Classifying BIM (Building Information Model) Objects using Graph Neural Networks
Mujtaba, Choudhary Muhammad (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025052114100
https://urn.fi/URN:NBN:fi:amk-2025052114100
Tiivistelmä
The classification of Building Information Model (BIM) objects plays a crucial role in enhancing data management and automation in the construction industry. Manual classification remains a time-consuming task while rule-based classification is not very accurate. This task was assigned by a BIM software company Datacubist Oy with the goal of investigating the applicability of deep learning for classifying BIM model objects based on their geometric features and neighborhood information. The objective was to evaluate whether incorporating contextual information through Graph Neural Networks (GNN) could improve classification accuracy com-pared to traditional approaches.
The work was carried out using real-world BIM data from several building projects. The process involved ex-porting geometry data as Oriented Bounding Boxes, and the neighborhood information as edge index from the source software. A feature extractor based on convolutional layers was implemented, followed by three separate classification pipelines using Graph Convolutional Networks (GCN), Graph Isomorphism Networks (GIN) and Graph Attention Networks (GAT). Each model was trained on the same dataset and with the same training configuration. The models were compared based on accuracy scores and the results were further analyzed using F1-scores and confusion matrices.
The results showed that GNN-based methods performed better than a baseline CNN model, hence proving the potential of utilizing contextual information for the classification of BIM objects. Furthermore, among the tested graph architectures, the GAT model achieved the highest classification accuracy.
The results demonstrated the feasibility of using deep learning and GNNs for BIM object classification and indicated potential applications including automation of BIM workflows and model errors detection.
The work was carried out using real-world BIM data from several building projects. The process involved ex-porting geometry data as Oriented Bounding Boxes, and the neighborhood information as edge index from the source software. A feature extractor based on convolutional layers was implemented, followed by three separate classification pipelines using Graph Convolutional Networks (GCN), Graph Isomorphism Networks (GIN) and Graph Attention Networks (GAT). Each model was trained on the same dataset and with the same training configuration. The models were compared based on accuracy scores and the results were further analyzed using F1-scores and confusion matrices.
The results showed that GNN-based methods performed better than a baseline CNN model, hence proving the potential of utilizing contextual information for the classification of BIM objects. Furthermore, among the tested graph architectures, the GAT model achieved the highest classification accuracy.
The results demonstrated the feasibility of using deep learning and GNNs for BIM object classification and indicated potential applications including automation of BIM workflows and model errors detection.