This paper presents an advanced deep learning methodology to overcome critical limitations in machine learning and computer vision approaches for heritage building information modelling (HBIM) point cloud classification, including the inability to handle fixed-scale features, misaligned neighbourhoods and semantic overlap inherent in heritage structures.
The experimental setup involved training on a diverse, multi-site corpus of synthetic and real-world HBIM sites, including the Maritime Museum, Park Row and Brass Foundry. Input data, characterized by coordinates, normals and colour, were classified using a consolidated 13-class taxonomy. Performance was assessed using metrics including overall accuracy and mean intersection over union (mIoU).
Compared to limited training configurations, expanding the training domain from two to five sites significantly enhanced cross-site generalization on the entirely unseen Queen’s House test set. This comprehensive approach yielded a quantitative improvement in mIoU (excluding geometrically ambiguous flat classes) from 26.23% to 59.19%, validating the method’s substantial potential for establishing robust and scalable scan-to-HBIM workflows.
Our novel approach leverages the state-of-the-art Point Transformer v3 (PTv3) architecture, known for capturing multi-scale geometric details and broad contextual patterns, combined with low-rank adaptation and Point Prompt Training. This integration facilitates resource-efficient domain adaptation necessary for heterogeneous and label-scarce HBIM environments, addressing the synthetic-to-real data gap and class imbalance issues.
