# Measured accuracy, and what it was measured against.

> Accuracy measured against references the platform did not produce: a median of 1.7 cm at 46 checkpoints of the public EuroSDR benchmark, terrain against national models, classification against hand-checked clouds.

Page: https://finitor.cl/accuracy

Measured 2026-09-29 – 2026-10-04.

## Position of photogrammetry results

Surveyed targets that the run never saw, found in the photos and triangulated from the run's own cameras. Horizontal and vertical are given separately; RMSE is the root-mean-square error over the targets.

- Survey camera with 8 ground control points: Median 1.7 cm horizontal · 1.6 cm vertical. RMSE 12.5 · 13.8 cm: 5.3 · 4.9 cm over the 39 targets seen in 10 or more photos; the 7 at the edge of the block, seen in fewer, reach 0.6 m DJI Zenmuse P1 (35 mm) at 50 m, 1,024 photos, 0.7 cm ground sample; photo positions from plain GPS; 8 targets as control, 46 held back as checkpoints; solved from the photos' recorded positions and angles, a mode the composer does not offer yet Targets surveyed by Newcastle University (EuroSDR RPAS benchmark)
- RTK drone, no ground control: 8.7 cm horizontal · heights 0.64 m low Autel Evo II Pro RTK, 176 photos at 50 m, nadir only; 5 targets as checkpoints Targets surveyed with GNSS RTK (1–2 cm)
- RTK drone with 3 ground control points: 10 cm horizontal · heights 0.56 m low The same flight; 3 targets as control, the other 2 held back as checkpoints Targets surveyed with GNSS RTK (1–2 cm)
- Consumer drone GPS, no control: Shape within 13 cm (median) · position as good as the drone's GPS (0.54 m vertical here) 18 photos; shape compared after removing the GPS offset The dataset's published reconstruction

## Terrain models

Automatic ground classification on national LiDAR with every class removed, gridded on the national terrain model's own cells and compared cell by cell.

- Flat urban park (Geneva): 0.24 m RMSE — the same as swisstopo's own points across its two survey years (0.235 m) swisstopo LiDAR, 1.1 M points swissALTI3D, 0.5 m
- Steep forest and village (Grindelwald): 95% of cells within 9.4 cm; median 0.1 cm · finds 82% of the ground on the steepest slopes swisstopo LiDAR, 10.2 M points swissALTI3D, 0.5 m
- Cloud LiDAR run, end to end: 0.29 m RMSE Geneva, coloured from the orthophoto swissALTI3D, 0.5 m

## Point classification

Scored point by point. F1 combines how many points given a class truly are that class, and how many of that class were found; 1.00 is perfect.

- Deep classifier (add-on): Ground 0.94 · vegetation 0.94 · buildings 0.86 Autzen stadium, 10.65 M points Hand-checked professional classification
- Standard classifier: Ground 0.93 · vegetation 0.56 · buildings 0.22 Autzen stadium, 10.65 M points Hand-checked professional classification
- Standard classifier, coloured from the orthophoto: Ground 0.98 / 0.88 · vegetation 0.98 / 0.67 · buildings 0.55 / 0.70 Geneva (flat) · Grindelwald (steep) swisstopo's own classes

## Volumes and measurements

Objects whose size is known exactly, measured with the desktop viewer's own tools.

- Stockpile volumes: 753.8 m³ vs 754.0 · 268.4 m³ vs 268.1 (within 0.1%) Synthetic piles on noisy ground, irregular point spacing Exact cone volumes
- Volume over time: 452.5 → 754.4 → 1,022.1 m³ vs 452.4 → 754.0 → 1,022.1 The same pile over three dated surveys Exact volumes
- Cloud and desktop give the same result: Point clouds agree to 4.2 cm (median) The same 18 photos processed in the cloud and on a laptop Each other

## What decides accuracy on your survey

- Ground control decides absolute accuracy. Without it and without RTK, positions are only as good as the drone's GPS — metres — even when the shape is right to centimetres.
- RTK photos are used at their own accuracy automatically, DJI and Autel alike. Without that step, the same Autel flight measured 0.62 m instead of 8.7 cm.
- A single-height, nadir-only grid can bias heights: the RTK flight above came out about 0.6 m low. Where height matters, fly a cross pattern or add oblique photos, and use ground control.
- Where nothing was measured — water, deep shadow, the edge of the flight — elevation models are filled from the surroundings unless you turn that off.
- Accuracy falls at the edge of a flight, where a point is seen in few photos: on the benchmark above, targets seen in fewer than 10 photos were more than five times worse than the rest. Fly past the area you need, and place control near its edges.
- You can measure your own survey the same way: name some surveyed points CHK in the ground-control file and the run reports the error at each, on its page and as a PDF.
- LiDAR without colour classifies vegetation poorly; colour it from an orthophoto first. Buildings are the weakest class for the standard classifier.

## Reference data

- EuroSDR RPAS benchmark datasets: J. P. Mills, M. V. Peppa, A. Alma'Amari, L. Davidson, J. Goodyear, N. T. Penna, 2023, doi:10.5281/zenodo.10059050, CC BY 4.0.
- helenenschacht and Brighton Beach: OpenDroneMap sample datasets (ODMdata).
- Geneva and Grindelwald LiDAR, swissALTI3D and SWISSIMAGE: © swisstopo.
- Autzen Stadium: PDAL sample data, with a hand-checked classification.

## More about Finitor Cloud

- [Finitor Cloud: photogrammetry and LiDAR in the cloud](https://finitor.cl/)
- [Cloud photogrammetry](https://finitor.cl/fotogrametria-en-la-nube)
- [LiDAR classification](https://finitor.cl/clasificacion-lidar)
- [Point cloud viewer](https://finitor.cl/visor-nube-de-puntos)
- [Pricing](https://finitor.cl/pricing)
- [Features](https://finitor.cl/features)
- [Demonstration: a finished survey](https://finitor.cl/demo)
- [How it works](https://finitor.cl/how)
- [Tutorials](https://finitor.cl/learn)
- [Finitor Desktop, the desktop application](https://finitor.cl/viewer)
- [Contact](https://finitor.cl/contact)
- [Terms and privacy](https://finitor.cl/terms)