OrientedDet v0.1.1

Added

• ProbIoU ROI regression for Rotated Faster R-CNN (roi_box_reg_main_loss_type: probiou + Smooth L1 aux).
• Hub slugs rotated_faster_rcnn_dota_le90_3x (83.42% eval-val mAP50) and rotated_faster_rcnn_dota_le90_1x (77.57% eval-val mAP50).
• dataset.train_includes_val config flag (Airbus Playground: train on all folds; val fold for monitoring only).
• Source provenance metadata in training runs (git_commit, package version, config hash).
• Eval reports under docs/eval-reports/; make eval-val full-tile protocol documented.

Changed (MMRotate parity)

• ROI regression loss: encoded-space Smooth L1 on all 5 channels (MMRotate), replacing radian periodic angle loss that under-weighted angle gradients vs MMRotate.
• Oriented R-CNN: MMDet avg_factor for midpoint RPN and oriented ROI losses; training RPN proposals no longer score-filtered; ROI matching defaults to rotated IoU (roi_use_hbb_for_matching: false); oriented RoIAlign uses first 4 FPN levels only.
• Rotated RetinaNet: separate cls/reg 4-conv towers with 3×3 prediction heads; P6/P7 via LastLevelP6P7 on C5; rotated IoU assignment; encoded L1 reg loss with avg_factor normalization.

Breaking

• RetinaNet checkpoints from before this release are incompatible (head.convs / 1×1 heads / extra_fpn_conv removed). Re-train or use Hub weights published after this change.

OrientedDet v0.1.0

OrientedDet v0.1.0

First public release of OrientedDet.

Install from PyPI:

pip install oriented-det

Highlights

• Core geometry (Polygon, QBox, RBox) and transforms
• Rotated IoU, NMS, and optional GPU kernels
• DOTA loader, tiling, augmentations, and oriented mAP evaluation
• Airbus Playground CSV dataset support
• Three baseline detectors:
  • Oriented R-CNN
  • Rotated Faster R-CNN
  • Rotated RetinaNet
• JSON config training via odet train
• odet CLI for training, inference, metrics, dataset tools, and pretrained downloads
• Pretrained weights on Hugging Face Hub: dl4eo/oriented-det-pretrained

Model Zoo

DOTA le90 pretrained checkpoints include:

Model	Slug	eval-val mAP50
Oriented R-CNN 1×	oriented_rcnn_dota_le90_1x	74.79%
Rotated Faster R-CNN 3×	rotated_faster_rcnn_dota_le90_3x	76.41%
Rotated RetinaNet 3×	rotated_retinanet_dota_le90_3x	71.52%
Rotated RetinaNet 1×	rotated_retinanet_dota_le90_1x	64.14%

Download weights:

odet pretrained list
odet pretrained download oriented_rcnn_dota_le90_1x

Links

• PyPI: https://pypi.org/project/oriented-det/
• Source: https://github.com/DL4EO/oriented-det
• Pretrained weights: https://huggingface.co/dl4eo/oriented-det-pretrained