Djinious
Any standard LiDAR
Point clouds ingested, CRS-aware, and tiled to COPC with a browser preview. Survey scanners, airborne LiDAR and phone captures enter the same pipeline.
LAS · LAZ · E57 · PLY
DjiniousWorld3D replicasA replica is an object,not an output folder.
3D replicas · LiDAR · photogrammetry · videoModelIntegrate
DjiniousWorld captures the physical world, maintains it as a living digital replica, and turns that replica into engineering decisions and simulation environments. Every replica is versioned, multi-representation, multi-epoch and auditable — and every measurement traces to a capture, a run and an accuracy grade it has actually earned.
DjiniousWorld
- applications in the run catalogue
- 24applications in the run cataloguereconstruction, analysis, export, simulation
- accuracy grades
- 4accuracy gradesindicative < relative_cm < georeferenced < survey
- OpenUSD layers per replica
- 6OpenUSD layers per replicareview, analysis, physics, semantics, geometry, source
- Provider API operations
- 13Provider API operationsplus an embedded MCP server
What it is
Real places, reconstructed and kept
Upload a point cloud, a set of photographs or a video; one API call runs the reconstruction and registers every result as a representation of the replica. Capture hardware is a spectrum, not a gate: what differs between a phone and a survey scanner is a declared, propagated accuracy grade — never a hidden quality difference.
Photogrammetry from photos and video
Structure from motion from an image archive or a video, with sequential matching for forward-motion footage, then dense multi-view stereo on a GPU host.
COLMAPGaussian splats
Photoreal splats trained from posed images on the GPU, or built directly from a coloured LiDAR cloud — and LiDAR-anchored training that inherits metric scale.
gsplat · 3DGSSurvey deliverables
Orthomosaics, textured meshes, DSM and DTM rasters with contour lines, and annotation export to DXF, GeoJSON, KML and CSV.
GeoTIFF · OBJ · GLB · LAZMeasure and compare
Distances, areas, volumes with cut and fill, and multi-epoch change detection — every figure carrying the uncertainty its grade allows.
M3C2Simulation-ready scenes
A replica mesh exported as an OpenUSD physics scene with a static collider, then settle-tested with a rigid probe so it is proven simulatable, not just parseable.
OpenUSD · UsdPhysicsInside the product
See it working.
Every capture below is the running product.
01 · Honest grading
A replica is never graded higher than it has earned
Grades are totally ordered — indicative, relative_cm, georeferenced, survey — and every derived representation, measurement and finding inherits the minimum grade across its full lineage. Pure photogrammetry has no metric scale, so it is declared indicative, never higher.
- GCP georeferencing by a closed-form Umeyama fit, with the grade capped by the residual — and RANSAC rejection of a blundered control point
- Metric scale from known lengths: agreeing references earn relative_cm, disagreeing ones stay indicative
- Registration propagates a reference’s grade only as far as the fit and the reference’s point spacing allow
- A measurement on an indicative replica is reported in model units, not metres
02 · Photogrammetry
From photographs to a surveyed site
The five-stage survey application runs structure from motion, dense multi-view stereo, meshing, mesh texturing and the orthomosaic in one call, dispatched to a GPU worker. Ground control is marked in the viewer with live residuals, and the fit moves the camera poses too — so the orthomosaic and textured mesh are rebuilt on the new datum.
- Textured meshes as OBJ, GLB and PLY; orthomosaics as GeoTIFF
- EXIF geotags read from the imagery, with the control’s own accuracy taken into account
- A quality report PDF for every finished run
DjiniousWorld
03 · Gaussian splatting
Photoreal splats, with metric scale when LiDAR is there
The flagship reconstruct application runs structure from motion on the CPU and trains a Gaussian splat on a DGX Spark GPU, the splat stage consuming the SfM bundle automatically. The trainer holds out every Nth image and scores the unseen views, so fidelity is measured rather than asserted.
