Svd-Based Point Cloud 3d Stone by Stone Segmentation for Cultural Heritage Structural Analysis – the Case of the Apollo Temple at Delphi

Unprecedented growth in 3D reality-based capturing devices - such as Terrestrial Laser Scanners (TLS) or Unnamed Aerial Vehicles (UAVs) equipped with high-resolution cameras, followed by photogrammetry 3D metrics reconstruction - has revolutionized the way Architectural, Engineering, and Construction (AEC) sector operates. Building Information Modelling (BIM) naturally emerged as the only viable alternative for managing this type of heterogeneous information. Unfortunately, despite the foreseen increase in BIM used in the heritage sector, known as Historic BIM (HBIM), high fidelity and increased resolution mesh output still need to be revised in BIM platforms. This inability of end users to handle meshed surfaces operates as a bottleneck to HBIM expansion and progression. Therefore, various research endeavours are exploring the latest state-of-the-art Machine Learning (ML) and Deep Learning (DL) algorithms for large-scale class recognition, leaving thus a gap in how to dismantle low-represented small objects. In this approach, this manuscript presents and validates a concrete methodology that utilizes Singular Value Decomposition (SVD) for boundary plane detection and mesh segmentation, which, as validated herein, reaches stone-level accuracy. The presented methodology leaves geometry intact, does not require mesh conversion, data transformations or additional channel input such as or colour or texture. Based on the findings of this research, state-of-the-art research efforts now consolidate towards high applicability and easy-to-generalize procedures that could reach stone-by-stone reverse engineering alternatives speeding up and broadening the otherwise time-and-labour plus resource-intensive Scan-to-BIM plans. The results disclose that, under a real case scenario of a heavily ruined monument where common architectural treaties are at stake, the developed algorithm can detect and detach elements robustly from the scene and then, as the next step, it efficiently applies a stone-by-stone segmentation for each columnresembling instance, with an overall accuracy of 90%. This easy-to-implement and simple-to-optimize approach is a viable auxiliary tool for multi-layer and multi-resolution automation frameworks, with or without machine learning deployment.(c) 2023 Consiglio Nazionale delle Ricerche (CNR). Published by Elsevier Masson SAS. All rights reserved.