STN 02Rocher de la Tortue — perforated sandstone boulder
- Epoch
- 2026-09
- Position σ
- 0.014m
- Δ since last
- -0.082m
current epoch — three stations restated. Pulling an epoch re-reads every station on the sheet — nothing reloads.
Sigmas, epoch deltas and baseline lengths are illustrative of the method, not surveyed results.
Position without a satellite fix.
Every reconstruction here was placed in a real coordinate reference system without a GNSS fix at capture time. Position is recovered visually: a generated world model of the scene, fused with street imagery and open terrain data, resolved into a visual positioning system. Under heavy canopy, in a quarry, beside a cliff face, or in a street where the sky is a sliver, that is the difference between a survey and a guess.
Because the fix is visual, it repeats. A site visited again resolves into the same frame as its first visit, which is what makes change measurable rather than merely visible.
From capture to a measurement you can defend.
- 01CaptureDrone flythrough, street-level panorama, or national LiDAR. No surveyed control point and no GNSS fix is required at capture time.
- 02World modelOpen world-model reconstruction (the Lyra 2.0 / World Labs class of model) turns the footage into geometry, and high-density splats carry the fine surface.
- 03FusionThe reconstruction is registered against fused open geodata — street network, national terrain and elevation models, existing LiDAR coverage — so it lands in a real coordinate reference system.
- 04Visual positioningA visual positioning system resolves where each capture sits without a satellite fix, and re-resolves it on every later visit. Position becomes a measurement rather than an assumption.
- 05Surface and changeThe LiDAR HD reconstruction mesh is the ground-truth base layer, and splats are superimposed on it. Differencing two epochs gives the movement, with its uncertainty stated.
Where the fusion pipeline actually stands.
The stages below are the ones running in this repository. Each reports what has been measured; anything not yet run says so rather than being implied. The LiDAR HD watertight mesh is the ground-truth base layer, and the splats are placed on it.
- 01CaptureCaptures from a world model, a Panoramax panorama and a YouTube flythrough. No surveyed control point required. [measured]
- 02EDWSR watertight mesh73 watertight surface tiles for tile 0650/6813, verified present in gs://fontainebleau-copc-tiles/wasure-checkpoints. [measured]
- 03Georef5 scenes carry a georef record; 0 measured. The rest are labelled manual — a hardcoded yaw, not a survey. [pending]
- 04RecalibrateRefinement against COP30 / AW3D30 elevation and OSM features. Pending a GPU run — no independent residual yet. [pending]
- 05Mesh pruneGaussians inside the watertight surface are rejected (0.25 m threshold). Runs as ETL stage 2.5; no report published yet. [pending]
- 06EncodeSPZ v3 for the Spark viewer plus a SOG archive, GPU-compressed on the L4. [measured]
- 07PublishCurrent published set is v3. The mesh-constrained v4 set awaits the prune. [pending]
What has been built.
Not concepts. Each of these is a working system in the repository, and the Fontainebleau reconstructions are live and viewable in the browser.
Circuit des 25 Bosses, Fontainebleau
Nine reconstructions of a sandstone boulder circuit, geo-registered and published as Gaussian splats that stream straight into a browser: 26.8 million gaussians, 299.7 MB. It proves the open pipeline on terrain that is genuinely difficult: low relief, heavy canopy, and no satellite reception under the trees.
Reference survey · drone · panoramax · flythroughIGN LiDAR HD integration
Ingestion and processing of the IGN LiDAR HD product as the survey-grade ground-truth surface. Splat reconstructions are superimposed on the LiDAR-derived mesh instead of floating free of it.
Base layer · national coverage · franceWasure surface reconstruction
Regularised surface reconstruction from point clouds — the open implementation we build on to turn LiDAR into watertight, measurable surfaces rather than loose point soup.
Reconstruction · open source · point cloudsVisual positioning system
Pose estimation from imagery alone, served as a geospatial pose: position plus orientation with an accuracy figure, for capture that has no satellite fix.
Positioning · openvps · panoramax · streetviewCapture and ETL pipeline
Ingestion, tiling, geo-registration, reconstruction and serving, run as resumable stages on GPU. It is the same pipeline the reconstructions above were produced with.
Infrastructure · gpu batch · serverless
Hazards it is built to read.
Surface change is the signal underneath all of it. The same differencing that measures a moving boulder face also carries fire fuel load, erosion and structural movement — different questions asked of one measurement.
- Forest fire risk — fuel load, canopy structure, and how both shift between seasons.
- Climate change — erosion and drainage changing the shape of a slope across epochs.
- Civil infrastructure — settlement and movement on assets that must not move.
- Ancient cities and heritage — architectural faults and slow structural drift.
- Construction — as-built against design, with progress measured rather than estimated.
Sovereign by construction, not by promise.
The pipeline is assembled from open components and runs on infrastructure you or your jurisdiction controls. There is no requirement to ship site data to a vendor cloud, because there is no vendor cloud in the path.
That is also the commercial difference. Proprietary reconstruction services — the Esri and Pix4D Cloud class of product — license per seat or per project, and the data leaves your estate. Ours is open where it matters, and priced against what you actually measure.
LiDAR and point-cloud processing
PDAL, COPC, tile preparation
pdal-processorGaussian splat pipeline
reconstruction to SPZ and SOG
splat ETLPanoramax ingestion
French street-level imagery
ign-panoramaxMapillary ingestion
crowd-sourced street imagery
mapillaryStreet View ingestion
third imagery source
streetviewData fusion
street graph, GIS, terrain, boundaries
fusionTile serving
streaming to the browser
tile-serverTerrain and elevation
open DEM, hillshade, slope
opentopography
Who is building it.
A small team. Between us we cover the science, the commercial model and the engineering — which is why the pipeline is short and nothing in it is a black box.
- CEO
Charlie Durand
Sets the direction and the commercial model — sovereign, open infrastructure with usage-based pricing against the per-seat reconstruction clouds.
- CSO
Laurent Caraffa
Co-founder. Owns the science: reconstruction, registration, and the change-detection methods the hazard work rests on.
- Technical director
Dhruv Malik
Fullstack engineer. Builds and runs the pipeline, the services, and the browser-based viewer end to end.
Point it at a site you already care about.
Tell us the site and what you need to know about it — movement, erosion, fuel load, or simply a surface you can measure against next season. If we cannot answer it, we will say so.