Every decade a survey crew climbs my viaducts with a Leica and comes back down with a lie of abundance — hundreds of millions of points where a few thousand would have told the truth. A terrestrial scan of a single arch returns more geometry than any BIM model on the desk can carry, and yet almost none of it is information. It is the same flat spandrel sampled ten thousand times. Point cloud simplification is the discipline that decides which of those samples I actually need to keep documenting a structure meant to carry the morning crush for another century.

This is a foundation, not a compression setting. The scanner is generous to the point of excess; the engineer has to be precise about what to forget.

←TODAY: In 2026 one terrestrial scan of a rail viaduct returns hundreds of millions of points — far more than any scan-to-BIM model can carry. →3012: By 3012 the as-built is continuous rather than a survey event, and the risk shifts from too many points to a structure that quietly forgets its own redundancy. Fulcrum: A scan is documentation only if you can defend every deleted point with a tape measure against the wall.

What it is

Point cloud simplification is the deliberate thinning of a dense 3D scan so that it stays light enough to work with and honest enough to trust. A laser scanner or photogrammetry rig measures many points on the external surfaces around it; the raw output is redundant, noisy, and unevenly spaced. Simplification answers exactly one question, as the PAZ concept library puts it: which of these points can we throw away without losing the object? The engineering name for the good answer is rational decimation — remove points where the surface is planar and predictable, keep them where curvature, edges, or salient features carry structural meaning. A flat wall can be described by a handful of points. A window reveal, a cornice, a column capital, the shear key at the foot of a pier cannot.

Why it works

It works because geometry is not uniformly informative. On a flat spandrel, every point is predictable from its neighbours, so most of them add nothing but weight. At an edge or a curve, each point is the only witness to where the surface turns. Simplification exploits that asymmetry: it estimates local surface variation — curvature — and spends its point budget where the geometry earns it.

Modern pipelines run in two moves. First they extract feature points via curvature estimation, so edges are protected before anything is deleted. Second they partition the remaining cloud — k-means plus Possibilistic C-Means, or graph filters — so density stays uniform while topology survives. The probabilistic-membership variants published in Frontiers in Physics in 2024 let a single point belong partially to several clusters, which is precisely what stops a cluster boundary from being cut cleanly through a feature. The two failure modes are symmetric and both are fatal: an over-decimated scan silently deletes a wall’s true face, so the model is light but wrong; an under-decimated one crushes the modeller and the machine, so the model is honest but unusable. Feature-preserving simplification is the narrow band where a scan-to-BIM workflow stays both light and geometrically honest.

Origins

The field is roughly three decades old and it has pivoted twice. The first generation, from the late 1990s and 2000s, was purely geometric. Curvature- and clustering-based decimation thinned dense scans while trying to hold onto sharp edges and surface topology — Song & Feng’s New Point Cloud Simplification Algorithm, built on mean-curvature segmentation, is the archetype: keep points where the surface bends, drop them where it doesn’t.

Around 2019 the ground moved. Graph-spectral formulations — Qiao et al. at ICME’19 — recast the cloud as a signal living on a graph and filtered it, borrowing the mathematics of signal processing wholesale. Then the learning-based networks arrived: PCS-Net in 2022 made the trade-off between feature preservation and density uniformity an explicit, trainable objective rather than a hand-tuned heuristic. The most recent work, GP-PCS in 2024, models surface variation with Gaussian processes on Riemannian manifolds for one-shot, reconstruction-free thinning — it decides what to keep without ever building the mesh first. Three decades, one unchanging problem: honest deletion. It is the same instinct a railway keeps about its own network. You do not maintain every metre of rail with equal attention; you spend the maintenance window where the load and the geometry demand it.

In practice

Atelier: In the PAZ Atelier we treat simplification as an editorial act, not a slider. Before a single point is dropped, we scan the envelope with the Leica, merge and clean in CloudCompare, and decide — by feature class — where honesty is non-negotiable: reveals, junctions, load paths, the true face of a wall. Only then do we thin, and the Rhino + Grasshopper model that comes out is deliberately light, but every deleted point was a decision we could defend if someone put a tape measure to it. This matters most exactly where naive decimation fails: on ornament and on structure. Clustering-and-saliency work in npj Heritage Science (2025) showed that saliency-weighted thinning keeps carved detail an ordinary voxel filter would read as noise; the same instinct protects a corbel or a rivet line on a heritage bridge. City-scale LiDAR work in the International Journal of Digital Earth (2025) makes the sharper point for anyone thinning a whole corridor: one global voxel size ruins façade, ground, and vegetation at once — the real engineering is choosing density per feature class. PAZ has covered the sister discipline before, in our 3D-reconstruction notes on the Notre-Dame point cloud — Andrew Tallon’s one billion laser measurements from around 2015, captured for scholarship and then, after the 2019 fire, the reconstruction reference. That is the whole argument in one asset: the as-built model no one thought they would need is the one that lets you rebuild to the thing itself rather than to a guess. Your Monday move: pick one feature class on your next scan — the reveals, say — and set its target density before you touch the global voxel size, so the office is deciding what to keep rather than what the default happened to keep.

Hack

Hack: Thin one viaduct scan by a single voxel cell and read the honest deletion as one ratio before you trust any of it. Load a raw cloud off the scanner — the source’s 100k-point reference is a fair proxy for a real arch — collapse every 0.03 m (3 cm) cell down to one point with Open3D’s voxel_down_sample, and print the before-and-after count. The ratio tells you instantly whether you are looking at real geometry or ten thousand copies of a flat spandrel. This step is topology-blind on purpose: the voxel grid softens every edge equally, which is exactly the baseline you want to feel before you reach for the curvature-weighted thinning of Song & Feng’s mean-curvature lineage, which spends its budget where the surface bends. Run it on a viaduct scan and the flat arch barrel evaporates while the numbers stay legible.

import open3d as o3d
pcd = o3d.io.read_point_cloud("viaduct_scan.ply")
thin = pcd.voxel_down_sample(voxel_size=0.03)  # one 3 cm cell -> one point
ratio = len(pcd.points) / max(1, len(thin.points))
print(len(pcd.points), "->", len(thin.points), f"({ratio:.1f}x)")

If that ratio is ten-to-one and the model still holds its edges, you were carrying nine redundant points for every real one. If the edges go soft, you have just seen — in a single number — why curvature-aware thinning exists.

The trajectory is clear: from curvature heuristics, through graph spectra, to learned and manifold-modelled thinning that reconstructs nothing and preserves everything that matters. For the engineer documenting a century-scale asset, the promise was never smaller files — it is scans that stay honest at any weight, so the leap from site to BIM stops being a lossy compromise and becomes a faithful translation. Before your next scan-to-BIM job, write down the one feature class you would defend under a tape measure, and thin everything else around it. The scanner will keep being generous. Our job is to stay precise about what we choose to forget.