of input point clouds. This paper presents a two-stage hybrid method to detect structural objects. The first stage detects areas of interest that are likely to contain an object, while the second stage finds precise objects. The method benefits from data-driven and model-driven approaches to achieve high accuracy for large-scale, highly cluttered and occluded real-world environments. We evaluate our method on the Stanford3D S3DIS dataset to show that the method detects from 83% to 98% of structural objects, such as columns, doors and windows.