Gaussian Splats for Surveyors

Most of the Gaussian Splatting content online comes from the games, VFX, and creative coding communities. The captures are beautiful. None of them are surveys.

Surveyors looking at splats for the first time tend to land on one of two reactions. Either: “this is the future of deliverables, I want it everywhere.” Or: “this isn’t metric, it’s useless to me.” Both are partly right.

Here is where splats actually fit in a survey workflow — what they’re genuinely good at, what they cannot do, and how to integrate them into the work you’re already doing.

Why a surveyor would care

Surveyors capture sites. Clients want to understand sites. The gap between those two things is where deliverables live, and it’s bigger than most non-surveyors realise.

A LAS point cloud is the highest-fidelity record of a site geometry you can practically produce. It’s also visually impenetrable to a non-technical client. You can hand a developer 50 GB of LiDAR data and they will look at a Potree viewer for thirty seconds, nod, and ask if you can also send a few photos.

This is not a failure of the data. It’s a failure of the medium for the audience. A point cloud is the right format for engineering work. It is the wrong format for “show me what this site looks like.”

Gaussian splats fill that gap. They are photorealistic, interactive, navigable, and they communicate site state to non-technical viewers in a way that point clouds cannot. They are not a replacement for point clouds. They are a parallel deliverable for a different audience.

For background on the technology itself, see Gaussian Splats: The Next-Gen 3D Reconstruction.

Want to try it now? Open a Gaussian splat in the browser — free, no signup, your file stays on your computer.

What splats are genuinely good at

A short, honest list:

Site walkthroughs. Drone-captured or handheld-captured splats of a site let a client navigate continuously through the space in a browser. Far better than a 360 photo tour. Better than a video flythrough, because the client controls the camera.

Condition documentation. A monthly splat capture of a facade, a bridge, a heritage building, or a piece of industrial equipment creates a visual record of condition over time that is dramatically more informative than photos. Cracks, weathering, vegetation growth — all visible in 3D.

Before-and-after comparison. Two splats taken months apart side-by-side communicate change far more compellingly than a written report. Useful for restoration projects, construction progress, environmental monitoring.

Marketing and proposals. A 30-second splat flythrough rendered out as video, or a live splat embed in a proposal document, lifts a survey deliverable from “PDF with photos” to something prospective clients actually remember.

Stakeholder engagement. Public consultation for infrastructure projects, planning submissions, heritage proposals — anywhere non-technical stakeholders need to understand a site without specialist tools.

What splats cannot do

The honest, blunt list. Splats do not replace measurement-grade outputs, and selling them as if they do will cost you trust.

Measurement. Gaussian splats are not metric. The geometric fidelity varies across the scene, and there is no guarantee the coordinates you read off a splat correspond to real-world distances. Do not measure off splats. If the client needs a measurement, the splat is the wrong tool — give them the point cloud, the photogrammetry mesh, or the CAD output.

Geospatial alignment. Out of the box, splats sit in their own arbitrary coordinate system. Some workflows can co-register a splat to a geographic CRS using the source photogrammetry, but it’s an extra step that loses precision. Don’t assume a splat will overlay correctly on a basemap or GIS.

Engineering deliverables. Splats are not contour drawings, volume calculations, cross-sections, or CAD-ready outputs. They are a visual communication format. The engineering work still needs the underlying point cloud or mesh.

Hard surfaces and edges. Splats can struggle with sharp geometric edges — corners of buildings, the lip of a kerb, the edge of a roof. The rendering will look fine at distance but blur up close. Don’t use splats for inspection work that needs millimetre detail on edges.

Repeatability of geometry. Two splats of the same site captured a week apart will not have identical geometry. The 3DGS training process is stochastic. For tasks that need rigorous geometric comparison, you need point clouds and a proper change-detection pipeline.

If you remember nothing else: splats are a communication format, not a measurement format. Treat them like an excellent rendering, not like a survey deliverable.

The hybrid workflow

The way splats actually fit into professional survey work is alongside the geometric deliverables, not instead of them.

A typical hybrid capture for a construction progress survey:

1. Drone flight for photogrammetry. Standard grid pattern with ground control points. Produces an orthomosaic, a DSM, and a textured mesh.
2. Drone flight for splat training. A second flight at lower altitude with denser overlap and varied angles — orbits around features, oblique passes, occasional close approaches. The frames feed splat training.
3. Photogrammetry processing. Standard pipeline. Outputs are the measurable, georeferenced deliverables.
4. Splat training. Run on the splat-targeted footage. Output is a .ply archive plus a compressed delivery file.
5. Combined delivery. Client portal contains the ortho, the DSM, the contours, the report — and the splat as a visual communication layer alongside.

The marginal cost of producing the splat, once you have photogrammetry capture working, is a second short flight and a few hours of GPU time. The marginal value is significant — it’s often the deliverable the client actually looks at first.

For more on combining splats with other survey deliverables, see The Surveyor’s Guide to Professional File Delivery.

Capture tips for survey-scale sites

The internet is full of capture guides for small, indoor, hobbyist splats. Survey-scale capture is a different problem. A few things that matter at this scale:

Overhead drone coverage

A standard photogrammetry grid (parallel lines at 80% forward and 70% side overlap) is a fine starting point. It will produce a recognisable splat. It will also produce a flat, ceiling-perspective splat with poor coverage of vertical faces and undersides.

