Why do photogrammetry and Gaussian splatting scans have no real-world scale, and how do you fix it?

The short answer: photos give you shape, not size

Photogrammetry and Gaussian splatting scans have no real-world scale because a set of ordinary photos pins down the shape and proportions of a scene but not its absolute size. One photogrammetry guide from the 3D scanner maker Artec puts it plainly: photogrammetry provides proportion but not scale, and to scale a model there must be at least one known distance.

The reason is geometric. A camera flattens a 3D scene into a 2D image, and a room photographed from a few feet away looks the same as a perfect miniature of that room photographed from a few inches away. Structure-from-motion, the step that works out camera positions and a sparse point cloud from overlapping photos, can recover where the cameras were relative to each other and to the scene, but every one of those distances can be multiplied by the same number without changing a single image. Developers call this the scale ambiguity of monocular reconstruction.

What arbitrary units mean in practice

When the reconstruction finishes, its coordinates are in whatever units the software happened to settle on. On the CloudCompare user forum, the program's developer describes photogrammetry as a scale-less method and notes that point cloud tools simply work in the implicit units of the data they are given. Any distance read straight off the file is a number with no unit attached.

Gaussian splatting inherits the same problem. A 3D Gaussian Splatting scene is usually initialized from a structure-from-motion run, so its colored blobs sit in the same arbitrary coordinate frame. A 2026 preprint on the topic, METRIC-GS by Sooyeun Yang and Jongseong Brad Choi, sums up the consequence: such a reconstruction cannot be measured or used to place augmented reality content until someone supplies a scale.

The standard fixes: scale bars, known distances and markers

The oldest and most reliable fix is to put something of known size in the scene. Professional photogrammetry uses scale bars, calibrated rods with printed targets at known spacings, made from materials whose length barely changes with temperature. Software finds the targets and stretches the model until the gap between them matches the certified length.

A cheaper version works in almost any tool. Measure the distance between two points in the model, divide the real distance by the measured one, and multiply the whole model by that factor. CloudCompare users do exactly this with its point picking tool and its scale command, using a reference as small as 10 centimeters.

Printed fiducial markers can automate the step. The open-source aruco-estimator package for COLMAP detects the four corners of a printed ArUco marker in the photos, works out where those corners sit in 3D from the camera poses, and compares their spacing with the marker's printed size to get the scale factor. Its documentation says the marker must be flat, visible in at least two images, and preferably in five or more for accuracy.

Estimating scale automatically, and why it is hard

Often nobody placed a ruler in the shot, so researchers try to read scale from the scene itself. METRIC-GS tests the two common shortcuts on five indoor scenes with tape-measured ceilings and reports that a floor-to-ceiling heuristic was off by 27 percent and aligning the scene to a metric depth network was off by 35 to 71 percent, depending on the network.

Its alternative labels the Gaussians with a frozen 2D recognition model, then uses typical real sizes of things like doors, tables and the floor-to-ceiling gap as evidence, fused with a robust statistical estimate. The key idea is that the method can refuse: when the evidence is too thin, it abstains instead of guessing, and those scenes can be handed to a heavier metric model called MASt3R. Across 47 ScanNet++ scenes with laser ground truth, the authors report 16.6 percent mean error for the estimator alone, 9.3 percent on the 28 scenes where it chose to act, and a hybrid that matches running MASt3R everywhere while calling it 60 percent less often.

The same paper carries a warning for anyone trusting automatic numbers: a feed-forward metric model that scored 2.3 percent on a benchmark in its training data scored 28 to 35 percent on new captures. For measurements that matter, a known distance in the scene is still the safest scale.