AI point cloud classification
AI point cloud classification, with a confidence for every point
AI point cloud classification in Neurones 3D is one click, offline, with a machine-learning model matched to the capture: indoor scans, aerial LiDAR, utility corridors or street scenes. Each point gets a class and a confidence, a person can review and correct the result in a grid, and the class travels into a LAS export in its proper slot.
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.
- Four trained classifiers ship in every build, including the free one.
- Indoor, ten classes (ceiling, floor, wall, structure, window, door, table, seat, storage, clutter): 81.9 % and 82.3 % held-out per-point accuracy.
- Outdoor (LiDAR), aerial land cover (ground, low, medium and high vegetation, building, other): 86.2 % held-out.
- Powerlines, for utility corridors: 87 to 88 %, with instances so every tree, span and pole is its own object.
- Streetscape, fifteen classes of mobile-LiDAR street furniture: 31 of 32 traffic-sign objects found.
- Every classification writes a per-point confidence: the share of the random forest’s trees that agreed.
- The ground filter separates ground with F1 90.3 % against IGN LiDAR HD’s own labels, with no training data.
- In a LAS export the class fills the classification byte; segment and instance ids ride as 32-bit extra bytes.
How it classifies, and how you check it
Segments first, then classes
The scene is segmented into coherent surfaces first, and each segment is described by its shape, orientation, height and neighbours, together with point and neighbourhood features. A random forest labels it. Classifying segments rather than loose points is what keeps a wall one class along its whole length.
A confidence you can use
A forest votes. The share of trees that agreed on each point is kept as a confidence column,
so the uncertain parts can be coloured, thresholded and reviewed first, instead of trusting a
single accuracy figure measured on somebody else’s data.
A person in the loop
The labelling grid reviews components with brushes, box and lasso, and can keep the model’s labels as a baseline beside the human ones. That is what can be signed; a model marking its own work is not.
Questions
Which point cloud classification models does N3D include?
Indoor (ten classes), Outdoor aerial LiDAR land cover, Powerlines for utility corridors, and Streetscape for mobile-mapping street furniture (a development preview). All four ship in every build.
How accurate is it?
Held-out, per point: indoor 81.9 % and 82.3 %, outdoor 86.2 %, powerline 87 to 88 %; streetscape found 31 of 32 traffic-sign objects. The confidence column shows where the model was unsure on your data.
Does it need a GPU or an upload?
Neither. The models are random forests run on the CPU, on your machine.
How do I correct mistakes?
In Label Components: brushes per class, box and lasso in the view, a bulk selector by size and place, and locked classes. Keep a baseline first and the model’s labels are exported beside yours.
Can I train it on my own data?
Not inside the app. Labels can be corrected and per-segment features exported with classify-features for training elsewhere.
The guides behind it
Each guide gives the menu path, the steps, the command-line equivalent and the limits.
confidence column, and split a powerline corridor into individual trees, spans and poles.
Understand the scene
Label components
A supervised review grid over the components in hand (primitives, segments or classes), with a brush per class, box and lasso in the view, a bulk selector by height, size and footprint, and a baseline so the model and the human can be compared.
Understand the scene
Unsupervised segmentation
Break the scene into coherent surface patches (walls, floors, table tops, pipe runs) with no training data and no parameters to tune, and get a segment column for every point.
Understand the scene
Ground filter
Drop a simulated cloth on the upside-down cloud to fit the ground, label every point ground or off-ground, and measure each point’s height above that ground, with no training data and parameters you can adjust live.
Deliver
Export a point cloud (LAS, PLY, CSV, USD, web)
One dialog for every point cloud export - format, subsample, colours, label columns and scalar fields - with a LAS writer that puts the class in the classification byte and every id in an exact 32-bit extra byte.