A Comprehensive Understanding of Machine Learning and Deep Learning Methods for 3D Architectural Cultural Heritage Point Cloud Semantic Segmentation

As a result of the development of Artificial Intelligence (AI) techniques, in recent years, machine learning (ML) and deep learning (DL) approaches have been widely used to semantically enrich 3D architectural cultural heritage (ACH) point clouds. While existing approaches for analyzing and interpreting point clouds continue to improve, the generalizability of pre-trained ML and DL methods to various types of historic buildings remains uncertain. In this context, a comprehensive understanding of both methodologies can enable us to make more effective use of AI techniques in the ACH domain (e.g., data exploitation, model definition, analysis, and preservation). This work presents and compares two very different approaches for the 3D ACH semantic segmentation task. Specifically, we train and test a ML method based on the Random Forest (RF) classifier on the point cloud of three chapels part of the “Sacromonte Calvario di Domodossola” and on the two test scenes of the ArCH dataset. Then, we employ dynamic graph convolutional neural network (DGCNN) as our DL method, training on the ArCH dataset and testing on both the two unseen test scenes of the ArCH dataset and on the “Sacrimonti” chapel point clouds. We provide empirical experiments to illustrate the efficiency of applying ML and DL methodologies to ACH point clouds. Following that, the advantages and limitations of these two approaches are evaluated through a systematic study of the classification results. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.