Skip to content

Changelog

All notable changes to OrientedDet will be documented in this file.

[Unreleased]

Added

  • Random rotate (train only) — preprocessing.enable_random_rotate / random_rotate_prob / random_rotate_angle_range (degrees). MMRotate PolyRandomRotate after flips (auto_bound=False). FCOS HRSC 1×/3× and Oriented R-CNN / Faster R-CNN HRSC 3× use p=0.5 ±20°. Oriented R-CNN / Faster R-CNN 1× and DOTA stay off. ±180° on FCOS 6× diverged after epoch 12.
  • HRSC2016 dataset loader (dataset.format: hrsc2016) — official XML + ImageSets, single-class ship, le90 via the DOTA polygon path. Recipes: configs/oriented_rcnn/hrsc2016_le90_1x.json, configs/oriented_rcnn/hrsc2016_le90_3x.json, configs/rotated_faster_rcnn/hrsc2016_le90_1x.json, configs/rotated_faster_rcnn/hrsc2016_le90_3x.json, configs/rotated_fcos/hrsc2016_le90_1x.json, configs/rotated_fcos/hrsc2016_le90_3x.json. Optional DOTA export: odet hrsc-to-dota.
  • resize_mode: keep_ratio — MMRotate-style long-edge scale without square pad; training/inference then apply pad_size_divisor (bottom-right). HRSC Oriented R-CNN, Faster R-CNN, and FCOS 1×/3× use this canvas. Oriented R-CNN uses Smooth L1 main + ProbIoU aux 0.1. Faster R-CNN keeps DOTA ProbIoU main + Smooth L1 aux 0.1. HRSC FCOS recipes use decoded rIoU (lr 2.5e-3), not L1. FCOS HRSC 3× uses lr_scheduler_gamma: [0.1, 0.5].
  • training.lr_scheduler_gamma accepts a number (same factor every drop) or a list (one factor per lr_scheduler_milestones entry).
  • Hub slug oriented_rcnn_hrsc2016_le90_3x (90.41% eval-val mAP50) — Oriented R-CNN 3× on HRSC2016 test (rotate ±20°, gamma 0.1); report docs/eval-reports/oriented_rcnn_hrsc2016_le90_3x/.
  • Hub slug rotated_faster_rcnn_hrsc2016_le90_3x (88.77% eval-val mAP50) — Faster R-CNN 3× on HRSC2016 test; report docs/eval-reports/rotated_faster_rcnn_hrsc2016_le90_3x/.
  • Hub slug rotated_fcos_hrsc2016_le90_3x (88.34% eval-val mAP50) — FCOS 3× decoded rIoU on HRSC2016 test; report docs/eval-reports/rotated_fcos_hrsc2016_le90_3x/.
  • dataset.drop_easy_empty_tiles — with tile_metrics_csv, drop train tiles that are vacuous true negatives (tp=fp=fn=0) before max_train_samples and hard-tile oversampling. Empty tiles with false positives stay and can be oversampled. Keep filter_empty_gt: false so those hard empties remain in the loader.
  • dataset.class_tile_oversample_classes — optional train-tile oversampling by GT class presence (no metrics CSV). class_tile_oversample_factor (default 1.0) and class_tile_oversample_min_count (default 1). Reuses the hard-tile sampler; weights multiply when both paths are on. Lookalike routing labels are never a match; unknown names are warned once and ignored.
  • Class-agnostic final NMS on Rotated FCOS and Rotated RetinaNet via model.nms_class_agnostic / production.nms_class_agnostic (default false, same contract as two-stage). DOTA / HRSC recipes stay class-aware.

Fixed

  • make eval-val / odet preds score floor — no longer uses production.score_threshold or train-val evaluation.score_threshold (often 0.3). Resolver is CLI → evaluation.preds_score_threshold → 0.05. Deploy / train val are unchanged.
  • Diagonal flip angle — MMRotate RRandomFlip(direction='diagonal') mirrors box centers and keeps θ (early return). We incorrectly applied π − θ, so diagonal samples (and diagonal + random rotate) had boxes at the wrong orientation. Horizontal / vertical flips were already correct.
  • keep_ratio + pad_size_divisor 32 no longer crashes on P6: torchvision max_pool on an odd feature map (e.g. 576×800 → 9×13) is anisotropic; stride derivation now falls back to configured [4, 8, 16, 32, 64] (MMRotate).
  • keep_ratio collate bottom-right pads a batch to a shared H×W (e.g. 480×800 + 576×800) so torch.stack works. Per-image content_size is unchanged; odet preds stays one image + divisor pad.

Removed

  • Hub slugs oriented_rcnn_dota_le90_1x, rotated_faster_rcnn_dota_le90_1x, rotated_faster_rcnn_dota_le90_3x_ce, rotated_retinanet_dota_le90_1x, and rotated_fcos_dota_le90_3x_kfiou_aux. DOTA zoo is one 3× slug per architecture; eval reports stay under docs/eval-reports/.
  • HRSC2016 6× recipes (oriented_rcnn/hrsc2016_le90_6x.json, rotated_fcos/hrsc2016_le90_6x.json). 3× ±20° is the long schedule; 6× was only +0.2 mAP on Oriented R-CNN and FCOS 6× never beat that 3×.
  • Notebook geometric transform classes (Rotate, HorizontalFlip, VerticalFlip, DiagonalFlip, Compose, OrientedTransform). Image+box augs are only apply_random_train_flips / apply_random_train_rotate and apply_flip_to_* / apply_rotate_to_*.
  • export/ (ONNX / TensorFlow tooling), the odet export-* subcommands, and the oriented-det[export] extra. This repo no longer ships a TF/ONNX export pipeline.

