ArchShapesNet: a novel dataset for benchmarking architectural building information modeling element classification algorithms

Abstract

Abstract Recent studies in the domain of semantic enrichment have employed artificial intelligence (AI) approaches to distinguish and classify building information modeling (BIM) elements to check their conformance with open standard data formats. Training AI algorithms requires the development of well-balanced and robust datasets of BIM elements. However, collection is difficult as sources are limited to existing models and sample libraries. This study developed a parametric augmentation approach to create synthetic copies of BIM elements, and thus rapidly supplement manually collected samples. The approach was used to create ArchShapesNet, a dataset consisting of 11 common architectural elements with an equal size of 4,000 samples per class. Two multi-view convolutional neural networks (CNN), a geometric deep learning algorithm, were trained and tested separately on ArchShapesNet and an initial dataset with sample imbalances. Results showed significant improvement in the accuracy and F1 scores, providing evidence of the utility of ArchShapesNet. The size and scope of the dataset are considered to be the first of their kind and provide a benchmark for testing the semantic integrity of BIM models. The augmentation approach also provides a general framework to create custom datasets for different specialties in the Architectural Engineering and Construction industry.