For better results, add:

• A second pass at oblique angles (camera tilted 30–45° below horizontal)
• Orbits around prominent features (buildings, structures, equipment)
• Lower-altitude passes for areas the client cares about most

The more varied your camera angles, the better the splat handles view-dependent effects like glass, water, and reflective metal.

Walked-around handheld coverage

For ground-level detail — facade close-ups, interiors, areas the drone can’t reach — handheld phone or camera footage works well. Walk slowly, keep the camera moving smoothly, and overlap your coverage by at least 50%.

Common mistakes: walking too fast (motion blur kills feature matching), walking in a straight line (no parallax), keeping the camera fixed on one feature (the splat lacks coverage of everything else).

Lighting

Splats prefer diffuse, consistent lighting. Overcast days are ideal. Harsh midday sun creates strong shadows that get baked into the splat, and the shadows then move incorrectly relative to the scene during viewing.

Avoid mixed lighting (e.g., flying half the site in sun and half in shadow as a cloud passes overhead). The splat will struggle to reconcile the lighting changes and you’ll see artefacts.

For interior captures, avoid bright window backlighting. Either capture at night, capture early morning with diffuse light, or scrim the windows. Bright windows in the background of an interior splat almost always produce floaters.

Reflective surfaces

Where photogrammetry struggles, splats can excel — water, glass, polished metal, dark vehicles. The view-dependent effects in 3DGS handle reflectivity in a way mesh-based photogrammetry cannot.

This is genuinely useful for survey work that includes water bodies, glass facades, or industrial sites with reflective equipment. Test it before relying on it, but in many cases the splat will produce a better visual result than the photogrammetry mesh for the same scene.

Vegetation

Splats handle vegetation considerably better than photogrammetry meshes. Trees, bushes, grass — all reconstruct as recognisable, navigable forms rather than the blobby, melted shapes a mesh produces.

If your site has significant vegetation that matters to the client (parks, heritage gardens, landscaped developments), splats are often the best visual deliverable available.

Coverage gaps

Splats are unforgiving of coverage gaps. If you didn’t capture a particular angle of a particular face, that face will look blurred, unstable, or filled with floating artefacts in the splat. Photogrammetry meshes degrade gracefully into a textured surface even with sparse coverage. Splats just look wrong.

Plan your capture to cover every face the client might want to look at, from multiple angles. If you’re not sure, capture more.

Processing realities

Splat training is GPU-heavy. A few practical notes:

• Reference implementation: the Inria code is the canonical training pipeline, but it’s slow and memory-hungry. A 4M-Gaussian scene needs a GPU with 24 GB or more.
• Production implementations: gsplat, Brush, Nerfstudio wrap the training process in more practical interfaces and are faster.
• Cloud GPU rental: for teams without dedicated GPU hardware, a cloud RTX A6000 or H100 instance for a few hours per splat is the pragmatic path. Typical training cost: $5–15 per scene.
• Training time: 30 minutes to 4 hours depending on scene size, image count, and target quality.

Expect a learning curve. Your first few splats will have artefacts, missing surfaces, or poor lighting. Capture more aggressively than you think you need to, and the second iteration almost always lands.

Delivery

This is where Swyvl fits. Splat delivery has the same pain points as any spatial deliverable — large files, specialist viewers, non-technical clients — and the workflow is the same. Upload the splat alongside the rest of the survey deliverables, organise into a capture session for the site, share a single branded link. The client gets a portal with the orthomosaic on a map, the point cloud in Potree, the splat in a browser viewer, and the report inline. No installs, no downloads, no support tickets.

If you’re delivering splats by emailing files or dropping them in shared folders, your clients are not seeing them. See How to Share Gaussian Splats with Clients for a deeper look at the delivery problem and how to solve it.

Real applications

Where survey teams are actually using splats today:

Heritage survey before restoration. A splat of a heritage building before restoration work is a permanent visual record. Combined with the LiDAR point cloud, it’s both a measurement archive and a visual one. Particularly valuable for buildings that may be altered or partly demolished during restoration.

Construction progress. Monthly splat captures alongside the photogrammetry deliverable give project managers and clients a navigable record of the site’s evolution. Particularly effective when the client is remote — splats reduce the need for site visits.

Mining face progression. Open-pit mine faces change weekly. A monthly splat capture, alongside the volumetric photogrammetry, communicates the progression to non-technical stakeholders (board members, regulators, community liaison) better than any other format.

Environmental monitoring. Coastal erosion, vegetation change, river morphology — anywhere the visual state of a site matters as much as the numbers, splats are a more compelling deliverable than orthomosaics alone.

Insurance and dispute documentation. A splat of a site at a particular moment is a defensible visual record. If a dispute later arises about the condition of a structure or site at the time of survey, the splat is much harder to argue with than a series of photos.

Marketing and proposals. Splats of past work, embedded in proposals or shown in pitch meetings, demonstrate capability in a way that case study PDFs cannot.

The take-away

Gaussian splats are not a replacement for the geometric deliverables that pay your bills. They are an addition. The cost of producing them, once you have a photogrammetry workflow running, is modest. The value of having them in your delivery — particularly for non-technical clients — is significant.

Be honest about what they can’t do. Don’t sell them as measurement tools. Don’t promise geospatial accuracy. Treat them as a visual communication layer, not a survey product, and they will earn their place in your deliverable stack.

The clients who see one will start asking for them on every job. That is the surveyor’s case for splats in one sentence.