Changed

  • DOTA 3× Hub deploy floors — production.score_threshold is eval-val global F1 − 0.05 on all four slugs: Oriented R-CNN 0.7 (F1 0.75), Faster R-CNN 0.6 (F1 0.65), RetinaNet 0.45 (F1 0.50), FCOS 0.2 (F1 0.25). Recipes, Hub sidecars, and docs updated. make eval-val still uses 0.05.
  • HRSC 3× Hub deploy floors — same F1 − 0.05 rule: Oriented R-CNN / Faster R-CNN 0.85 (F1 0.90), FCOS 0.2 (F1 0.25). Recipes, Hub sidecars, and docs updated.
  • Hub slug rotated_faster_rcnn_dota_le90_3x refreshed from runs/rotated_faster_rcnn/20260901-095802 (83.46% eval-val mAP50, was 83.42%). Weight stem is now rotated_faster_rcnn_r50_fpn_dota_le90_3x-9951acc6. Deploy production.score_threshold 0.6 (eval-val F1 0.65 − 0.05). Report docs/eval-reports/rotated_faster_rcnn_dota_le90_3x/.
  • loss.focal_weighted now scales Rotated FCOS and Rotated RetinaNet sigmoid focal loss per class (same loss.class_weight_* as ROI heads). focal stays unweighted. background_weight is ignored on one-stage one-hot focal.
  • make wizard / --wizard FCOS FPN and box-reg nudges are class-agnostic: pooled GT width vs finest stride (keep P3 + decoded kfiou when boxes occupy few cells), not DOTA class names such as small-vehicle.
  • Hub slug rotated_fcos_dota_le90_3x refreshed from runs/rotated_fcos/20260831-052647 (82.32% eval-val mAP50, was 81.58%). Weight stem is now rotated_fcos_r50_fpn_dota_le90_3x-6e383331 (dropped leftover _riou in the filename). Report docs/eval-reports/rotated_fcos_dota_le90_3x/.
  • NMS split (DOTA + HRSC) — model.final_nms_iou_threshold: 0.1 (train val), production.final_nms_iou_threshold: 0.3 (deploy / image_demo), and new evaluation.final_nms_iou_threshold: 0.1 for odet preds / make eval-val (MMRotate test parity). Resolver: resolve_preds_final_nms_iou_threshold.

  • Rotated FCOS DOTA dota_le90_3x.json is now the decoded rIoU 3× recipe (was dota_le90_3x_riou.json). The previous L1 3× is dota_le90_3x_l1.json. Hub slug rotated_fcos_dota_le90_3x (was rotated_fcos_dota_le90_3x_riou).

  • ROI box-reg aux keys — roi_box_reg_iou_weight / roi_box_reg_iou_loss_type / roi_box_reg_smooth_l1_aux_weight are now roi_box_reg_aux_weight + roi_box_reg_aux_loss_type (smooth_l1 | probiou | riou | kfiou). Schedule fields are roi_box_reg_aux_schedule_*. Weights are unchanged (no retrain). Old keys still load with a deprecation warning. Hub slugs are unchanged.

  • training.lr_scheduler_cosine_t_max remaps to lr_scheduler_cosine_epochs (same integer; deprecation warning). Mixing both with different values is an error.

  • HRSC2016 eval (resize_mode: pad / keep_ratio) uses the same whole-image scale forward as training. DOTA fixed/crop eval-val still native-tiles oversized rasters.

  • HRSC two-stage max_detections_per_image 2000 (was 100). Final NMS for published eval-val stays 0.1 via evaluation.final_nms_iou_threshold; recipes now ship production NMS 0.3 like DOTA.

[0.2.0] - 2026-08-25

Added

  • Rotated FCOS (model_type: rotated_fcos) — anchor-free single-stage detector: DistanceAnglePointCoder, center-in-OBB assigner, centerness, and L1 / KFIoU / decoded rIoU box regression. Recipes under configs/rotated_fcos/.
  • Differentiable polygon IoU (oriented_det.ops.diff_iou_rotated) for FCOS box_reg_loss_type: riou (1 - IoU). Distinct from sampling pairwise_rotated_iou. Recipes dota_le90_1x_riou.json / dota_le90_3x.json (lr 2.5e-3).
  • Hub slug rotated_fcos_dota_le90_3x (81.58% eval-val mAP50) — Rotated FCOS 3× decoded rIoU; report docs/eval-reports/rotated_fcos_dota_le90_3x/.
  • Hub slug rotated_fcos_dota_le90_3x_kfiou_aux (77.18% eval-val mAP50) — Rotated FCOS 3× L1 + KFIoU aux; report docs/eval-reports/rotated_fcos_dota_le90_3x_kfiou_aux/.
  • TF/ONNX export mode rotated_fcos_pre_nms — Rotated FCOS decode + pad uses the same Keras detect bundle as two-stage models.

Removed

  • FCOS 1× ProbIoU-aux recipe (dota_le90_1x_probiou_aux.json). Train-time mAP50 66.8% vs 76.5% for 1× KFIoU aux on the same protocol.

[0.1.1] - 2026-07-11

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.

[0.1.0] - 2026-05-27

Added

  • Core geometry (Polygon, QBox, RBox) and transforms
  • Rotated IoU, NMS, and optional GPU kernels
  • DOTA loader, tiling, augmentations, oriented mAP
  • Airbus Playground CSV dataset support
  • Oriented R-CNN, Rotated Faster R-CNN, Rotated RetinaNet
  • JSON config training via odet train
  • Pretrained weights on Hugging Face Hub (dl4eo/oriented-det-pretrained), including Oriented R-CNN 1× DOTA le90 (74.79% eval-val mAP50) and Oriented R-CNN 3× DOTA le90 (79.40% eval-val mAP50)
  • MkDocs user guide and API reference

Notes

  • Public home: https://github.com/DL4EO/oriented-det