HRSID in oriented-det — the larger SAR ship benchmark at 78.55%

HRSID is the large radar ship set: 5,604 chips, harbours as well as open water.

A Faster R-CNN finetune reached 78.55% on the official held-out test. The curves are below.

HRSID test — dense harbour chip with COCO-polygon ground truth (research illustration)

The dataset Link to heading

Official 65/35 train/test. No val — recipes evaluate on test. MS COCO polygons → le90 rbox (n-gons use min-area rectangle). Point data_root at HRSID_JPG (JPEGImages/ + annotations/*2017.json). Whole-image keep-ratio 800 + pad-32. No tiling (chips are already 800×800). Native training does not need odet coco-to-dota.

odet train --config configs/rotated_faster_rcnn/hrsid_le90_1x.json
make eval-val EXPERIMENT=runs/rotated_faster_rcnn/<timestamp>

Oriented R-CNN / FCOS 1× recipes exist; those trains are not yet run. No HRSID notebook — this post is the tutorial.

	SSDD	HRSID
Chips	1,160	5,604
Canvas	keep-ratio 608	keep-ratio 800
Split	last-digit train/test	official 65/35 (no val)
Local FRCNN 1×	90.34%	78.55%
Hub	none	none

HRSID test — open-water chip with GT (fewer ships, clearer heading)

Local 1× Faster R-CNN: 78.55% Link to heading

From hf://rotated_faster_rcnn_dota_le90_1x (1-way ship head re-init), RTX 3090 Ti, 1h 32m (runs/rotated_faster_rcnn/20260918-134544). Held-out test: 1,962 chips, 5,918 ships. Score ≥ 0.05, rotated IoU 0.50, NMS IoU 0.10. Report: docs/eval-reports/rotated_faster_rcnn_hrsid_le90_1x/.

Epoch	Train loss	In-train val mAP50 (score ≥ 0.3)
4	0.290	70.71%
8	0.270	69.94%
12	0.245	71.53%

make eval-val (score ≥ 0.05) is 78.55% — same checkpoint. The 0.3 in-train floor drops many true positives (recall 0.83 at 0.05 vs GT cover 79% at 0.3). That is a threshold. Best-F1 deploy threshold on that sweep is 0.60 (P 0.90 / R 0.74 / F1 0.81). Mean best IoU vs GT is 0.68.

Wei et al. (IEEE Access 2020, Table 4) report horizontal box AP above 84.7% on this split. The 78.55% here is rotated overlap on the ship’s own rectangle.

Precision–recall on HRSID held-out test (Rotated Faster R-CNN 1×)