- Middlebury temple: 23,383 gaussians, PSNR 34.05 dB and SSIM 0.9958 on held-out views
- LiDAR-anchored training on ETH3D “pipes”: metric extent matching the 17 m scan, +4.3 dB PSNR over SfM initialisation
- Splats straight from LiDAR: each coloured point becomes a gaussian, recentred for float32 precision
04 · Time
What changed, with evidence
A replica holds epochs — its state from one campaign after another. M3C2 change detection compares a baseline and a current cloud and colours every core point by its signed distance: added or raised, removed or receded, or below the detectable threshold. A change smaller than the minimum detectable is reported as no change, never as zero.
- A timeline and side-by-side comparison of epochs
- Cut, fill and net volume above a base surface, each with an accuracy band
- Annotations with 2-D and 3-D figures, on the 3-D view and on the map
05 · Simulation
A replica that opens as a physics scene
The replica is an OpenUSD stage, so a mesh becomes a physics scene without a lossy export step: gravity, a static triangle-mesh collider and a ground plane, with convex decomposition available for movable objects. A settle test drops a rigid probe with MuJoCo onto the collider and judges the trajectory — supported, settled, no tunnelling.
- Verified on a 129,642-vertex LiDAR mesh: valid collider, georeferenced origin preserved
- Geometric segmentation into ground, structure and vegetation
AI & agents
Agents operate the platform
Everything the web interface does, an agent can do through the API: create a site, upload a capture, launch a run, poll it and fetch its results. The entire video-to-splat pipeline has been driven end to end through the API with an API key alone.
Provider API
Thirteen declared operations over sites, replicas, epochs, representations, captures, runs and artifacts, with idempotency keys, jobs and provenance.
Embedded MCP server
API key-secured JSON-RPC that lets AI agents operate the app and understand its architecture.
Built-in assistant
A tool-calling assistant with read access to the data, approval-gated write access, and the ability to ask for a choice mid-task.
Trust
Verified means run
A feature is recorded as verified only when a test asserts the behaviour and that test was run — through a browser where it is something a person uses. What cannot run on the build host is recorded as written and unrun, not glossed over.
Validated on open data
Real airborne LiDAR, real multi-view photographs and a real co-registered laser scan, all driven through the API.
Lineage on every result
Each representation traces to the run that produced it, and an annotation is graded from the representation it was drawn on — naming a representation is never naming a grade.
Permissive licensing
COLMAP, PDAL and gsplat are permissive; copyleft tools run only as subprocesses, and non-commercial models are excluded.
Organisations and sharing
Organisations with roles, read-only share links, signed webhooks, and usage metered against quotas.
In the digital thread
What it takes in. What it hands on.
DjiniousWorld does its part of the engineering loop and passes its evidence along — and it works just as well on its own.
Receives
DjiniousEngineeringThe site the project needs as context, and the captures of it to reconstruct.DjiniousWorld3D replicasDelivers
DjiniousEngineeringA graded, versioned replica of the real environment — the physical surroundings of the twin.
DjiniousWorkshopOpenUSD physics scenes with settle-tested colliders, ready for integration and simulation.
DjiniousMapGeoreferenced orthomosaics, DSM and DTM rasters and contour lines of the surveyed site.On its own
On its own, DjiniousWorld is a reconstruction and survey platform: bring LiDAR, photographs or video, and get point clouds, splats, meshes, orthomosaics and measurements with an accuracy grade you can trust.
Use cases
DjiniousWorld in use.
1 worked case.
Book a demo
See it on your problem.
See your own captures become a replica — and see exactly how far each number can be trusted.
- Ingesting a LiDAR cloud and reconstructing from photographs or video
- The replica workspace: point cloud, splat, mesh and orthomosaic views
- Ground control, metric scale and how accuracy grades are earned
- Measurements, volumes and multi-epoch change detection
- Exports — GeoTIFF, OBJ, GLB, DXF, OpenUSD — and driving it all through the API and MCP
Along the thread
Works well with
DjiniousEngineeringOrchestrationThe AI systems engineer: it takes the assignment and carries the whole project through a traceable System Ledger.
DjiniousWorkshopSystem integrationComponents wired into a complete system and closed through GPU physics in NVIDIA Isaac Sim.
DjiniousMapAgent GISA governed geospatial control plane: data is harvested, measured and only then drawn.
DjiniousWorkflowAgent automationGraph automation any stage can call on: build a process, run it, watch it and audit it.