Sheela Raju Kurupathi¹, Park Dongryul¹, Sebastian Bosse¹ and Peter Eisert^1,2, ¹Fraunhofer Heinrich Hertz Institute (HHI), Germany, ²Humboldt University of Berlin, Germany

The construction industry faces challenges in extracting and interpreting semantic information from CAD floor plans and related data. Graph Neural Networks (GNNs) have emerged as a potential solution, preserving the structural integrity of CAD drawings without rasterization. Accurate identification of structural symbols, such as walls, doors, and windows, is vital for generalizing floor plans. This paper investigates GNN methods to enhance the classification of these symbols in CAD floor plans, proposing an entity-as-node graph representation. We evaluate various preprocessing strategies and GNN architectures, including Graph Attention Networks (GAT), GATv2, Generalized Aggregation Networks (GANet), Principal Neighborhood Aggregation (PNA), and Unified Message Passing (UniMP) on the CubiCasa5K dataset. Our results show that these methods significantly outperform current state-of-the-art approaches, demonstrating their effectiveness in CAD floor plan entity classification.

BIM, CAD, Floor Plans, GNN, Entity-as-Node, Multi-Classification.