Reality captureGeospatial & territory

Open data in, graded replicas out: airborne LiDAR, real photographs and a laser scan

DjiniousWorld was validated end to end on public datasets — the Autzen airborne LiDAR cloud, the Middlebury temple photographs and the ETH3D “pipes” scene — every step driven through the API, every result carrying the accuracy grade it earned.

DjiniousWorld3D replicas

The Autzen airborne-LiDAR replica in DjiniousWorld: a 110,000-gaussian splat built from the LiDAR cloud, beside analyses for change detection, terrain rasters, cloud registration, metric scaling and GCP georeferencing, each with its accuracy grade.

LiDAR points, Autzen: 110,000LiDAR points, AutzenRGB, intensity, classification
mesh vertices from the LiDAR cloud: 129,642mesh vertices from the LiDAR cloudPoisson, 259,374 faces
PSNR on held-out views: 34.05 dBPSNR on held-out viewsSSIM 0.9958
from LiDAR-anchored training: +4.3 dBfrom LiDAR-anchored trainingETH3D “pipes”

Why open data

A reconstruction platform is easy to demonstrate on its own fixtures. DjiniousWorld was instead run on the canonical public test sets, so every figure below can be checked against data anyone can download.

Airborne LiDAR: the Autzen Stadium cloud

110,000 points in projected coordinates, ingested and then worked through the analysis catalogue.

Ingest

COPC and a browser preview of all 110,000 points.

pointcloud.ingest

Mesh

A Poisson surface of 129,642 vertices and 259,374 faces.

analysis.mesh

Segmentation

Ground 87,392 · structure 1,403 · vegetation 21,205 points.

analysis.segment

Terrain

DSM and DTM GeoTIFFs at 1 m and 34,798 contour segments at 5 m.

analysis.raster

Volumes

About 2.0 million m³ above a 411.09 m base, ±2,973 m³ — the band that the grade allows.

analysis.measure

Physics

The mesh as an OpenUSD physics scene, settle-tested with MuJoCo: valid collider, georeferenced origin preserved.

export.usd_physics_validated

The grade ladder, exercised

Georeferencing

Four ground control points recovered the transform exactly and earned georeferenced. With a fifth, blundered by 18.8 m, RANSAC rejected it and the grade held.

Registration

A 1.5° rotation recovered to about 0.02° — and honestly capped at indicative by the reference’s 1.34 m point spacing.

Metric scale

Two known lengths of 100 m and 50 m agreeing on 2.0× — relative_cm, metric but local.

Photographs to a splat

Structure from motion
The 47 Middlebury temple photographs into a 7,444-point sparse cloud and an SfM bundle.
Splat training
On a DGX Spark GPU, densified from about 7,400 initial points to over 23,000 gaussians.
Held-out scoring
With 41 images for training and 6 held out: PSNR 34.05 dB and SSIM 0.9958 on views the trainer never saw.