Before you choose how to capture it

Every capture method fails somewhere. Here's where, and what it costs to avoid it.

There is no perfect way to capture a building. A phone, a drone, a $50,000 laser scanner and a tape measure all produce real, useful data — and every one of them can quietly produce something wrong if you don't know what to watch for. This is an honest look at how each route actually fails, the software pipeline it puts you through before you get a usable model, and what it costs at every tier — measured against what happens when the same building goes through Skanna instead.

Four ways in

Everything below starts from one of these. What comes out of it — a mesh, a point cloud, a splat, or just numbers on paper — depends on which one you pick, and on the software you run it through afterwards.

Tape & laser distance meter

The oldest method, and still the most honest one. A laser distance meter gives you an accurate individual dimension in seconds — no scale problem, ever, because nothing is being reconstructed. What it never gives you is geometry: no mesh, no point cloud, nothing to walk through. Every wall, corner and opening still has to be drawn by hand from the numbers.

$20–$150

Photogrammetry

Any camera, phone or drone. Walk or fly around the subject taking overlapping photos, and structure-from-motion software works out the camera positions and rebuilds the surface from parallax. No special hardware — which is also its main weakness: the geometry it produces has no inherent size or orientation until you give it one, by hand, afterwards.

Camera or phone you likely own · software from free

Consumer LiDAR

iPhone 12 Pro or later, iPad Pro 2020 or later. The built-in sensor measures actual depth as you walk, so what comes out already carries real-world scale and a spatial position for every point — no photo-matching guesswork. Range is short, a few metres, and it struggles outdoors in bright sun. For interiors it's the fastest accurate route there is.

Device you likely already own · app free–$100/yr

Professional laser scanning & surveying

Terrestrial laser scanners, robotic total stations, and drone-mounted LiDAR. Millimetre-grade accuracy over ranges consumer sensors can't touch, at the cost of five to six figures in hardware and, for a total station, a trained operator to run it. This is the tier legal surveys and boundary work are actually done at.

$15,000–$90,000+

Where NeRFs and Gaussian splats fit

They come up constantly and they aren't a fifth capture method — they're two of the possible outputs of photogrammetry's own input: the same walk-around photos or video, run through a different reconstruction algorithm.

ReconstructionWhat it actually isBest forFalls apart when…
Classical mesh A triangulated surface with a photo-baked texture — the traditional photogrammetry output. Exporting to CAD/BIM, measuring, 3D printing. Textureless or reflective surfaces; anything that moved during capture.
NeRF (Neural Radiance Field) An implicit volumetric model — a trained function that predicts colour and density from any viewpoint. There is no explicit shape to extract. Photoreal fly-through video, view synthesis. You need a clean, editable mesh; the scene is large or unbounded; you need it to render in real time on an ordinary device.
Gaussian splat An explicit cloud of coloured, oriented ellipsoids fitted to the same photos — fast to render, not a solid. Real-time photoreal viewing in a browser. Anything that needs true geometry underneath: measuring, editing, or exporting to BIM without a lossy mesh-extraction step first.

All three can come from the exact same walkthrough video. The difference is only which algorithm you run afterwards — and that choice trades photorealism for usability. A splat looks the most impressive in a browser and is the least usable the moment you need a real measurement, a wall you can move, or a file a structural engineer can open.

What actually goes wrong

Not theoretical edge cases — the failure modes that show up on real jobs, across every method above.

Drift over a long walkthrough

Every frame-by-frame capture — LiDAR or photogrammetry — accumulates a small registration error as you move. Walk far enough, close a loop, and the last wall doesn't quite meet the first one: a visible seam, a corridor that's slightly bowed, a room a few centimetres out of square. Software can partially correct it — loop closure, plane-snapping — but that's a workaround, not a cure. The longer and larger the capture, the more it shows.

Scale and rotation are never free with photos

A photogrammetry, NeRF or splat reconstruction has no idea how big it is or which way is north until you tell it. Get the reference wrong — a mis-measured control point, a GPS reading with a few metres of drift — and every downstream dimension inherits that error, silently. A LiDAR capture doesn't have this problem: depth is measured directly, not inferred from parallax.

Large sites get stitched, cropped and edited

A property too big for one capture session gets assembled from several — and the joins are where accuracy problems hide. On a large US plot, the finished export was several separate captures stitched together, with edges cropped and touched up by hand to hide the seams. That's normal for a job that size — it also means the file you receive is partly an edit, not purely a measurement, and the person who stitched it made judgment calls you can't see in the final render.

The output depends on more than the hardware

Two people scanning the same building with the same phone can get meaningfully different results — overlap between passes, walking speed, lighting consistency, and which processing software (and which settings) turned the raw capture into a mesh all move the number. An accuracy figure you read online is a best case from one dataset, not a guarantee for yours.

Reflective, transparent and blank surfaces

Photogrammetry needs texture and contrast to match points between photos — glass, mirrors, gloss paint and blank white walls give it nothing to lock onto, so it either fails outright or invents geometry that isn't there. LiDAR measures time-of-flight instead of matching patterns, so it handles all of these; its own failure mode runs the other way — very dark, light-absorbing surfaces can return too little signal.

Anything that moves during capture

A person walking through frame, a car passing, curtains moving in the wind — photogrammetry treats every photo as one fixed scene, so anything that moved between frames shows up as a ghost, a blur, or a hole in the mesh. Frame-rate depth capture (LiDAR) is far more tolerant of this: each frame is measured independently, in real time.

