BridgeDiff: A two-stage conditional diffusion model for bridge point clouds completion

Scanned point clouds of civil infrastructure are often incomplete due to scanning limitations, making point cloud completion an essential task for Scan-to-BIM workflows. The complex geometries of structures further challenge point cloud completion methods. Currently, most completion models focus solely on synthetic small object datasets or autonomous driving scenarios. In this paper, going beyond conventional methods, we introduce a conditional diffusion based point cloud completion model BridgeDiff for bridges. The model adopts a two stage strategy, where a coarse completion stage reconstructs the global geometry from incomplete inputs, followed by a refinement stage that enhances local geometric details such as sharp edges and smooth surfaces. To support training and evaluation, a synthetic dataset containing 1000 single-span masonry arch bridges is created and released as open source. Experimental results show that BridgeDiff achieves a Chamfer Distance of 0.025 and an Earth Mover’s Distance of 0.118, outperforming existing point cloud completion methods. Further validation on real scanned bridge point clouds demonstrates high geometric fidelity and strong generalisation capability for infrastructure reconstruction. Our method has the potential to accelerate the digital transformation of the built environment by enabling scalable and data-driven infrastructure reconstruction across design, construction, and asset management.