MultiSet accepts metric-scaled 3D Gaussian Splats as a native VPS input.
Upload a .ply splat file alongside a poses.json with camera trajectories. The platform generates an XR-ready map while preserving the photorealistic fidelity that makes Gaussian Splatting revolutionary.
These four are often discussed as competing ways to reconstruct a space. They are better understood as four different things to store, each good at a different job. Only two of them are usable as a localization map without further work.
| Gaussian splatting | Photogrammetry mesh | NeRF | Point cloud | |
|---|---|---|---|---|
| What it stores | Oriented gaussians with colour | Triangles and texture | A learned radiance field | Discrete measured points |
| Renders in real time | Yes | Yes | Usually not | As points |
| Metric scale by default | No | Usually | No | Yes |
| Editable | Hard | Yes | No | Yes |
| Usable for localization | Only if metric-scaled | Yes | Rarely | Yes |
| Best at | Photoreal viewing at speed | Geometry you can measure and edit | Research-grade view synthesis | Survey and as-built accuracy |
This is the row that decides whether a Gaussian splat is useful to a machine, and it is the one most easily missed, because it is invisible in the thing you are looking at.
A splat reconstructed from images alone is internally consistent but scale-free. The geometry is right in proportion and wrong in absolute terms: the room might come out at 0.8 or 1.3 times its true size and the render looks perfect either way. For viewing, that does not matter at all. For localization it is fatal, because a pose returned in the wrong scale puts a label in the wrong place by an amount that grows with distance.
A map that is not metric is not a map. It is a picture of a place.
Metric scale comes from something in the capture that knows real distance: LiDAR depth, a calibrated stereo rig, a scanner's own measurements, or control points surveyed in the space. If a splat was produced by a pipeline that had access to one of those, it carries scale. If it came from phone video and nothing else, it usually does not.
The practical test is to measure something you know in the splat, a door height or a pallet width, and see whether it comes back correct. MultiSet ingests metric-scaled splats directly. For one that has lost its scale, the fix is to recover it from a reference measurement or from the source capture rather than to re-shoot the space.
This is the scan-agnostic argument in its clearest form, and Gaussian splatting is where it shows best.
The same capture produces a splat a human can look at, walk through and show a client, and a feature map a machine localizes against to a few centimetres. One is human-readable, the other machine-readable, and they come from one pass through the building rather than two.
Most of the market produces one or the other. Visualization tools give a beautiful reconstruction that no device can position itself against. Localization vendors give a map that works and that nobody wants to look at. Producing both from a single capture is what removes the second site visit, and on a live plant or a secure facility that second visit is the expensive part.
