Point cloud change detection
Point cloud change detection software, with M3C2 and a level of detection
Neurones 3D is point cloud change detection software for Windows. It compares two surveys with M3C2 and a per-point level of detection, so noise is not reported as movement, and measures a scan against a mesh or an IFC model as an exact point-to-triangle distance. In the published test it measured an injected 50 mm lift as +0.04998 m.
By Dr. Florent Poux, 3D Geodata Academy · Updated
At a glance
Short statements, each one checkable. Where there is a number, it was measured, and the guides say on what.
- On a real vegetated tile with an injected 50 mm lift, M3C2 measured a median of +0.04998 m.
- A cloud compared with itself reports 0.0 % changed and no change patches.
- On the demo pair, a 120 mm settlement is found and a 10 mm one is correctly reported unchanged against a measured 14.9 mm level of detection.
- M3C2 gives four verdicts: unchanged, gained, lost and not observed, so an occlusion is not called a demolition.
- Nearest point compares against another survey, a mesh or an IFC model; against a model the distance is exact point to triangle.
- Conformance bands cut at 1, 2 and 4 times the tolerance give a verdict rather than a gradient.
- Out-of-tolerance regions are listed worst first; click one to fly to it.
- An as-built review report splits the site into 8 × 8 zones with stable A1-style labels, each pass, fail or no-data.
- Comparing and viewing are free and offline; the region CSV and the report need a paid edition.
How N3D tells change from noise
Two questions, two methods
Process › Compare & detect change… offers two methods, because “how far is this from that?” and “what changed between these two dates?” are different questions.
| Nearest point | M3C2 | |
|---|---|---|
| Question | Does the survey match the reference? | What was gained or lost between two surveys? |
| Reference | another survey, a mesh or an IFC model | an earlier survey |
| Distance | nearest point, or exact point to triangle on a model | along the normal, averaged in a cylinder |
| Verdict | share within tolerance (0.05 m by default), bands at 1, 2 and 4 times | unchanged, gained, lost, not observed |
| Typical use | as-built checks, deviation against BIM | settlement, erosion, stockpiles, demolition |
Both write a column to the cloud (deviation or m3c2_change) that can be coloured, reviewed
and, with the paid edition, exported.
A level of detection for every point
Two scans of an unchanged wall never agree exactly: the scanner has noise, the surfaces are rough, and the registration is off by a little. M3C2 averages both surveys in a cylinder along the local normal and gives each point the change it would take, at 95 %, to be believed. On the demo pair that threshold was measured at 14.9 mm: a 120 mm settlement is reported, a 10 mm one is reported unchanged. The card gives the four verdicts with their counts and shares, the median level of detection, the largest change and the number of change patches. A cloud compared with itself reports 0.0 % changed, which is the first test any change detector should pass.
Not observed is not lost
A nearest-neighbour comparison always finds a neighbour somewhere, so it cannot say I did not see this. Delete every point in a box from the second survey and it reports a demolition. M3C2 reports the box as not observed: in the published test that share rose from 2.7 % to 5.9 % and no change patch was reported. Material that is gone can only be seen from the earlier survey, which is why the cores default to it.
Align first, then compare
Change detection assumes the two surveys share a frame. Process › Align to reference cloud… registers one onto the other with trimmed ICP and shows the RMS together with the inlier share, the iterations and whether it converged, and reads check this when fewer than 30 % of the points matched. When the two start far apart, the scene graph matches their rooms first. The graph can also compare two surveys room by room, with the change in area, volume and objects and rooms flagged gone or new. On outdoor data, the ground filter makes a change map easier to read; the LiDAR terrain page covers the rest of that workflow, and scan to BIM the check against a design model.
Limits
M3C2 reports a mean displacement in a cylinder, not a peak: a 1.5 m dome reads about 1.28 m at its apex. Against a mesh or BIM model, only Nearest point measures the exact distance. If one out-of-tolerance region spans the whole site, the list has stopped meaning anything, and the card says so: lower the separation (0.25 m by default).
Questions
Is there free point cloud change detection software?
Yes. In Neurones 3D, comparing two clouds, M3C2 change detection and deviation against a mesh or IFC model all run in the free build, offline, on Windows. Writing the region list as CSV and building the report are part of the paid edition.
What is M3C2?
A change detection method that measures the distance between two surveys along the local surface normal, averaging both surveys inside a cylinder. Each point also gets the change it would take to be believed, so noise is not reported as movement.
What is the level of detection in change detection?
The change a point would have to show to be believed. N3D computes it for every point, at 95 %, with the registration error between the two surveys as a setting (0.01 m by default). A difference below it is reported as unchanged.
How does it avoid false alarms from noise and occlusion?
Two ways. The level of detection keeps noise from being reported as movement, and a region with no points in the other survey is reported as not observed, not lost. In the published test, deleting every point in a box from the second survey raised the not-observed share from 2.7 % to 5.9 % and reported no change patch.
Can I compare a point cloud with a BIM or IFC model?
Yes. Add the IFC or the mesh as a layer and use Nearest point: the distance is exact, point to triangle, with conformance bands and a worst-first list of out-of-tolerance regions. With M3C2 the model is sampled rather than intersected.
Do the two surveys have to be aligned first?
Yes, they must share a frame. Align to reference cloud registers one onto the other and reports the RMS with the inlier share; the scene graph start matches two surveys by their rooms with no initial placement. Alignment is rigid and does not rescale.
Can the result go to a client?
Yes, with the paid edition. File › Build report… writes a self-contained HTML as-built review with 8 × 8 pass, fail and no-data zones and a provenance block. On the command line, report exits with 0 on a pass and 2 on a fail, so a batch of sites sorts itself.
The guides behind it
Each guide gives the menu path, the steps, the command-line equivalent and the limits.