In May I announced that Oriented-det was coming. Today it is here.

Oriented-Det v0.1.0 is officially released under the Apache 2.0 license, available on PyPI as oriented-det, and fully documented. This post covers the main ideas behind the framework and how to get started.

What is Oriented-det? Link to heading

Oriented-det is a lightweight PyTorch library for rotated object detection in aerial and satellite imagery. It targets the workflows where orientation actually matters — ships in a harbor, aircraft on an apron, vehicles in a dense parking lot — and where axis-aligned boxes simply miss the point.

The framework covers the full detection loop: geometry primitives, rotated IoU and NMS, DOTA dataset loading, three baseline models, and a clean training pipeline with mixed precision, checkpointing, and TensorBoard support.

Install in one line (Python 3.10+; PyTorch is installed separately for your platform):

pip install oriented-det

Four things that drove the design Link to heading

1. Oriented boxes as the default, not an afterthought Link to heading

Standard YOLO workflows use axis-aligned bounding boxes. Oriented detection is often bolted on as a patch. In Oriented-det, the core representation is the rotated bounding box — five parameters: center x, center y, width, height, angle — and everything else is built around it. The geometry module, IoU computation, NMS, dataset loader, and model heads are all natively oriented.

2. Sovereignty by design Link to heading

Oriented-det is built to run wherever you need it: on-prem, private cloud, air-gapped environments. There is no hosted inference requirement, no platform lock-in, and no call home. You own the weights, the data path, and the deployment surface.

This matters in practice for institutional and government EO programmes where data residency and supply chain auditability are requirements, not nice-to-haves.

3. A clear license for real deployments Link to heading

The framework is released under Apache 2.0. For teams building commercial or institutional products, license clarity is part of the technical decision. There is no ambiguity about what you can ship.

4. Lean enough to stay in production Link to heading

MMRotate is the reference for oriented detection research. It is also heavy: MMCV, MMDetection, MMEngine, compatibility matrices. Oriented-det is lightweight and evolvable, yet also stable and maintainable.

Three baseline models Link to heading

Oriented-det v0.1.0 ships three detectors, all trained on DOTA:

Oriented R-CNN — horizontal RPN with midpoint-offset proposals routed to an oriented ROI head. The current recommended baseline.

Rotated Faster R-CNN — oriented RPN with oriented ROI head, a two-stage detector that is more expensive but useful for dense scenes.

Rotated RetinaNet — single-stage detector with oriented anchors and focal loss, useful as a fast baseline.

All three produce true oriented boxes: the angle is a predicted output, not zero-initialized and ignored.

Pretrained weights are hosted on the Hugging Face Hub and can be downloaded with:

odet pretrained download oriented_rcnn_dota_le90_1x

Getting started Link to heading

Install the package, then download a pretrained checkpoint to run your first inference:

pip install oriented-det
# PyTorch must be installed separately for your platform:
# https://pytorch.org/get-started/locally/

odet pretrained download oriented_rcnn_dota_le90_1x

Training from scratch or fine-tuning on your own data follows the same pattern — edit a JSON config, point it at your DOTA-format dataset, and run:

odet tile-dota /path/to/your/dataset/train
odet train --config configs/oriented_rcnn/dota_le90_3x.json

The full workflow — tiling, training, prediction, evaluation — is covered in the Getting Started guide.

What’s next Link to heading

The next post in the series is a hands-on walkthrough — oriented object detection on macOS, in pure Python — bus detection on the bundled DOTA demo image using the pretrained Oriented R-CNN, from uv pip install to result.jpg on Apple Silicon (--device mps). The Oriented R-CNN 3× checkpoint announcement covers the final DOTA le90 pretrained model and eval-val results.

Beyond that, the series will cover:

  • Technical deep-dive: geometry conventions, angle normalization, and the le90 convention on DOTA
  • Lessons learned on DOTA baselines: what transfers to production EO, what doesn’t
  • Splitting large EO images without label leakage
  • Oriented mAP: what to report and how to read it
  • Post-processing strategies for dense harbor scenes
  • Packaging a sovereign detector for production

If you are evaluating oriented detection for an operational EO system and want to discuss requirements, feel free to reach out.

Written on June 22, 2026 by Jeff Faudi. Link to heading