Point cloud editing
Crop the cloud down to the region you care about, then voxel-downsample it to a workable size. Grey is the original, blue is the result; export to XYZ or CSV when the counts look right. Nothing is uploaded.
Drop a point cloud file here
Supports .pcd (ascii / binary) and .ply (ascii / binary, either endianness). Parsing and editing happen on this page.
Crop
Axis-aligned box; bounds are inclusive. If the scene sits at an angle, straighten it with the coordinate transform tool first.
Voxel downsampling
One point per cell. Smaller cells stay closer to the raw data; larger cells compress harder.
How it works
- Drop in a PCD or PLY; the page parses the point count and bounding box.
- Adjust the crop range or the voxel size — the preview updates live, grey for the original and blue for the result.
- Export XYZ or CSV once the count looks right.
Cropping is an axis-aligned box
Six numbers make an axis-aligned box, and bounds are inclusive: a point exactly on the boundary is kept. The fields start at the cloud’s own bounds, so cropping changes nothing until you pull a side inwards.
That also means it cannot crop at an angle. To cut along an object’s own direction, straighten that direction first with point cloud coordinate transform, then come back — one extra step, but every step stays unambiguous.
Voxel downsampling: centroid or first point
Space is split into cells of size size and each cell keeps one point. The two choices differ more than they look:
- Centroid: the average position of the points in the cell. Smoother surface and more even spacing, but the output coordinate is computed — no such point exists in the raw data. Usually fine for registration and mapping.
- First point: keeps a point that was actually measured. Use it for accuracy evaluation, or when raw intensity and timestamps must survive.
Start with 1/100 of the longest edge — the page suggests that value. The count drops roughly with volume; if it barely moves, the cell is smaller than the point spacing.
A sequence that works
- Straighten the scene first if it sits at an angle; skip if it does not.
- Crop away what you do not need: ground, distant walls, clutter behind the sensor.
- Voxel-downsample to a workable size.
- Export XYZ / CSV, or feed it straight into registration.
Things that happen
| Symptom | Usually this |
|---|---|
| Nothing survives the crop | Wrong units (a millimetre cloud with metre ranges), or min/max swapped |
| The count barely drops | Voxel size is smaller than the point spacing |
| Far fewer points than expected | Voxel too large, or the crop box already excluded most of the cloud |
| Blue is completely hidden under grey | The settings do not change the cloud, so the two overlap — expected |
| Intensity or colour looks blurred | Centroid mode averaged the extra fields too; switch to first point |
Why local only: point clouds often carry customer site coordinates or unpublished scene data. There is no backend here, so the file never leaves your browser.