Skip to Content

Drone photos: a local processing workshop

Orthophoto, terrain model, volumes and report, computed on a PC with an NVIDIA card and open source tools: what it makes possible, and its limits.
November 15, 2026 by
Drone photos: a local processing workshop
Federico Solutions, Myrko Federico

Drone photos: a local processing workshop

Field: photogrammetry, geospatial, open source software

Period: July to September 2026, then in production use

Client: internal project, Lab 77 (R&D)

A drone brings back a few dozen or a few hundred photos. What the client wants is not photos: it is an accurate map, a terrain model, a volume in cubic metres, a readable report. In between lies a heavy computation, usually handed to subscription software or an online service. I wanted to find out how far one can go with a single PC fitted with an NVIDIA graphics card and open source tools. Here is what I built, why locally, and its limits.

Context

This work started with DataTerra, my territorial intelligence project. I needed a workshop where I could turn the photos from my own flights into deliverables, without depending on a vendor for every project. The same processing code now runs on my PC and on the project's online platform: it was written once, and it runs in both places.

Problem

  • A dozen tools to chain together: photogrammetry, terrain model clean-up, volume calculation, layout, geographic information system. Each has its own format and logic.
  • A heavy computation: rebuilding a scene from photos needs a powerful graphics card, and hours if done badly.
  • Figures that must be honest: software always displays an accuracy. You still need to know whether it was measured in the field or merely computed by the software itself.
  • Repetition: every mission repeated the same manual steps, with as many chances of error.

Solution

1. One mission, one folder, one click

Each flight becomes a mission: one folder for the photos, one for the computation, one for the deliverables. A small local web portal creates the mission, runs the steps and downloads the results, with no command line. Processing recipes are written once and replayed identically.

2. Open source tools, each in its place

The photogrammetry engine is OpenDroneMap, through WebODM. Terrain models and volumes are reworked with GDAL and PDAL, the mapping and the report are produced by QGIS without its interface, and the 3D model is computed by OpenSplat, then cleaned in SuperSplat. Each tool lives in its own Docker container: one can be updated without breaking the others.

3. Deliverables a client understands

At the end of a mission, the deliverables folder holds an orthophoto (a distortion-free photographic map), a surface model and a terrain model, heights above ground, contour lines, a vegetation index, a volume calculation (cut, fill, net) with three ways to define the base or a comparison between two flights, a 3D model, a PDF report and a QGIS project that can be reopened. All measurements are also stored in a single file: they are computed once, and never recomputed elsewhere.

4. Figures that say where they come from

Without control points laid on the ground and precisely surveyed, the accuracy displayed by the software is only internal consistency between the photos. So I had every figure labelled: "internal" without control points, "checked" with them. And when a capture does not allow a measurement, for example façades that are invisible on a vertical flight, the report says so instead of hiding the data.

Result

The chain has been in production since summer 2026. On my reference mission, about twenty photos, photogrammetry takes around twelve minutes; regenerating all deliverables from that computation takes about two minutes. When I improve a recipe, I only rerun the second part.

Why locally

  • No subscription: all the tools are open source. The cost of one more mission is the electricity.
  • The data stays with me: the photos of a site do not leave the PC until I decide to deliver.
  • I understand every step: when a figure is surprising, I can trace it back to the file that produced it and check it.

The limits

  • You need the right hardware: Linux and an NVIDIA card. My card was so recent that the public Docker images of two tools did not recognise it: they had to be rebuilt.
  • Accuracy is earned in the field: without control points, measurements are consistent with each other, not guaranteed in absolute terms.
  • The volume is an earthwork calculation: it measures a stockpile or an excavation seen from above. The 3D model of a building is a different deliverable, which needs a flight all around it.
  • It is not a product yet: the tool serves my missions. Making it available to other operators is still to be prepared, licence included.

What I take away

  • Measure before adding. I compared extra options of the photogrammetry engine on the same mission: no gain, rather a slight setback. The only real lever was ground control points.
  • The bottleneck is not always where you think. The slowness of the 3D model came from overly ambitious default settings, not from the graphics card. "Preview, balanced, quality" presets settled the matter.
  • Check the files before fixing. Twice, a defect blamed on the processing was settled by reading the files produced: this avoided two false fixes.
  • Name things. "3D volume" meant an earthwork calculation to one person and a building model to another. A tooltip ended the misunderstanding.

Going further

Do you have a heavy, repetitive task that currently depends on subscription software? I design custom tools, built on open source software, that stay with you. See custom applications and automation, or book a 20-minute discovery call.

Sources

Tags: photogrammetry, drone, orthophoto, digital terrain model, earthwork volume, QGIS, OpenDroneMap, open source software

AI assistant memory: make it remember, on your server
Keep the essentials, find by meaning, link memories together: a persistent memory hosted on your own server.