One pass in.
A measured model out.

Tesseract turns a single drone flight into a textured 3D model on one laptop, with no second pass and no ground survey. Left of the line is the drone's own video. Right of it is the model, seen from where the drone was. Drag the line to compare.

Keyframe from the drone video — Photo 0.0 s into the pass Model

Held-out check: the model is rendered into photos it never used, and scored against them.

The pass

The problem statement asks for something drone mapping usually avoids: one flight, one angle, and a model accurate enough to measure. Survey workflows fly a grid, twice, with markers on the ground. Disaster response, border mapping and reconnaissance rarely get that chance.

The demo clip is 18.9 s of a consumer drone gliding over a construction site. Tesseract keeps 177 sharp, well-spaced frames of the 568, works out where the camera was for each one, and builds 458,297 textured triangles from them in 753 s.

The site

The reconstructed site seen straight down, with contour lines and the flight path — The model seen straight down: its own texture, contours drawn from its heights, and the drone's path in black with a dot for each second.

Method

Every stage writes what it did and how long it took. These are the demo run's own timings on an RTX 4060 laptop, 753 s in all. Most of it is the depth prior that gives the houses and trees their shape; without it the same clip takes 293 s.

ScreenRefuse or repair a clip before spending time on it: horizon, sky, overlays, cuts.3.3 s
IngestScore every frame for blur and motion; keep a sharp, evenly spaced set.4.1 s
CameraMapAnything reads the lens from 60 frames; no calibration needed.42.5 s
PosesCOLMAP matches features and solves where every frame was taken.74.0 s
DepthOpenMVS measures depth for every pixel on the GPU and fuses it.54.1 s
SurfaceA mesh through the fused points, then MapAnything's depth on every third frame gives roofs, walls and trees their shape and OpenMVS fits it back to the photos.452.2 s
TextureEach face takes its best photo; seams are levelled and unseen faces take the cloud's colour.103.1 s
Level and exportGravity up, then OBJ, GLB, FBX, PLY, LAS and GeoTIFF.19.5 s

Evidence

A number only counts here if something outside the model checked it.

Check	Result	How it was measured
Looks like the photos	24.4 dB, SSIM 0.69	Every tenth keyframe is held back from the build, then the model is rendered into it and compared, pixel by pixel.
Accurate to a metre	0.34 m median	A ten-minute synthetic flight with known ground truth and a DJI-style telemetry file; 98.5% of the cloud within 1 m.
Ten minutes of video	548 s	The same ten-minute file end to end, inside the 15-minute budget.
Every format asked for	6 of 6	OBJ, GLB, PLY, LAS and GeoTIFF from every run, and FBX where the Open Asset Import Library is installed.

The accuracy figures come from synthetic clips, because a real flight with surveyed ground truth wasn't available. Metre-level results need a gimbal angle in the telemetry, which many consumer drones don't record.

Measure a 20-second synthetic flight in metres, with latitude and longitude

Uses

The problem statement names eight applications. Each one needs a few measurements, and the workspace has them all, taken on the model's surface rather than on the screen.

Application	Distance	Height	Area	Volume	Profile	Line of sight	Coordinates	Notes
Border and strategic area mapping	●				●	●	●
Disaster damage assessment		●	●	●			●	●
Urban planning and smart cities	●	●	●		●
Infrastructure inspection	●	●			●			●
Construction progress monitoring		●	●	●	●
Archaeological documentation	●		●		●		●	●
Digital twin generation							●
Military reconnaissance and mission planning	●	●			●	●	●	●

Open the workspace Everything you measure exports as GeoJSON and CSV.

Limits

One angle. A single pass sees the far field only at a grazing angle, and never sees the back of anything. That's physics, not software.
The sky crop. Frames with a horizon lose their top third to keep the sky out of the depth maps, and with it most of the distant houses.
Seams. Each patch of texture comes from a different photo, so exposure changes can show at patch borders.
Scale. Without GPS the model has no metres until one known length is given. With GPS but no gimbal angle, the scale is only as good as the lens estimate.