oriented-det is Apache 2.0. You can install it, read it, change it, and run it on your own machines.

The detector you ship is the one you train on imagery your contract allows. You annotate your classes, you train, you keep the weights. DOTA and HRSC are how the demos were made. Your archive is how you operate.

The dataset terms are below.

What Apache 2.0 covers Link to heading

The oriented-det source, the PyPI package, the training and inference code, the configs, and the documentation are released under Apache License 2.0.

That is the sovereignty story we have been telling since the May announcement and the v0.1.0 release:

  • You run the stack where you need it — on-prem, private cloud, air-gapped, regulated.
  • There is no hosted-inference requirement and no platform lock-in.
  • The code is auditable and forkable. You can keep it for as long as your programme lasts.
  • You may modify it and retrain it on your own data.

Apache 2.0 is a permissive software license. It tells you what you may do with our code. It does not re-license Google Earth screenshots, GF-2 tiles, CycloMedia aerials, or anyone else’s annotations. It does not turn a research checkpoint into a commercially cleared weight file.

A license tag on a Hugging Face repository does not change that. Putting apache-2.0 next to a .pth trained on DOTA or HRSC does not give you DOTA’s images, HRSC’s images, or a commercial right to deploy that network as if the training set were yours.

DOTA: academic use only, commercial use prohibited Link to heading

DOTA is explicit. From the official dataset page:

All images and their associated annotations in DOTA can be used for academic purposes only, but any commercial use is prohibited.

The same page states that images come from Google Earth, GF-2 and JL-1, and CycloMedia, and that Google Earth imagery must respect the Google Earth terms of use.

That is not a Creative Commons license. It is not Apache. It is not “fair use” as a slogan. US fair use and EU text-and-data-mining exceptions are jurisdiction-specific, fact-specific, and a poor foundation for a product you intend to sell or operate as a service. Do not assume they cover a commercial oriented detector trained on DOTA.

In practice:

  • Do not treat DOTA images or labels as training data for a product.
  • Do not treat a DOTA-trained Hub checkpoint as a production model you can ship because oriented-det itself is Apache 2.0.
  • Do not redistribute DOTA imagery under Apache, or imply that we did.

We use DOTA the way the authors allow it to be used: academic baselines, architecture comparisons, and public demos labelled as research illustration — see the static optical demo and the zoo announcements. That is a benchmark. It is not your detector.

HRSC2016: a research ship set, not a commercial grant Link to heading

HRSC2016 (Liu, Yuan, Weng, Yang; ICPRAM 2017) is the standard oriented ship benchmark. oriented-det includes a native loader and Hub recipes because the literature uses it.

The images are collected from Google Earth. The paper presents the set as a public research dataset for ship recognition. We are not aware of an Apache, MIT, or commercial-use license from the authors that would let you treat those images — or a network trained on them — as a product asset.

Google Earth terms still sit underneath the pixels. A third-party mirror that stamps CC BY on a re-upload does not rewrite the original collection.

Treat HRSC the same way we treat DOTA for anything that is not a paper:

  • Fine for reproducing mAP, debugging the loader, and checking that a recipe converges.
  • Not a substitute for licensed maritime imagery in a commercial or operational pipeline.
  • HRSC-trained Hub weights are research checkpoints, not a cleared ship detector.

If your counsel reaches a different conclusion on HRSC than on DOTA, that is their job. Ours is not to blur the two into “public, therefore shippable.”

FAIR1M, SSDD, and HRSID are the same line Link to heading

v0.3.1 added loaders for three more research sets. Apache 2.0 still does not cover their pixels, and none of them has a Hub zoo.

  • FAIR1M. The Kaggle dump in the FAIR1M note is CC BY-NC-SA. The official Gaofen test is closed.
  • SSDD and HRSID. Research SAR chips. The SSDD and HRSID notes train locally and do not redistribute the images.

Same posture as DOTA and HRSC: fine for a benchmark, not a production training set.

What the Hub weights are for Link to heading

The pretrained zoo exists so you can:

  • reproduce published zoo numbers (DOTA: official Task 1; other datasets: held-out val when train does not include it),
  • try odet image-demo on a laptop,
  • compare Rotated RTMDet, Rotated Faster R-CNN, Oriented R-CNN, FCOS, and RetinaNet,
  • start academic fine-tuning experiments.

It does not exist so you can skip buying imagery.

Weights are not magic dust that forgets their training set. A detector trained on DOTA has seen DOTA. A detector trained on HRSC has seen HRSC. We do not Apache-license those datasets, and we do not claim the checkpoints wash the restriction away.

If you need a detector you can operate as a business, train on data your organisation has the right to use.

The production path: your imagery, a compatible license, then oriented-det Link to heading

The solution is not a clever reading of DOTA. The solution is the one EO programmes already know:

  1. Acquire imagery under a contract that allows training machine-learning models and deploying the result (tasking, archive, or a vendor licence that says so in writing).
  2. Annotate oriented boxes on your classes — ships, aircraft, vehicles, tanks, or whatever the programme actually needs.
  3. Train with oriented-det on that dataset. The Apache license is the right instrument for that step: you may run the framework, modify it, and keep the resulting weights as a product of your data, not of DOTA.
  4. Qualify and deploy on-prem. You own the stack, the data path, and the checkpoint.

“Purchased” here means licensed for this use, not merely downloaded. A research dump, a screenshot, or a dataset whose page says academic-only is not a compatible license. If the imagery vendor forbids model training or commercial inference, oriented-det cannot fix that. If the vendor allows it, Apache 2.0 does not stand in the way.

That is sovereignty in the operational sense: a framework you can keep, on data you can defend.

We help teams do exactly this — workshops, consulting, and project delivery around oriented-det, typically from two weeks to three months. The public zoo is how we show the tooling. Your archive is how you ship.

A short checklist Link to heading

AssetLicense postureProduction?
oriented-det code, configs, docsApache 2.0Yes — this is the stack
DOTA images and annotationsAcademic only; commercial use prohibitedNo
HRSC2016 images and annotationsResearch set; Google Earth source; no commercial grant we rely onNo
FAIR1M images and annotationsResearch set; Kaggle dump is CC BY-NC-SA; no Hub zooNo
SSDD and HRSID images and annotationsResearch SAR sets; not redistributed with oriented-det; no Hub zooNo
Hub checkpoints trained on DOTA or HRSCResearch / demo artifactsNo as a shipped product
A model you train on imagery you licensed for MLYours, subject to your data contractYes — this is the path

This article describes how DL4EO draws the line between our software license and third-party datasets. It is written for engineers and programme managers who have to make a procurement decision. It is not a legal opinion, it is not a licence grant beyond Apache 2.0 on the code, and it does not replace advice from a lawyer who has read DOTA’s terms, Google Earth’s terms, HRSC’s paper and distribution conditions, your imagery contracts, and the law that applies to you.

If you are unsure, seek counsel before you train, fine-tune, or deploy.

What we can state without hedging: oriented-det’s Apache 2.0 license is the right license for using the framework and for retraining it on data you own or have licensed. DOTA, HRSC, FAIR1M, SSDD, and HRSID do not become that data because they are famous, public, or convenient.

Written on September 10, 2026 by Jeff Faudi. Link to heading