HRSC2016 in oriented-det — recipes, trains, and a held-out 90.41%

DOTA is the pretrain zoo. HRSC2016 is the small-data ship zoo.

Three oriented-det families now have native HRSC recipes and published 3× Hub weights: Oriented R-CNN 90.41%, Rotated Faster R-CNN 88.77%, Rotated FCOS 88.34% mAP50. Those numbers are make eval-val on ImageSets test. Test is not in training. That sentence is the whole point of this dataset in our stack — DOTA eval-val is leaky; HRSC eval-val is a real holdout.

This post is the dataset, the recipes, the L4 trains, and the reports. It is a research benchmark. It is not a production ship detector — see the licensing note.

In-train val mAP50 every 4 epochs on the three HRSC 3× runs (non-empty ImageSets test)

The dataset Link to heading

Liu, Yuan, Weng, Yang, ICPRAM 2017. High-resolution optical ships from Google Earth, oriented boxes, one class we actually train: ship. Fine-grained Class_ID values in the XML are ignored.

The original escience.cn page is often offline. Use a copy of the official 2016 release, not HRSC2016-MS:

Source	Notes
IEEE DataPort	~3.5 GB `HRSC2016_dataset.zip`; free IEEE account
Baidu AI Studio	Same layout MMRotate points at
Kaggle `guofeng/hrsc2016`	`kaggle datasets download -d guofeng/hrsc2016`

Point dataset.data_root at the folder that contains FullDataSet/ and ImageSets/ (a wrapping HRSC2016/ directory is accepted):

HRSC2016/
  FullDataSet/AllImages/*.bmp
  FullDataSet/Annotations/*.xml
  ImageSets/{train,val,test,trainval}.txt

The paper quotes 1,061 images as 436 train / 181 val / 444 test. The ImageSets we actually train against are trainval 617 (436 + 181) and test 453 (15 of those empty). filter_empty_gt drops the empties from the training-loop val loader (438 kept); make eval-val scores all 453.

Split	Images	Ships	In training?
ImageSets trainval	617	1,748	Yes
ImageSets test	453 (15 empty)	1,228 (1,188 in VOC AP)	No

XML mbox_cx/cy/w/h/ang is a rotated box with angle in radians. The loader (dataset.format: hrsc2016) converts through the same polygon → RBox path as DOTA, so training is le90.

Unlike DOTA, these images fit a whole-image canvas. No odet tile-dota for the Hub recipes. Optional export exists (odet hrsc-to-dota) if you want DOTA folders anyway.

I am not republishing the Google Earth pixels here. Run odet image-demo on a test BMP if you want to see boxes.

Recipes Link to heading

Six configs, two schedules, three families. Hub publishes 3× only.

Family	1×	3× (Hub)
Oriented R-CNN	`hrsc2016_le90_1x.json`	`hrsc2016_le90_3x.json`
Rotated Faster R-CNN	`hrsc2016_le90_1x.json`	`hrsc2016_le90_3x.json`
Rotated FCOS	`hrsc2016_le90_1x.json`	`hrsc2016_le90_3x.json`

Shared canvas: keep_ratio, long edge 800, pad_size_divisor 32. Horizontal + vertical + diagonal flips. Batch 2, ResNet-50 FPN, ImageNet backbone, frozen_stages: 1. Train from scratch (load_from_checkpoint: null) — these are not DOTA finetunes.

What changes per family:

Knob	Oriented R-CNN	Faster R-CNN	FCOS
Box loss	Smooth L1 main + ProbIoU aux 0.1	ProbIoU main + Smooth L1 aux 0.1	decoded rIoU
LR	0.005	0.005	0.0025
3× `lr_scheduler_gamma`	0.1	0.1	`[0.1, 0.5]`
Random rotate 1×	off	off	on p=0.5 ±20°
Random rotate 3×	on p=0.5 ±20°	on p=0.5 ±20°	on (inherited)
Deploy score	0.85	0.85	0.20

3× is 36 epochs, milestones [24, 33]. FCOS keeps a milder second drop (0.5 instead of 0.1) so the decoded-rIoU head does not get wrecked at epoch 33.

A 6× schedule was trained and removed. It was +0.2 mAP on Oriented R-CNN; FCOS 6× never beat this 3×. ±20° is the long schedule. ±180° on FCOS 6× diverged after epoch 12.

NMS is 0.1 everywhere that matters: train val, make eval-val, deploy. Two-stage HRSC keeps max_detections_per_image 2000. That is MMRotate HRSC test-cfg, not the DOTA deploy floor.

Trains Link to heading

Three published runs, single NVIDIA L4, PyTorch 2.3, AMP off, batch 2, 309 steps/epoch on 617 trainval images.

