Gaussian Splatting vs Photogrammetry vs NeRF: Which One Can a Machine Use?

Photogrammetry reconstructs actual surfaces, a mesh or point cloud you can measure against, and is usually metrically scaled because survey workflows include control points. Gaussian splatting represents the scene as many small oriented blobs, renders in real time, and looks better, but has no surface to measure and is not scaled by default. One is built for measurement, the other for viewing.

Yes, once scale is resolved and coverage is good enough. MultiSet ingests 3D Gaussian splats as a native input and produces a localization map from them, which devices query for 6-DoF pose at sub-5 cm median accuracy. What you cannot do is point a device at an unscaled splat and expect a trustworthy position, because pose error scales directly with the scale error.

For most practical work, splatting has replaced it. NeRF produces excellent view synthesis but is generally too slow to render in real time, and it carries the same unknown-scale problem. If you already have NeRF output, the useful question is the same one: what fixed the scale?

Measure something whose real size you know: a doorway, a pallet, a standard rack bay. If the reconstruction returns arbitrary units, or metres that are consistently off by a percentage, it is not metric. Do this before judging a reconstruction by how it looks, because visual quality tells you nothing about scale.

If people will measure and document against it, photogrammetry or laser scanning, because you need real surfaces. If people will mostly look at it and walk through it, splatting looks considerably better. If devices need to localize in it, that is a separate artifact derived from either, and it should come from the same capture rather than a second site visit.

Yes. A splat of one area, an E57 laser scan of another and a phone capture of the corridor between them stitch into a single continuous coordinate system. Devices localize across all of it without knowing which method produced which section.