Segment Anything Model for Point Cloud Temporal Change Detection on Construction Sites

For the task of construction progress monitoring, precise and automated change detection for individual building elements is a key challenge. 3D laser scanning and reality capture technologies allow the physical geometry of a construction site to be captured efficiently; however, current methods for accurately extracting object boundaries for the purpose of change detection remain tedious and error prone. Existing segmentation and clustering algorithms tend to work well for simple objects with known geometry, but may fail for scenes with complex geometry, occlusion, and clutter. To address these issues, this paper proposes the use of Segment Anything Model (SAM), paired with 2D–3D projection and geometrical reasoning, to perform automated temporal change detection on construction sites from laser-scanned point clouds. The key advantage of SAM is the ability to segment arbitrary objects in a zero-shot manner, without additional training or fine-tuning. In this work, the input point cloud is first projected onto virtual panoramic images, then a pre-trained SAM is applied to predict object instance labels for each panorama. The object instance labels are reprojected back into the point cloud, which allows the 3D scene to be subdivided into individual object instances. This process is repeated for each stage of construction, and a geometry matching algorithm is used to identify changed and unchanged building elements across multiple stages. For experimental evaluation, this work uses the Nothing Stands Still (NSS) dataset, which consists of 3D point clouds collected from real-world jobsites across multiple construction stages, involving significant additions and removal of construction-related elements. Experimental results show that the SAM method outperforms PointNet and rule-based clustering for change detection based on the F1-score metric.