From drone video to a 3D model in under 10 minutes: the steps and the timings

People often assume 3D capture needs a survey team, specialist cameras and a day of processing. For a photo-real model you can walk around in a web browser, that is no longer true. We timed every step from a short drone clip to a finished 3D Gaussian splat on one desktop graphics card. The total was 9 minutes 37 seconds. The result is live as our apartment viewer.

The clip

A 29-second, 1080p stock drone clip: a slow arc past a 10-storey apartment tower beside a park, with a city skyline behind. Slow, steady movement around a subject is close to ideal. Fast turns, zooms and long stretches looking at the sky are the things that hurt.

Step 1: frames (part of 3 min 58 s)

We read the video at 12 frames a second and kept the sharpest frame from each third of a second, which gave 88 frames. Picking by sharpness rather than at fixed intervals matters with drone footage, because small vibrations blur some frames and not others.

Step 2: camera solve (3 min 58 s with step 1)

COLMAP, the standard open-source structure-from-motion tool, works out where the camera was for each frame. All 88 of 88 frames were placed, with an average error of 0.48 pixels. Because neighbouring frames in a video overlap, sequential matching is much faster than matching every frame against every other.

One catch: we let COLMAP estimate lens distortion, and the trainer then refused to start until we switched on its mode for distorted cameras (--gut in LichtFeld Studio). Leave distortion in and use that mode, or undistort the images first.

Step 3: training the splat (5 min 39 s)

LichtFeld Studio, an open-source trainer, optimises up to a million small coloured 3D blobs (Gaussians) until renders of the model match the frames. The usual 30,000 training steps took 26 minutes 20 seconds. A quarter of that, 7,500 steps, took 5 minutes 39 seconds. Scored on every eighth frame, held out from training, the short run reached 34.7 dB against 36.1 dB for the full run. By eye, the two are hard to tell apart, so the short run is the one we published.

The total

• Frames and camera solve: 3 min 58 s
• Training, 7,500 steps: 5 min 39 s
• Clip to finished 3D: 9 min 37 s (30 min 18 s with full training)
• Hardware: one RTX 4070. Nothing was uploaded anywhere.
• Download for the browser: 14.4 MB, compressed in the SOG format and shown with the open-source Spark renderer

Turning a model into measurements

One camera gives shape but not size. We set the scale by counting the tower's ten storeys and assuming a typical 3.25 m each, which puts the tower at 34 m, ±12%. A check: the neighbouring building's roof sits at 0.79 of the tower's height, which matches its eight storeys to ten. We tried a passing train as a ruler, but it was too faint in the model to measure, so we went back to storeys.

Tips for your own footage

• Fly a slow orbit or arc around the subject, 20 to 40 seconds, at a steady height.
• Overcast days give even light with no hard shadows baked into the model.
• Avoid zooming. Keep the subject in frame, and keep the sky to a small part of it.
• Every side you want to see must be filmed. What the drone never saw comes out blurred.
• Check with held-out frames, not just by spinning the model around.

On top of this model we built an apartment finder with floors, availability and an AI search. We wrote that up separately in Gaussian splatting for real estate.