Turning a video into something walkable

There are two routes. They are not equivalent, and the difference is worth two minutes of reading before you film anything.

Lane A — your phone recommended

Scaniverse reconstructs real Gaussian splats on the phone, free and unlimited. Photoreal result, minutes of work, nothing to install here.

Use this for anything you will show someone.

Capture guide

Lane B — this machine proof of concept

Plain .mp4 in, 3D out, reconstructed here with COLMAP on the CPU. Produces a sparse point cloud rendered as splats — recognisably the place, not photoreal.

Use it when you have footage and no phone app, or to see the geometry.

Lane B — the one command

Film a slow walk around or through the space, then run:

node tools/reconstruct.mjs --video "C:\path\to\walk.mp4" --name my-building

It extracts frames, runs structure-from-motion, converts the result to a splat .ply, and adds it to your library — one step, no decisions. When it finishes it prints the URL to open.

What it is doing

Frames — FFmpeg pulls ~2 frames/second and drops near-duplicates and blurry ones (mpdecimate), capped at 150. Feeding every frame in is the classic way to make this take all night and produce nothing better.
Structure-from-motion — COLMAP finds features, matches them between frames, and solves for where the camera was for each one, plus a 3D point for every matched feature. All on the CPU; this is the slow part.
Splat conversion — each 3D point becomes one small isotropic Gaussian, so the cloud opens in the same viewer as a phone scan.

What it will and will not give you

Will — correct geometry and camera path, real colours, walkable, measurable after calibration.
Will not — solid surfaces or photoreal fill. That needs Gaussian-splat training, which requires a CUDA GPU. This machine has an Intel Arc iGPU. The step is skipped rather than faked.

Useful options

Option	Default	When to change it
--fps	2	Raise to 3–4 if you walked quickly; lower for a slow pan.
--max-frames	150	Raise for a large building; each extra frame costs matching time.
--matcher	sequential for video exhaustive for photos	Chosen for you by input type. Video frames are ordered, so sequential is correct and far cheaper — exhaustive compares every pair and was 65% of the reference run's time. Override to exhaustive if your walk doubled back a lot.
--min-track	3	Lower to 2 to keep more points if the result is sparse; raises the floater count.
--images	—	Use a folder of photos instead of a video.

Filming for Lane B

Walk slowly. Motion blur is the number-one cause of a failed reconstruction.
Keep consecutive frames overlapping heavily — think 70%.
Even, diffuse light. Avoid shooting into the sun or through glass.
Texture helps; blank white walls give the matcher nothing to hold onto.
Return to where you started, so the loop closes.

Measured timings from the reference run on this machine are in LEARNINGS.md, and the full log of any run is written to data/laneb/<name>/reconstruct.log.