Gaussian Splatting for Surveyors: Georeferenced 3D Scans on the Map

Every few years a 3D capture technique jumps from research papers to job sites. Gaussian splatting is making that jump now, and it is worth understanding before a client asks for one, because a splat pinned to surveyed coordinates is a genuinely useful site record, while a splat floating in arbitrary space is a screensaver.

What Gaussian splatting is, in plain terms

Gaussian splatting builds a photoreal 3D scene from ordinary photos or video. Instead of stitching pixels into a mesh of triangles, it represents the scene as millions of small, soft, colored blobs (3D Gaussians), each with a position, size, orientation, color, and transparency. Render all of them together and you get an image that looks strikingly like a photograph from any viewpoint, including view-dependent effects like reflections and thin structures that meshes handle poorly.

Two properties matter for field work. First, capture is just walking the site with a camera; there is no scanner tripod choreography. Second, splats render fast, including in an ordinary web browser, so a client can orbit a site condition on their laptop without installing anything.

Splats vs photogrammetry meshes vs LiDAR point clouds

Surveyors already have two 3D formats in their toolkit, so the fair question is what a third one adds.

• Photogrammetry meshes are watertight-ish triangle surfaces with baked textures. They are great for measurement and CAD workflows, but they struggle with vegetation, wires, glass, and anything shiny, and a good mesh takes real processing time.
• LiDAR point clouds are direct distance measurements: geometrically trustworthy, excellent for extracting hard surfaces, but visually sparse. Nobody mistakes a point cloud for a photo.
• Gaussian splats sit at the other end: exceptional visual fidelity and fast rendering, at the cost of being a "soft" representation. A splat is a description of how the scene looks, not a set of certified surface measurements. Precise geometry still comes from LiDAR or a controlled photogrammetry workflow.

The honest framing: splats are the best communication format we have had for site conditions, and they complement, rather than replace, the formats you measure from.

Why georeferencing is the whole ballgame

A splat by itself lives in an arbitrary coordinate system with arbitrary scale. Pin it to surveyed coordinates and everything changes: the scan of the eroded headwall now sits at its true position on the project map, in your project's datum, alongside your points and linework. Now it can serve as documentation of existing conditions on a date, context for design, or evidence when someone later disputes what was there.

That last word matters. Survey records get pulled into disagreements years later, and "photoreal 3D capture, tied to RTK-surveyed control, with datum and epoch recorded" is a fundamentally stronger artifact than an unanchored model. It also means all the usual discipline applies: the splat's anchor points should be surveyed with a fixed RTK solution (see what RTK GNSS surveying is), and you should know what reference frame those coordinates are in, because NAD83 and ITRF differ by 1–2 meters in the US. A beautifully rendered splat georeferenced into the wrong frame is a beautifully rendered 2-meter bust.

The open ecosystem is arriving

Until recently, splats were trapped in bulky research formats and bespoke viewers. That is changing quickly:

Together these do for splats what GeoTIFF and 3D Tiles did for imagery and meshes: a capture from any tool can land in any viewer, at real-world coordinates, without a proprietary pipeline in between.

The commercial state of play

Georeferenced reality capture on a phone is already a commercial product category. PIX4Dcatch Pro runs $1,590 per year, and pairing it with the viDoc RTK rover for centimeter-grade georeferencing adds $3,240 in hardware, as of mid-2026 per vendor pricing. That stack works, and it also illustrates the current cost of entry: several thousand dollars before the first capture, with the georeferencing intelligence living in a closed pipeline.

Meanwhile the ingredients are commoditizing from below: capable RTK receivers keep getting cheaper (free NTRIP corrections cover much of the US), phones carry LiDAR, and the splat formats above are open. The gap is software that connects them.

Where Open Survey fits

This is the signature feature we are building Open Survey around: attach a 3D asset, a LiDAR scan or a Gaussian splat, to any survey point. Because the point carries true coordinates, fix quality, datum, and epoch, the attached scan inherits that georeferencing and appears on the web workspace map as a 3D asset at its real position. Capture a splat of the culvert, attach it to the point you shot at its invert, and anyone with the browser link can spin around the culvert exactly where it sits on Earth.

Laan Labs has shipped 3D capture at scale before (the team built 3D Scanner App, with 70M+ downloads across its apps), and Open Survey's approach is deliberately open: open-source software, open formats for export, and a self-hostable backend so the scans that document your sites remain your data.

Practical takeaways

• Treat splats as a visualization and documentation layer; measure from LiDAR or controlled photogrammetry.
• Georeference every capture to surveyed control, and record the datum and epoch with it.
• Prefer open formats (SPZ, glTF with the splat extension, 3D Tiles) so your captures outlive any one vendor's viewer.