v0.4.0

Rotated RTMDet. rotated_rtmdet_dota_le90_3x (CSPNeXt-S + PAFPN) is on the Hub at official Task 1 77.60%, the top of the DOTA zoo. Native YOLO-OBB is 0.4.1.

Added

• Rotated RTMDet (model_type: rotated_rtmdet) — clean-room CSPNeXt backbone + CSPNeXt-PAFPN neck, SepBN head (shared convs, per-level BN), dynamic soft-label assigner, Quality Focal Loss + decoded rotated IoU box loss. Base fragments: rotated_rtmdet_cspnext_s.json and rotated_rtmdet_cspnext_tiny.json (tiny has no Hub weights).
• RTMDet-S 3× Hub — rotated_rtmdet_dota_le90_3x. Official Task 1 77.60% (AP75 52.01, COCO mAP 49.02; deploy 0.45; leaky eval-val 82.75%). Recipe: dota_le90_3x_s.json (36 epochs, AdamW, EMA, random rotate p=0.5 ±180°, mosaic off). Report: docs/eval-reports/rotated_rtmdet_dota_le90_3x/.
• EMA — training.use_ema / ema_momentum / ema_gamma. Validation and best_mAP_*.pth use the EMA weights; epoch checkpoints keep raw weights plus ema_state_dict for resume.
• AdamW and weight-decay groups — training.optimizer: adamw (adamw_betas), weight_decay_norm_mult / weight_decay_bias_mult. max_grad_norm: null disables clipping. Defaults (SGD, multipliers 1, clip 1.0) are unchanged.
• lr_scheduler_type: cosine_annealing_late — flat LR after warmup until lr_scheduler_cosine_start_epoch, then cosine to lr_scheduler_cosine_eta_min.
• preprocessing.random_rotate_rect_labels — tiles containing those classes rotate by 90° steps only (RTMDet uses storage-tank and roundabout).
• Cached mosaic (preprocessing.enable_mosaic and mosaic_* keys), off by default and off in every published recipe.
• tools/bench_latency.py — single-tile forward latency (backbone + head + decode + NMS) for several detectors on the same canvas and device.
• OSSDD FCOS and Faster R-CNN 3× — ossdd_le90_vh_3x.json and ossdd_le90_vh_3x.json inherit VH 1× finetunes of the DOTA 1× hubs (36 epochs, MultiStep milestones 24/33, 512 crop, batch 4).
• OSSDD mAP — evaluation.map_iou_thresholds averages the area under the precision-recall curve over rotated IoU 0.10 to 0.95 step 0.05. Set on the OSSDD recipes only. best_metric mAP uses that average. iou_threshold stays 0.5 for GT-cover and the precision-recall operating point.
• RTMDet ONNX — odet export onnx --mode rotated_rtmdet_pre_nms. CSPNeXt + SepBN head + box decode in the graph; rotated NMS stays in Python. Fixed H×W, batch 1, same pre-NMS tensors as FCOS.

Changed

• Evaluation split into val_* and test_*. Periodic val and make eval-val use val_score_threshold (often 0.3). test_map and make eval-test use test_score_threshold (null → 0.05). The 1× schedule sets test_map: true and val_map_at_end: false. DOTA and FAIR1M point test_tiles_dir at the val tiles (leaky). dota-submit still requires the unlabeled official test and does not read test_tiles_dir. Older keys (train_val_score_threshold, preds_score_threshold, compute_map_final, compute_map_test, …) remap on load. Published Hub sidecars in pretrained/ use the new names.
• odet preds / make metrics metrics margin defaults to 0 (every GT and detection is scored). It no longer reads production.ignore_margin_pixels, tile overlap, or a margin stored in predictions.json. Pass --metrics-margin-pixels for a positive band. Deploy / image_demo still use production.ignore_margin_pixels.
• Tutorial notebooks write to work_dirs/ in a checkout (was tutorial_work/). ORIENTED_DET_TUTORIAL_WORK and the Kaggle /kaggle/working default are unchanged.

ONNX export covers Rotated FCOS, Rotated RTMDet, Oriented R-CNN, and Rotated Faster R-CNN.

Full changelog: https://github.com/DL4EO/oriented-det/blob/v0.4.0/docs/changelog.md