Towards Robust 3D Segmentation in Construction: Tackling Raw and Incomplete Scans

The demand for precise 3D data capture in the architecture, engineering, and construction (AEC) industry continues to grow, driven by the need for accurate progress monitoring and change detection. However, applying 3D segmentation models to raw individual scans remains challenging due to non-uniform point densities in dynamic construction environments. This study investigates point density normalization strategies and benchmarks two backbone networks, point transformer and sparse U-Net. Three downsampling strategies—baseline, sparse-aware, and dense-aware—were compared, with dense-aware consistently delivering the best performance. Notably, the dense-aware approach excels in recognizing temporary, non-structural objects like ladders, even in sparse point regions. The results confirm that allowing models to focus on dense areas where objects are complete significantly enhances segmentation accuracy. Future work will focus on scaling the study to outdoor environments and expanding dataset volumes to improve the generalizability of 3D segmentation models in real-world construction scenarios.