The pipeline behind each route

For everything except a tape measure, capture is step one of several before you have a usable, scaled, shareable model — and each step is usually a different piece of software, with its own file format on the way out.

Step Photogrammetry / NeRF / splat Survey-grade laser scanning Skanna
Capture Photos or video, any camera or drone Registered scan positions, a licensed operator Walk-through with the free iPhone app
Register & process Structure-from-motion in RealityCapture, Metashape, Luma AI or similar — often hours on a GPU Scan registration in the scanner's own software (Cyclone, SCENE) On-device, on the walk
Clean & decimate CloudCompare, MeshLab, or the capture tool itself — large clouds need reducing to stay workable Same tools, at a larger scale Automatic on import
Scale & orient Manual — place control points or a known dimension in the same software before export Inherent — the instrument measures real-world coordinates Not needed for LiDAR; a photogrammetry/splat reference brought in must already be scaled elsewhere
View & edit the point cloud Potree, CloudCompare, or a paid viewer licence Same, at a larger scale Trace it directly in the browser
Convert formats PLY, E57, LAS/LAZ, OBJ, FBX, glTF — usually a different tool for each hop Same, plus the scanner's own proprietary formats GLB throughout; PLY accepted directly as a reference
Model it into BIM/CAD Redrawn or traced in a separate CAD/BIM seat Redrawn or traced in a separate CAD/BIM seat Traced directly in BIM Author — the same file
Drawings, twin, exports Commissioned or built separately, per deliverable Commissioned or built separately, per deliverable Generated from the same model

What it costs

Ballpark list prices in USD, current at the time of writing — they move, so check the vendor before you budget against them.

TierHardwareSoftwareTypical all-in
Tape / laser distance meter A meterNone$20–$150
Consumer LiDAR (phone) iPhone 12 Pro+ / iPad Pro, $700–$1,300 if buying new Free (Skanna, Scaniverse) to ~$100/yr (Polycam) $0 marginal if you already own the phone
Prosumer photogrammetry Mirrorless camera or drone, $500–$3,000 Metashape Standard $179 / Professional $3,499 (perpetual); RealityCapture (now RealityScan) pay-per-output or subscription $700–$6,500+
Handheld / SLAM 3D scanner Matterport Pro3 ~$6,000, or handheld units (Revopoint, Einstar) $500–$2,000 Matterport subscription from ~$65/mo; handheld units bundle their own software $1,000–$7,000+ hardware, plus subscription
Terrestrial laser scanner FARO Focus $15,000–$85,000 by model; Leica BLK $20,000–$50,000+ Vendor software (Cyclone, SCENE), often licensed per seat $20,000–$90,000+
Robotic total station $15,000–$50,000 Vendor field/office software, often licensed per seat $15,000–$50,000, plus a trained operator
Drone LiDAR DJI Zenmuse L2 payload ~$12,000–$14,000, on a Matrice 300/350 RTK airframe ~$10,000–$20,000 DJI Terra or third-party processing, often subscription $25,000–$40,000+ system

Sources checked at time of writing: manufacturer and reseller list prices for Matterport, FARO, Leica, DJI, Agisoft and Epic Games (RealityCapture / RealityScan). Used-market and rental prices run well below new-list for the professional tiers.

Where Skanna sits in all this

Not a claim that Skanna is more accurate than a $50,000 laser scanner — it isn't, at range. The claim is narrower: for the size of building most people own, it collapses the whole pipeline above into one free app and one browser tab.

Capability Skanna Photogrammetry / NeRF / splat Survey-grade laser scanning
Real-world scale, no extra step LiDAR measures depth directly Needs a control point or known dimension, by hand The instrument measures real coordinates
Handles reflective / blank surfaces Time-of-flight, not pattern-matching Needs texture and contrast to lock onto
Cost to get started Free app, a phone you likely own ~Camera/drone plus software, free tiers are limited $15,000–$90,000+
View, trace and scale point clouds ~Pre-scaled PLY/splat accepted as a tracing reference — no scale or levelling tools built in; fix that upstream first ~CloudCompare, Potree or a paid viewer — a separate tool ~The vendor's own desktop software — a separate tool
Convert between formats GLB/PLY straight into the same file ~PLY, E57, LAS, OBJ, FBX — usually several tools ~Proprietary scanner formats, exported per job
Model it into a real BIM/IFC object Traced directly in BIM Author Redrawn separately, in a different tool Redrawn separately, in a different tool
Drawings from the same file Generated on your title block Commissioned separately Commissioned separately
Share it as an interactive twin One link, no app ~If the platform you used supports it Not what the instrument is for
Handles very large sites ~Stitch and align multiple scans in BIM Author, floor by floor This is where photogrammetry earns its keep — no ladder, no walking every metre This is where survey-grade earns its keep — one instrument, many known points

Pick the route that matches the job

None of this means reach for the most expensive tool. It means know what you're trading before you commit a building to it — and it's fine to mix routes on the same site: LiDAR for the interiors, a drone photo pass for the roof and façade, a tape measure for the one room you can't get a phone into.

For the wider argument — why buildings lose their information in the first place, and what one accurate model produces instead of a pile of one-off deliverables — see the long version.