Automated Extraction of Pipe Geometry Using SAM for Mixed Reality Inspection Tasks
Accurate detection and measurement of building elements are essential for efficient automated inspection and quality assessment in construction. This study evaluates the effectiveness of the Segment Anything Model (SAM) for pipe segmentation using a Mixed Reality-based dataset and introduces an auto...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Copernicus Publications
2025-07-01
|
| Series: | The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences |
| Online Access: | https://isprs-archives.copernicus.org/articles/XLVIII-G-2025/417/2025/isprs-archives-XLVIII-G-2025-417-2025.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Accurate detection and measurement of building elements are essential for efficient automated inspection and quality assessment in construction. This study evaluates the effectiveness of the Segment Anything Model (SAM) for pipe segmentation using a Mixed Reality-based dataset and introduces an automated method for pipe 3D centreline reconstruction and diameter estimation. The impact of the input point prompt distribution and number on segmentation accuracy is analyzed, identifying optimal configurations for improved performance. Using depth data and pose information from the MR device, the proposed approach reconstructs the 3D centreline and estimates pipe diameters with high reliability. The method is evaluated in a real experimental pipe network. The results indicate that the use of five-point prompts in a uniform distribution achieves approximately 90% precision and recall for pipe segmentation, with median position and diameter errors of 33 mm and 10 mm, respectively. The findings highlight the ability of the MR system to achieve accurate pipe positioning and diameter estimation, particularly in pipe networks with moderate complexity and fewer thin pipes, where segmentation and measurement challenges are minimized. |
|---|---|
| ISSN: | 1682-1750 2194-9034 |