Model	Run	Wall	Mean epoch	Checkpoint
Oriented R-CNN 3×	`runs/oriented_rcnn/20260830-163857`	2h 34m	4m 17s	`best_mAP_0.90.pth`
Faster R-CNN 3×	`runs/rotated_faster_rcnn/20260831-035851`	1h 5m	1m 50s	`best_mAP_0.89.pth`
FCOS 3×	`runs/rotated_fcos/20260831-020019`	34m	57s	`best_mAP_0.89.pth`

FCOS is ~4.5× faster to train than Oriented R-CNN on this set. Same ratio we saw on DOTA 1× vs FCOS on L4. Oriented RoIAlign over ~2000 proposals is still the tax.

Periodic mAP every 4 epochs (the plot above) is in-train val: 438 non-empty test images, score ≥ 0.05. It is not the Hub number — but it is how the runs actually moved:

Oriented R-CNN is already at 62% by epoch 4 and 89% by epoch 12. The rest of 3× is a slow polish to 90.4%.
Faster R-CNN crawls until the first LR drop, then jumps (75% at epoch 24 → 88% at epoch 28).
FCOS has a real dip at epoch 20 (37%, down from 74% at epoch 16), then recovers after the milestone and sits with Faster R-CNN in the high 88s.

Best checkpoints are not the last epoch. Oriented R-CNN peaked at epoch 32 (90.41% in-train); epoch 36 was 90.39%. Faster R-CNN peaked at 88.92% then 88.31%. FCOS 88.62% then 88.22%. The Hub weights are those best_mAP_*.pth files, re-scored with make eval-val.

# edit dataset.data_root in the config, then:
odet train --config configs/oriented_rcnn/hrsc2016_le90_3x.json
odet train --config configs/rotated_faster_rcnn/hrsc2016_le90_3x.json
odet train --config configs/rotated_fcos/hrsc2016_le90_3x.json

Results Link to heading

Published metric: make eval-val mAP50 on ImageSets test, rotated IoU ≥ 0.50, NMS 0.1, score ≥ 0.05. 453 images. Held-out.

Model	Hub slug	mAP50	F1 peak	Deploy `--score-thr`	Mean best IoU
Oriented R-CNN 3×	`oriented_rcnn_hrsc2016_le90_3x`	90.41%	0.940 @ 0.90	0.85	0.848
Rotated Faster R-CNN 3×	`rotated_faster_rcnn_hrsc2016_le90_3x`	88.77%	0.916 @ 0.90	0.85	0.785
Rotated FCOS 3×	`rotated_fcos_hrsc2016_le90_3x`	88.34%	0.904 @ 0.25	0.20	0.746

Deploy floors are the same rule as DOTA: eval-val global F1 minus 0.05. Two-stage HRSC scores pile up near 1.00, so F1 peaks at 0.90. FCOS F1 peaks at 0.25. Copying --score-thr 0.85 onto FCOS will hide the scene. That is the same decoder ring as the macOS FCOS note.

F1 versus score threshold on held-out HRSC test — two-stage heads peak at 0.90, FCOS at 0.25

At each model’s F1 operating point:

	TP	FP	FN	Precision	Recall
Oriented R-CNN @ 0.90	1142	59	86	0.951	0.930
Faster R-CNN @ 0.90	1108	82	120	0.931	0.902
FCOS @ 0.25	1108	115	120	0.906	0.902

Faster R-CNN and FCOS miss the same 120 ships. FCOS pays for its one-stage head with extra false positives. Oriented R-CNN is ahead on both counts, and the boxes are tighter (mean best IoU 0.85 vs 0.79 vs 0.75). That is the accuracy pick on this set.

Precision–recall on held-out HRSC test (points with at least one true positive)

The PR plot is the calibration story in another shape. Two-stage curves live in a high-recall sliver because the scores are peaked. FCOS traces a long high-precision arc down to low recall — usable, just not at 0.85.

Eval reports (per-class AP is one row: ship): oriented_rcnn_hrsc2016_le90_3x, rotated_faster_rcnn_hrsc2016_le90_3x, rotated_fcos_hrsc2016_le90_3x.

What to download Link to heading

odet pretrained download oriented_rcnn_hrsc2016_le90_3x
odet pretrained download rotated_faster_rcnn_hrsc2016_le90_3x
odet pretrained download rotated_fcos_hrsc2016_le90_3x

odet image-demo path/to/test.bmp hf://oriented_rcnn_hrsc2016_le90_3x \
  --score-thr 0.85 --nms-thr 0.1 --out-file out_orcnn.png

Swap the slug for Faster R-CNN (0.85) or FCOS (0.20). Keep --nms-thr 0.1. HRSC recipes use resize_mode: keep_ratio / pad, so inference is one whole-image forward — no sliding windows.

make eval-val still uses score ≥ 0.05. Do not mix that protocol with the deploy floor.

License, again Link to heading

HRSC2016 is a research ship set collected from Google Earth. The Hub weights are research checkpoints. Apache 2.0 covers oriented-det’s code. It does not cover these images, and it does not turn a 90% mAP weight file into a commercial detector.

If you need ships in production, train on imagery you have licensed for that use. The recipes above are the starting point; the pixels have to be yours.