Oriented-Det v0.4.0 adds a detector I actually want to train: Rotated RTMDet.

One stage, a small CSPNeXt backbone, and 77.60% on official DOTA Task 1. The boxes are tight: AP75 52.01. The slug is rotated_rtmdet_dota_le90_3x. Deploy it at 0.45.

How to install it, and how fast it trains, are below.

Upgrade Link to heading

pip install -U oriented-det
# or pin:
pip install oriented-det==0.4.0

odet pretrained download rotated_rtmdet_dota_le90_3x
odet image-demo demo.jpg hf://rotated_rtmdet_dota_le90_3x --score-thr 0.45

PyTorch is still installed separately (pytorch.org). Weights stay on Hugging Face at dl4eo/oriented-det-pretrained. Deploy score is 0.45 (best F1 on the leaky sweep was 0.50; the zoo rule subtracts 0.05). NMS stays 0.1.

Quote Task 1 Link to heading

Recipes train on train+val tiles, so make eval-val scores tiles the model already saw. For this run that leaky figure is 82.75%. The gap to the hidden test is 5.15 points. Helicopter is the sharp version of the same trap: 61.6% on Task 1, 90.0% on leaky eval-val. Quote 77.60%.

ModelScheduleTask 1 AP50Task 1 AP75Deploy score
Rotated RTMDet-S3×77.60%52.010.45
Oriented R-CNN1×76.73%50.240.55
Oriented R-CNN3×74.88%51.230.55
Rotated Faster R-CNN1×74.42%41.900.60
Rotated Faster R-CNN3×74.48%45.390.60
Rotated RetinaNet (OBB)3×73.89%47.110.25
Rotated FCOS1×73.07%40.400.20
Rotated FCOS3×72.91%45.390.20
Rotated RetinaNet (OBB)1×71.72%43.460.25

RTMDet leads AP50 and AP75. The previous AP75 peak was Oriented R-CNN 3× at 51.23. The ResNet families still finetune from 1×. RTMDet has one Hub weight, and it is this 3×. That is the checkpoint a finetune loads. A CSPNeXt-tiny fragment is in the tree. It has no Hub weights.

On the hidden test, the strong classes are tennis-court 90.7, plane 89.2, ship 88.9, storage-tank 86.6, large-vehicle 82.9. Bridge is 53.5. Helicopter is 61.6. Report: docs/eval-reports/rotated_rtmdet_dota_le90_3x/. Run: runs/rotated_rtmdet/20261002-102116.

Same size, MMRotate’s table Link to heading

OpenMMLab’s RTMDet-s with ImageNet pretrain and random rotate, without multi-scale, is 76.93% mAP50 (AP75 50.59, mmAP 48.16). Same size class, same augmentation class, same server metric. Our 77.60 / 52.01 / 49.02 sits in that band. Their multi-scale + random-rotate line is 79.98% mAP50 (AP75 60.07). We did not train that recipe. Mosaic is off.

The log counts 8,857,000 parameters. Their table lists 8.86M and 37.62 GFLOPs for RTMDet-s. The ResNet-50 heads in this zoo are 36–41M (RetinaNet 36.6M on its 3× log, FCOS 36.2M, Oriented R-CNN 41.3M).

Speed Link to heading

Two measurements. They do not divide into each other.

Training wall, one RTX 3090 Ti Link to heading

Both runs below are the published 3× recipes: 13,691 train+val tiles, AMP off, mAP every 4 epochs, one RTX 3090 Ti. Tiles/s is 13691 / mean epoch. It includes those mAP epochs.

ModelEpochsBatchWallMean epochTiles/s
Rotated RTMDet-S3688h 58m14m 57s15.3
Rotated RetinaNet (OBB)36223h 23m38m 59s5.9

RTMDet moves about 2.6× more tiles per second of wall clock, at 4× the batch, with about a quarter of the parameters. Quiet epochs (the log minimum, a pass without the polygon match) are 11m 14s for RTMDet and 34m 22s for RetinaNet. The 8h 58m figure is the whole recipe, match included. Source: pretrained/rotated_rtmdet_cspnext_s_dota_le90_3x-3f6a1a0b.log and pretrained/rotated_retinanet_r50_fpn_dota_le90_3x-961aaf73.log.

The ResNet 1× walls already in this series are a different card (one NVIDIA L4, batch 2, 12 epochs): FCOS 7h 59m, Faster R-CNN 11h 30m, Oriented R-CNN 1d 12h 22m. Leave them on that card. A 3090 Ti number and an L4 number are not a speedup.

Inference Link to heading

The tiled-inference stitch already published is PyTorch, one 1024 forward per DOTA val tile, CPU polygon NMS, 7,669 tiles:

ModelTask 1 AP50Throughputms / tile
Rotated Faster R-CNN 1×74.42%6.25 img/s160
Oriented R-CNN 1×76.73%0.91 img/s1,100

That is the July throughput note. RTMDet is not in it. tools/bench_latency.py times a single-tile forward (backbone + head + decode + NMS) for several detectors on one canvas. This post does not invent a third row from a different protocol.

MMRotate’s published RTMDet-s latency is a third protocol again: 4.86 ms, TensorRT FP16, batch 1, with NMS, on an RTX 2080 Ti (their README). That is their checkpoint, not our PyTorch forward, and it is not a ratio against the 160 ms row.

What the head is Link to heading

model_type: rotated_rtmdet. Clean-room CSPNeXt-S and CSPNeXt-PAFPN, strides 8 / 16 / 32. The head is SepBN: shared convolutions, a batch-norm per FPN level, no centerness branch. Assignment is a dynamic soft-label assigner (top-k 13). Classification is Quality Focal Loss. The box loss is decoded rotated IoU at weight 2.0. The angle L1 term is off (rtmdet_angle_weight: 0). Boxes are le90.

The DOTA recipe always sees 1024×1024. Thirty-six epochs, batch 8, AdamW 2.5e-4, weight decay 0.05 on conv weights only, no gradient clip. Warmup is 1000 steps, the learning rate stays flat until epoch 18, then cosine to 1.25e-5. EMA (momentum 0.0002) is what validation and best_mAP_*.pth store. Random rotate is on (p=0.5, ±180°). A tile that contains storage-tank or roundabout rotates by 90° steps only. The backbone starts from ImageNet.

odet train --config configs/rotated_rtmdet/dota_le90_3x_s.json

Export Link to heading

Pre-NMS ONNX covers this head: CSPNeXt, the SepBN head, and box decode in the graph. Rotated NMS stays in Python. Fixed H×W, batch 1, same contract as FCOS.

make export-onnx EXPERIMENT=runs/rotated_rtmdet/<timestamp> EXPORT_MODE=rotated_rtmdet_pre_nms

The v0.3 ONNX post is FCOS, Oriented R-CNN, and Faster R-CNN. RTMDet export ships in this tag.

Also in 0.4.0 Link to heading

Native YOLO-OBB is the following tag, v0.4.1. This tag is RTMDet.

  • Periodic val and make eval-val use val_score_threshold. test_map and make eval-test use test_score_threshold (null → 0.05). Older key names still load.
  • odet preds / make metrics scores every ground truth and every detection (margin 0). Deploy / image_demo still use production.ignore_margin_pixels. On this slug that margin is 0.
  • tools/bench_latency.py for a same-canvas forward.

Where this weight sits next to the other heads is which detector to train.

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