Semantic Segmentation of building point clouds based on Point Transformer and IFC

Abstract

In the process of building construction, the semantic understanding of building point clouds provides potential solutions for efficient building quality supervision, progress monitoring, and sub-system deviation analysis. However, the lack of suitable public labeled datasets, the low degree of color discrimination and the particularity of multi-system coexistence in construction scenes have brought certain challenges to semantic segmentation of point clouds. But rich semantic information in related Industry Foundation Classes(IFC) can be used to synthesize effective labeled data. Our main contributions are as follows:1) We propose a synthetic conversion method from BIM model to point cloud, and construct a synthetic dataset for construction scenes based on IFC. 2) We segment the colorless point cloud of the construction scene into five types (IfcSlab, IfcBeam, IfcWall, IfcColumn, IfcDistributionFlowElement) with Point Transformer and use focal loss to improve the segmentation accuracy of small-area components with synthetic data for data enhancement.