Open datasets
Real surveys at the size AIDM is built for, published on Hugging Face under CC BY 4.0. Every image is exactly as the drone recorded it, with RTK-corrected positions. Each set has a manifest with the SHA-256 of every file and a stored zip you can send as is to the task API.
| Dataset | Images | Size | Drone | Used for |
|---|---|---|---|---|
| aidm-vineyard-small | 73 | 0.8 GiB | DJI Mavic 3M | Every test on real images, in CI and in development |
| aidm-vineyard | 145 | 1.6 GiB | DJI Mavic 3M | Matching on repetitive vegetation |
| aidm-industrial-rtk | 503 | 2.4 GiB | DJI Matrice 4E | The default benchmark: oblique 3D and accuracy, at the size of a normal job |
| aidm-solar-park | 905 | 7.9 GiB | DJI Mavic 3M | The large benchmark: repetitive panels, reflections, densification and meshing |
All images are 5280 x 3956 pixels. The sets were flown for flight plan testing under the open category, with no customer involved.
Licence
CC BY 4.0. Attribution: "AIDroneMap, provided by ITALDRONI". You can use them for anything, including commercial work and your own benchmarks, as long as the attribution stays.
Download
The images:
hf download AIDroneMap/aidm-vineyard-small --repo-type dataset --include "images/*" --local-dir datasets/vineyardThe zip, for a client that sends the engine one file:
hf download AIDroneMap/aidm-vineyard-small --repo-type dataset --include "*.zip" --local-dir .Pinning a version
Every version is a git tag, starting at v1. For tests and benchmarks, reference the commit hash the tag resolves to, never a branch, and verify the files against manifest.json. AIDM's own CI does exactly that.
Why these sets
The public test sets used by drone mapping projects are small and old. Measuring future hardware on an 18-image survey from 2013 says little. These sets have the image count, the resolution and the surfaces (vegetation, panels, roofs) that make today's jobs hard. A set of 1,500 or more images, for split and merge and distributed work, comes later.