Assessing IFC classes with means of geometric deep learning on different graph encodings

Fiona C. Collins1, Alexander Braun1, Martin Ringsquandl2, Daniel M. Hall3, André Borrmann1
1 Technical University of Munich, Germany
2 Siemens AG, Corporate Technology, Munich, Germany
3 ETH Zürich, Switzerland


DOI: 10.35490/EC3.2021.168
Abstract: Machine-readable Building Information Models (BIM) are of great benefit for the building operation phase. Losses through data exchange or issues in software interoperability can significantly impede their availability. Incorrect and imprecise semantics in the exchange format IFC are frequent and complicate knowledge extraction. To support an automated IFC object correction, we use a Geometric Deep Learning (GDL) approach to perform classification based solely on the 3D shape. A Graph Convolutional Network (GCN) uses the native triangle-mesh and automatically creates meaningful local features for subsequent classification. The method reaches an accuracy of up to 85\% on our self-assembled, partially industry dataset.
Keywords: geometric deep learning, classification, Industry Foundation Classes
Pages: 332 - 341
Paper:
Contribution_168_final

Presentation Video: