Digital twins you can see and measure

Digital twins you can see and measure

How Omniris turns ordinary drone photos of a site into a measurable 3D mesh or a photo-real Gaussian splat, what you can do with each, and how to fly a site so the result is worth having.

Overview

A digital twin, in Omniris, is a 3D copy of a real place built from the photos you already take with a drone. You can turn it, walk around it and look at it from angles nobody photographed, and keep it as a record of how the site looked on the day of the flight. Omniris builds two kinds, from the same photos:

Mesh or splat?

If you are not sure, Omniris suggests one. When an upload of more than 20 photos finishes, it reads the flight path from the photos' GPS, says what kind of flight it looks like and points to the product that suits it. A grid flown straight down suggests a map; an orbit, a perimeter flight or a walk around an object suggests a mesh, or a splat when the photos lack the GPS a mesh needs. The suggestion never starts a job or spends credits, and every product stays on offer.

What the two share

• The same first step. Both begin by working out where every photo was taken from and which way the camera pointed, using the photos themselves. This is called structure from motion. Once one product has done it, the other reuses it: a mesh can be built from a finished splat, and a splat from a finished mesh, without solving it again.
• The same check before you pay. A Data Quality panel scans the folder before anything is charged: how many photos, blur and exposure, GPS, mixed cameras, and the shape of the flight. Problems likely to ruin the result stop the job from starting (some, such as a low photo count, can be overridden after a clear warning); the rest are shown as advice.
• The same viewer. Both open in the Omniris 3D viewer on the web, in the desktop application, in the mobile app and on a shared link, with nothing to install.
• The same vault. Results land in a folder beside the photos they came from, each file with its own custody history.

flowchart LR
  A[Photos in a
vault folder] --> B[Data Quality
check]
  B --> C[Camera positions
solved from the photos]
  C --> D[Splat trained
and cleaned]
  D --> E[Splat, streamed
in tiles]
  C --> F[Scaled and placed
from the photo GPS]
  F --> G[Depth from every photo:
dense point cloud]
  G --> H[Surface built and
painted from the photos]
  H --> I[Mesh in detail levels,
surface and terrain models]
  

Meshes

A mesh is a 3D surface made of millions of small connected triangles, like a very fine net draped over the site. A texture is the picture painted onto that net, taken from your photos, so the model looks like the place. Because Omniris scales and places the mesh using the GPS recorded in the photos, distances on it are real distances.

Making a mesh from photos is called photogrammetry: measuring a place from overlapping photographs of it.

How a mesh is made

Choose Create 3D Model on a folder of photos, or on a finished splat to reuse the splat's first step. The dialog shows the price of each quality setting before you start; from a folder, it also runs the Data Quality check.

Quality settings

Two terrain options are under Terrain options:

• Ground classification: how ground is told apart from everything standing on it. This decides the terrain model and so every volume measured against it; the mesh itself and the surface model are unaffected. The standard method (SMRF) is the default and the better choice on flat or gently rolling ground. The alternative (cloth simulation, CSF) is worth trying where the default leaves too much vegetation or structure in the terrain. Neither has been validated on steep terrain.
• Contour interval: the height between contour lines; the default is 1.6 ft (0.5 m), and the dialog takes the value in metres. Below the reconstruction's noise level, about an inch (3 cm), extra lines trace noise rather than ground.

When a folder's photos sit in subfolders, as a drone's memory card often splits one flight, the subfolders are included automatically and the dialog says so; you can untick that. Thermal photos from DJI dual cameras are set aside automatically, and the dialog says how many. You can also name the site and customer, which lets measurements on that site prefer this model.

What you get

Each one downloads on its own from the model's Download menu, in the viewer or on the job. In the desktop application the files are downloaded from the model's output folder in the vault.

A 3D model from the map-making dialog

The dialog that stitches drone photos into a map can also build a 3D model, with the same viewer and the same measuring tools in 3D. Its Capture Mode has two choices:

• Map: a survey flown with the camera straight down, for an orthomosaic and elevation models. Tick Include 3D to add a 3D model to the map, for an extra charge.
• 3D model: photos taken at any angle, including oblique, cross-grid, orbit, ground-level and handheld. It produces a point cloud and a textured mesh, with no orthomosaic and no elevation model.

For DJI thermal drones whose photos come in visible and thermal pairs, this route can also add a thermal skin: a second copy of the 3D model painted with temperatures, switched with an RGB / Thermal toggle in the viewer, with no charge beyond the 3D model itself. The companion paper Temperatures you can measure covers thermal imaging, and the companion paper on stitching covers maps.

Measuring on a mesh

A mesh is measured two ways, and they answer different questions. In 3D, on the model itself, you can measure anything you can see from any angle: a wall's height, the gap between two columns, the area of a roof slope. On the 2D map, the mesh's surface and terrain models work like a survey: distances along the ground, profiles and volumes. A height map holds one height per spot, so it cannot describe a wall or an overhang; the 3D tools can.

In the 3D viewer

Results follow your choice of feet or metres. Both tools keep themselves honest:

• Only with a real scale. They are offered only on a model that was placed using its photos' GPS. On a model that could not be, the button is shown switched off, labelled Measure (no scale), with the reason, rather than giving numbers in an unknown unit.
• Faces must be flat. If the corners you click are not on one plane, the area is flagged as a projection that under-reads a folded surface. Points in a straight line are refused: they have no area.
• Ground area and volumes are left to the map. The 3D face area is the area of a face. The area of ground inside a boundary, and volumes, are measured on the 2D map, where they are computed against the full height model rather than a few clicked points.

On the 2D map

A mesh's surface and terrain models work in Map Tools on the 2D map just as a stitched map survey's do: distance along the ground, height profiles, and stockpile and cut-and-fill volumes with tonnage, saved and re-measured on the site's later surveys. Where a map survey and a mesh both cover the ground you draw on, the panel says which one it used, lets you pick, and labels every result with its source. The companion paper on stitching describes these map tools in full.

More in the viewer

• Levels of detail. Low, Medium and Full switch between the model's levels. The viewer opens on a light level, and a phone opens on the most detailed level it can hold.
• Point cloud. Show the dense point cloud the mesh was built from, exactly over the model, to check the model against its source. The measuring tools pick on points too.
• Findings. Place an observation (damage, a defect, a question) at a spot on the model with + Finding, and show the findings around the model as markers. The companion paper Findings you can act on describes findings.
• Show on map. Jump from the model to its location on the 2D map. The map's 3D Models layer works the other way: a pin for each model you have, which opens it in the viewer.
• Terrain. Show the surrounding landscape, and in many cities photorealistic buildings, from Google's 3D map around your model, for context. Off until you switch it on.
• Design overlay. Show an imported design model over the as-built mesh, to see where the work differs from the plan. The companion paper From design to measured progress covers designs.
• Source Data. Open the folder of original photos the model was built from.
• Screenshot. Save the current view as a PNG image.

Splats

A Gaussian splat is a newer way of turning photos into 3D. Instead of a surface, the scene is made of millions of small, soft, partly see-through blobs of colour (the "Gaussians"), each with its own position, size, shape and colour. Seen together from any angle, they look like the photographs. There is no surface to get wrong, so the things a mesh struggles with, such as foliage, fine detail and glass, come out looking much more like the real place.

The price is that a splat is a picture, not a measurement. It has no reliable real-world scale, so Omniris does not offer measuring tools on it, and it delivers no surface or terrain model.

How a splat is made

Choose Create 3D Splat on a folder of photos. The dialog shows the price, the Data Quality check and the settings below. GPS in the photos helps but is not required.

Editing a splat

Even a good splat can carry stray blobs, sky leaking into the ground, or a car that drove through the shot. Edit Splat on the 3D viewer's toolbar opens the Splat Editor on top of the model, for the splat's owner.

Selections can be inverted, cleared and deleted, with undo and redo. Your work is saved as a draft while you edit and can be resumed later. If someone else is editing the same splat you are told before you start; you can still go ahead, because each person's edits become a separate version.

Bake turns the edits into a new version, with a name you choose, filed beside the original. The original is never changed. Each baked version can be downloaded as PLY, SPLAT (a web-friendly format) or SPZ (a compressed format), and can itself be edited again. Baking carries no charge.

What you get

• The splat in the viewer, streamed in tiles, on the web, the desktop application, the mobile app and shared links.
• The splat file (PLY), the standard format that splat viewers and editors such as SuperSplat, Polycam and Blender add-ons open.
• Edited versions in PLY, SPLAT and SPZ.

Flying for a good twin

Processing cannot add what the camera never saw. When a twin comes out badly, the fix is almost always in the capture, not in the processing. The good news is that one well-planned flight can feed both a mesh and a splat.

Omniris publishes a free field guide, Capturing drone imagery for high-quality Gaussian splats, with an interactive 3D view of each flight pattern below. This section summarises it, adds what differs for a mesh, and describes the checks Omniris runs on your photos.

Common to both

• Fly straight down and at an angle. Photos taken straight down (nadir) see roofs and ground; photos taken at an angle (oblique) see walls. Nadir alone leaves walls weak; oblique alone leaves roofs and tops weak. On a flat site, a grid flown with the camera tilted to about 45° is the single most important pass for a splat.
• Overlap generously. Overlap is how much of each photo also appears in the next one, along the flight line and between lines. Aim for 80% both ways for a splat. The Omniris mission planner's 3D pattern starts at 80% along the line and 70% between lines.
• See every surface from several directions. At least three distinct viewing directions per surface, and at least three heights around anything tall.
• Keep photos sharp. Use a fast shutter: the field guide recommends 1/1000 s or faster. Fly slower if you cannot. Blur in even a few photos spoils the matching around them.
• Lock exposure and white balance. Set the camera to manual, meter once on the subject, and leave it. Automatic exposure and white balance make the same wall a different brightness and colour from one photo to the next.
• Fly in steady, even light. Overcast light is best. In sun, fly near midday and finish in one session; never split a capture across changing light.
• One camera, one lens setting. Mixing cameras, focal lengths or image sizes in one set causes drift; the Data Quality check warns about each.
• Do not hover or shoot bursts. Many near-identical photos from one spot add processing time without adding views.

Hard subjects

Some things defeat every method, and are worth spotting before you fly:

• Reflective surfaces (glass, polished metal, water) look different from every angle. Expect artefacts; extra coverage from several angles helps.
• Repetitive textures (uniform brick, gravel, car parks) are hard to match. Add low oblique passes from varied angles.
• Thin structures (wires, fences, antennas, bare branches) need closer, higher-resolution passes, and are never perfect.
• Moving things (traffic, people, trees in wind, water) become ghosts or floaters. Fly when the site is still.
• Sky has nothing to match. Frame the subject, not the sky.

What differs for a mesh

• GPS on every photo. A mesh is scaled and placed from the GPS in the photos. If nearly every photo (95% or more) carries GPS, it also guides the reconstruction; below that the reconstruction proceeds without it and is more likely to fail the GPS check. With almost no GPS, no measurable mesh can be made. A splat does not need GPS.
• Exposure shows in the texture. A mesh is painted straight from the photos, so differences in brightness between photos can show as patches on the model. Locked exposure matters for a mesh too.
• A map is not a 3D model. A map (orthomosaic) needs photos taken straight down: in the map-making dialog, photos more than 25° from straight down are left out, and a flight shot mostly at an angle cannot become a map. The dialog says so, shows the flight's camera angle, and offers to build a 3D model instead.

What differs for a splat

• More views, from more angles. A splat reproduces how surfaces look from each direction, so every surface should be seen from a wide spread of directions, not just from above and once at an angle.
• Light must be consistent. A mesh tolerates some lighting drift; a splat turns it into floaters.
• Gaps become artefacts, not holes. Where a mesh leaves a hole, a splat fills in blobs that look right from where the photos were taken and fall apart from anywhere else.
• A walk round at eye level helps. Photographing the façades on foot, under the same light, often improves a splat more than any change to the flight.
• Photo counts. The field guide suggests roughly 200 to 800 photos for a single building, 1,000 to 3,000 for a small site and 3,000 or more for a complex one. How well the views are spread matters more than the count.

Flight patterns

The field guide's patterns, each with a 3D visualiser in the guide. The heights and angles are starting points from field practice, not fixed rules.

Planning the flight in Omniris

The mission planner, on a job or a site, draws a flight over your area and exports it for DJI Pilot 2, DJI Fly, Litchi or AMC; you fly it in your own app. Its patterns are Grid (for a map), Crosshatch (for 3D: two grids at right angles with the camera tilted, −45° by default), Corridor (roads, pipelines, rivers), Perimeter (rings facing inward, for façades, stockpiles and structures) and Orbit (a circle around one structure). You set the detail you want as a GSD (ground sample distance: how much ground one pixel covers), and it works out the height, the line spacing and the photo spacing. It can hold a constant height above the ground, slows the flight so the camera can keep up, and warns when motion blur would exceed a pixel or the height would pass 400 ft.

Checked before you pay

When you start a splat or a mesh from a folder, the Data Quality check reads the photos first and reports what it finds, including:

• too few photos (it asks for at least 20; for a splat, 100 or more is recommended);
• blurry, over-exposed or under-exposed photos, which you can leave out of the job;
• for a splat, how far exposure varies across the set, a sign that automatic exposure was on;
• how many photos are near straight down: for a splat, when 90% or more are, it warns that walls will be weak, and suggests orbiting at about 45° with a pass tilted upward; for a mesh, a set almost entirely straight down is noted as right for terrain, contours and volumes but weak on walls, with a pass at about 45° suggested if you need them;
• missing GPS or camera information, and mixed cameras, lenses or image sizes; for a mesh, photos without GPS stop the job, and photos that form separate groups with a gap between them (often two sites in one folder) stop it unless you confirm it is one site;
• the shape of the flight (an orbit, a perimeter, a mapping grid, a walk round an object), read from the photos' GPS.

Afterwards, look at the result as the field guide suggests: floaters in mid-air, views that fall apart away from where the drone flew, soft edges on façades, and thin or missing areas. Each points to what the next flight needs more of. Treat the first capture of a new site as a trial run.

Viewing and sharing

Meshes and splats open in one 3D viewer that runs in a web browser, with nothing to install and nothing to download first. The viewer never loads a whole model at once: it fetches a light version first and finer detail where you are looking, so even a large site opens quickly.

Public Share 3D on a model's output folder creates a link that opens straight into the 3D viewer; the recipient never sees a folder listing. It can carry a password and an expiry date, and can be revoked at any time. You can also share the folder with colleagues and clients inside Omniris, like any other folder.

The same evidence vault

Meshes and splats are not kept in a separate system. Every result is a file in the vault, filed in a folder beside the photos it was made from, with everything that brings:

• Custody. Each delivered file is recorded in the tamper-evident custody record as an output of the job that made it.
• Originals untouched. Processing never changes your photos, and editing a splat never changes the splat it started from: each edit is a new version.
• Legal holds and retention. Running a job again on the same folder replaces its earlier results, but never a result under a legal hold or retention policy; that file is refused, with the reason.
• Encryption. Photos in an encrypted folder can be processed, results arriving in an encrypted folder are encrypted as they land, and the viewer decrypts an encrypted model as it streams it, for you and for the visitors of a public link alike, keeping no decrypted copy.
• Sites. Name a site when you start the job, and the model is grouped with that site's other flights, maps and measurements.

The companion paper Evidence you can prove describes the custody record, holds and retention in full.

Price

Meshes and splats are paid for in the same credits as every other Omniris product. The dialog shows the price before you start.

• A mesh is priced by the number of photos, then by the quality setting: Preview costs less than Standard, and High more.
• A splat is priced by the number of photos, then by the training iterations and the resolution you choose.
• Building one from the other (a mesh from a finished splat, or a splat from a finished mesh) skips the shared first step, so it finishes sooner; the dialog shows its price.
• Failure costs nothing. A job that fails, or that you cancel, is refunded in full.
• Editing a splat is free. Baking a new version carries no charge.
• Viewing, measuring and sharing are included.

Where it runs

Building a mesh or a splat needs a graphics card (GPU) for its heaviest steps. Viewing needs nothing more than a browser.

Once started, a job is carried through by Omniris itself, not by the browser that started it: you can close the page, and you get an email when it finishes.

What the results claim, and what they don't

A model is only as useful as the honesty of its limits.

• A splat is for seeing. It has no reliable real-world scale, so it is never measured in Omniris and delivers no surface or terrain model. If you need numbers, build a mesh from the same photos.
• A mesh's scale and position come from the photos' GPS. Ground control points are not used for a mesh built as its own 3D model project (the 3D model a stitch makes alongside its map does use that stitch's ground control). The GPS check makes sure the shape and scale agree with the drone's own track, which is what distances, areas and volumes depend on. The model's absolute position on the Earth is only as good as the drone's GPS: with a standard (non-RTK) drone it can be off by several feet (a few metres) as a whole, without changing anything measured on it.
• No GPS, no mesh. A folder whose photos lack GPS is refused before you pay, and a reconstruction that disagrees with its GPS is either not offered a mesh or stopped and refunded, rather than delivered with wrong numbers. A splat can still be made from those photos.
• There is a noise floor. A photo-based surface is not smooth below about an inch (3 cm). Contours closer together than that trace noise, not ground.
• The terrain model is automatic. Separating ground from what stands on it follows fixed rules and can be wrong, most often under dense vegetation. Neither ground method has been validated on steep terrain.
• Two surfaces of one site can disagree. A mesh's surface model and a stitched map's come from different methods and can differ by more than a volume's error allowance. Every map measurement says which kind of survey it came from, so a step in a volume history can be told apart from a change of source.
• Comparing dates is for maps. Side-by-side comparison of two flights as a change map works between stitched map surveys; a mesh's surface is measured, but not offered in date comparisons.
• The 3D face area is a face. It is the area on the plane of the corners you clicked. Corners that are not on one plane are flagged, because the number then under-reads the real surface.
• A splat's quality score is advice. It compares the splat with its photos and flags results below most captures. It is a prompt to look at the result, not a verdict on it, and the splat is delivered either way.
• Some subjects defeat every method. Glass, water, polished metal, thin wires and anything that moved during the flight produce artefacts in both meshes and splats.
• Phones see lighter copies. On the mobile app a mesh opens at the most detailed level the phone can hold, and a splat as a copy with fewer blobs.
• Availability. Everything described here, from upload through processing, measuring, editing and sharing, is in production.

Glossary

3D Tiles

An open standard for streaming large 3D models in pieces, so a viewer loads only what is in view, at the detail the screen can show.

Digital twin

A 3D copy of a real place, built here from drone photos.

Floater

A stray blob hanging in mid-air in a splat, usually caused by inconsistent light, moving objects or too few views.

Gaussian splat

A 3D scene made of millions of small, soft, coloured blobs that together reproduce what the camera saw from any angle.

Georeference

The scale and position that place a model correctly on the Earth.

GLB

A standard single-file format for textured 3D models, opened by most 3D software.

GSD (ground sample distance)

How much ground one pixel of a photo covers. A smaller GSD means finer detail, and usually a lower flight.

Iteration

One round of adjusting a splat's blobs to match the photos.

LAZ

The compressed form of LAS, the standard file format for point clouds.

Level of detail

One of several versions of a model, from light to full, that the viewer switches between so a model opens fast.

Mesh

A 3D surface made of connected triangles.

Nadir

A photo taken with the camera pointing straight down.

Oblique

A photo taken with the camera tilted, so it sees the sides of things.

Overlap

How much of each photo also appears in the next, along the flight line and between lines.

Photogrammetry

Measuring a place from overlapping photographs of it.

PLY, SPLAT, SPZ

File formats for splats: PLY is the standard one, SPLAT a web-friendly one, SPZ a compressed one.

Point cloud

Millions of individually measured points on the surfaces of a site, each with a position and colour.

RTK

A correction system that makes a drone's GPS position much more precise than standard GPS.

Structure from motion

Working out where every photo was taken from, and which way the camera pointed, by matching the photos to each other.

Surface model (DSM)

A height map of the top of everything: ground, buildings and vegetation.

Terrain model (DTM)

A height map of the bare ground, with buildings, vehicles and vegetation removed.

Texture

The picture painted onto a mesh, taken from the photos.