DOTA recipes train on a 1024×1024 canvas. Real scenes are often larger. When the input exceeds that canvas, odet image-demo switches to pad/tile automatically: overlapping 1024×1024 windows, detections mapped back to full-image coordinates, then merge NMS.

The Hub checkpoint in the commands below is oriented_rcnn_dota_le90_1x. It was retrained after a diagonal-flip bug in the data pipeline — the boxes and headings on this walkthrough come from that corrected weight.

pip install oriented-det
odet pretrained download oriented_rcnn_dota_le90_1x

In a training or inference config:

"load_from_checkpoint": "hf://oriented_rcnn_dota_le90_1x"

On a single tile (bundled recipe + sidecar config from pretrained/):

odet image-demo demo.jpg hf://oriented_rcnn_dota_le90_1x \
  --out-file result.jpg --device cuda \
  --score-thr 0.55 --nms-thr 0.1

Weights, config sidecar, and training log: pretrained/oriented_rcnn_r50_fpn_dota_le90_1x-725c244f.* · Recipe: configs/oriented_rcnn/dota_le90_1x.json

Large images: sliding-window inference Link to heading

Example — ship detection on demo/large.jpg (1299×1904):

odet image-demo demo/large.jpg hf://oriented_rcnn_dota_le90_1x \
  --out-file result.jpg \
  --score-thr 0.55 --nms-thr 0.1 \
  --window-batch-size 8 --classes ship

Typical CLI output:

Preprocessing: resize_mode=fixed, target_size=(1024, 1024) (model canvas 1024×1024)
Inference thresholds: score>=0.55, merge NMS IoU<=0.1, overlap_pixels=200, ignore_margin_pixels=0.0
  -> pad/tile (image 1299×1904 vs canvas 1024×1024, overlap_pixels=200, 6 windows)
  -> detections (score >= 0.55, NMS <= 0.1)
  -> after class filter ['ship']
Saved visualization to result.jpg

Ship detections on demo/large.jpg — Oriented R-CNN 1×

Ship detections on demo/large.jpg — Oriented R-CNN 1×

Each docked vessel gets a rotated box aligned to its hull, with no visible seams at the six window boundaries.

What to notice:

  • 6 windows for this image size — modest overhead compared with a single tile.
  • --window-batch-size 8 batches window inference on GPU (all six windows in one pass here).
  • --classes ship keeps one DOTA class after detection.
  • overlap_pixels=200 comes from the bundled recipe default — fine for DOTA-scale objects; increase it if your targets are larger than the overlap band, or they can be split across window boundaries.
  • --score-thr 0.55 is the Hub deploy floor for this 1× slug (production.score_threshold). The bundled demo.jpg bus tile uses the same default.

For a zero-shot maritime experiment on a Copernicus Sentinel-2 tile — zoom, overlap, and margin tuned for small ships — see Zero-shot ship detection on a Copernicus Sentinel-2 tile with Oriented R-CNN.

Demo thresholds (short note) Link to heading

--score-thr and --nms-thr on odet image-demo are post-decode filters. Values tuned on one architecture do not transfer to the others. Hub 1× deploy floors: Oriented R-CNN 0.55, Rotated Faster R-CNN 0.60, FCOS 0.20, RetinaNet 0.35. Copying 0.70 onto FCOS will hide most of the scene.

0.5 is the mAP matching IoU on DOTA Task 1, not detection NMS. Recipes use production.final_nms_iou_threshold: 0.1.

June 29, 2026 — Jeff Faudi