Deep learning framework for point cloud instance segmentation in construction sites

doi:10.1016/j.autcon.2026.107159

Deep learning framework for point cloud instance segmentation in construction sites

2026-08-04

0

PRE

AI

DOI：10.1016/j.autcon.2026.107159

Original

Original request for help

Save

Abstract

En 中文

• Introduced Site-Inst dataset with 13 construction categories. • Proposed SiteInstance framework tailored for construction site point cloud instance segmentation. • Integrated transfer learning, geometric–semantic–direction fusion, and construction-aware component augmentation. • Achieved mAP50 of 84.0% and mAP25 of 88.1%, surpassing state-of-the-art baselines. • Ablation studies further verified the effectiveness and efficiency of the approach.

Keywords:

Construction site

DL

Point cloud

Instance segmentation

Deep learning framework for point cloud instance segmentation in construction sites

Abstract

Journal

Organization