Virtual Disassembling of Historical Edifices: Experiments and Assessments of an Automatic Approach for Classifying...

Area of Science:

• Digital Heritage
• Computer Vision
• Geometric Computing

Background:

• Reality-based 3D recording techniques like photogrammetry and laser scanning are increasingly used for heritage documentation.
• 3D point clouds generated by these methods often lack semantic information, hindering analysis and monitoring.
• Manual semantic annotation of large point clouds is time-consuming and resource-intensive.

Purpose of the Study:

• To propose an automated pipeline for semantic segmentation and classification of multi-scalar 3D point clouds of heritage objects.
• To enable multi-level segmentation from historical complexes down to architectural elements like pillars and beams.
• To reduce the manual effort and resource requirements for point cloud labeling in heritage documentation.

Main Methods:

• Development of an algorithmic toolbox for processing large 3D point clouds.
• Implementation of a workflow for semantic segmentation and labeling of heritage complexes into individual buildings.
• Application of the same toolbox for further segmentation of buildings into architectural elements.

Main Results:

• The proposed automated pipeline successfully segments and semantically labels multi-scalar point clouds of heritage sites.
• The toolbox demonstrated promising results in segmenting heritage complexes and individual architectural elements.
• The system facilitates the semantic annotation of 3D point clouds, addressing the needs of machine learning algorithms.

Conclusions:

• The automated pipeline offers an efficient solution for semantic annotation of 3D heritage data.
• This approach supports detailed analysis, monitoring, and predictive modeling of heritage objects.
• The developed toolbox can significantly assist in manual point cloud labeling, especially for large datasets.