What happens to a point cloud

Six steps, in this order, on every submission. You can stop after any of them, run one on its own, or hand over a stage list of your own.

1. Inspect, before spending anything

   The header first: bounds, point count, returns, the classification histogram, the coordinate reference system and what is missing from it. Horizontal and vertical EPSG codes are read, stamped and reprojected as a first-class record rather than guessed at.

   This is a read of the header, not of the file, so it answers on a survey-sized cloud in the time it takes to open one. A file with no vertical datum, a unit that disagrees with its own header, or duplicate points across strips is something you want to know before a job runs rather than after.

2. Classify

   Noise out, ground separated from everything standing on it, vegetation split by height above ground, buildings found as planes. The ground filter is chosen from the site itself rather than fixed: a progressive morphological filter suits a flat corridor and a cloth simulation suits a steep forested slope, and picking the wrong one is the difference between a terrain model and a smoothed hillside.

   Measured on a five-point-eight-million-point parcel on a twenty-four per cent forested slope: sixteen seconds from raw, unlabelled points.

3. Model the ground and the surface

   Rasters at the cell size you ask for, each written with its coordinate reference system attached: bare earth, the surface including everything on it, and the height of the canopy between them. A colour preview is written beside each one so a reviewer can see the result without loading it into anything.

4. Contour and section

   Contours from the ground points at your interval, with index contours on their own layer. Cross-sections at stations along an alignment, each with a long profile, as drawings or as the underlying numbers.

5. Tile, for a browser

   An overview and a level-of-detail pyramid, so a survey-sized cloud opens in a browser rather than being downloaded. The same cloud can be written as a cloud-optimised point cloud instead, for a desktop client that reads pieces of a file over the network.

6. Report

   What the run did, what it found and what it wrote: the classification before and after, the parameters each stage actually used, the findings worth a reviewer's attention, and every artifact with its size. As a page to read and as a machine-readable record, because a deliverable that cannot be checked is a deliverable somebody has to trust.

What goes in and what comes out

In

Files come in as LAS or LAZ, and E57, PLY, CSV, XYZ and PTS are imported to LAZ.

Out

Deliverables go out as LAZ, GeoTIFF, GeoJSON, DXF, shapefile with a .prj, E57, PLY, CSV, OBJ, glTF, LandXML, COPC and 3D Tiles.

Installed dependencies

The engine is pure Rust and needs no PDAL or PROJ installation to run its native commands.

Three ways to drive it

While it runs

While a job runs it publishes a ledger a client can poll: state, phase, percent complete, events with stable codes, the run's own log lines and the report.

That matters more than it sounds. A job on a survey-sized cloud is minutes of work, and a progress bar that is a guess is worse than none; the phases and the percent come from what the run has actually finished.

How it is measured

Two real deliveries, and the machine each ran on. Neither is a benchmark and neither is a promise about your data — a point cloud's shape decides most of its cost.

A substation survey

Measured on a thirty-five-million-point substation survey on a ten-core laptop: the full preset — powerline, clearance at five metres, trees, footprints, roads and bridges, native COPC and the PDF report — takes seventy-eight seconds.

The same survey, relabelled

On that same substation survey, starting from every point relabelled as unclassified, the conductor and structure extraction reproduced the labelled run exactly.A reproduction of one delivered dataset, not an accuracy rate. No precision or recall figure is published for powerline extraction because none has been measured across more than one site.