What is point cloud segmentation?
A point cloud represents a scene with individual XYZ samples. Semantic segmentation assigns a class to each point, helping turn a dense survey into interpretable layers such as terrain, buildings or vegetation. The available labels depend on the model and training dataset.
Inputs and outputs
| Input | Output |
|---|---|
| 3D coordinates and the features expected by the checkpoint, such as RGB or intensity. Training also needs point labels. | A semantic class for each processed point, with scores where exposed by the implementation. Preserve the mapping back to the original cloud. |
Where it helps in GIS
- Separate candidate terrain points for further ground-surface processing.
- Review building and vegetation layers in airborne or terrestrial surveys.
- Organize large survey scenes for asset inspection and 3D visualization.
Semantic labels do not give every tree or building a unique identity. Instance segmentation is a separate task; a classified point cloud is also not automatically a clean mesh, building footprint or survey-grade terrain model.
Starting models and methods
RandLA-Net: an outdoor point-segmentation approach using random sampling and local feature aggregation. The Semantic3D implementation provides terrain, vegetation, building and other outdoor labels.
Superpoint Transformer: partitions a cloud into a hierarchy of coherent groups and uses attention between groups. Its DALES checkpoint is a relevant starting point for aerial LiDAR.
Prepare a useful experiment
- Record the CRS, coordinate units, vertical reference and any local origin shift.
- Match feature channels, normalization, sampling and label IDs to the selected checkpoint.
- Split by independent geographic areas before making training tiles.
- Keep point indices through sampling so predictions can return to the source cloud.
- Evaluate each class on local held-out labels and inspect boundaries, sparse areas and rare objects.
LAS and LAZ are containers, not a universal model input. Follow the model-specific loader or build a verified conversion step. Never silently substitute RGB, intensity or missing features. Map model class IDs to an appropriate export classification scheme explicitly.
Learn more
- RandLA-Net official implementation
- Superpoint Transformer official implementation
- Superpoint Transformer checkpoint collection
The category artwork illustrates the task. It is not a measured model result.
Panoptic segmentation and ground filtering
SuperCluster: combines semantic labels and instance groups for supported object classes. Use the DALES panoptic configuration; instance groups are not permanent asset identities across surveys.
CSF: a traditional cloth-simulation ground filter, not a neural AI model. It separates candidate ground and non-ground points without training weights. Creating a DTM requires a separate interpolation and validation step.
This category groups point-cloud labeling methods. The individual entry distinguishes semantic segmentation, panoptic segmentation and traditional ground filtering so their outputs are not confused.
