ONNX export¶

Export a trained OrientedDet checkpoint to a pre-NMS ONNX graph plus Python preprocess / rotated NMS. Consumers run ONNX Runtime only — no PyTorch and no oriented-det.

The producer CLI is odet export (ONNX only; same as python -m export). Folder README: export/README.md. Sliding-window tiling and the Tile Geo Process HTTP service stay on the Docker deploy path in this release.

Install¶

Producer (checkpoint → ONNX):

uv pip install -e ".[export]"

Consumer (infer / demo without this repo):

pip install -r export/requirements-runtime.txt

GPU ORT: install onnxruntime-gpu[cuda,cudnn] instead of onnxruntime (do not install both).

Export a graph¶

Default is Rotated FCOS (rotated_fcos_pre_nms). Download Hub weights or point at a training run:

odet pretrained download rotated_fcos_dota_le90_1x
make export-onnx

From a run directory:

make export-onnx EXPERIMENT=runs/rotated_fcos/<timestamp>
make export-onnx EXPERIMENT=runs/oriented_rcnn/<timestamp> EXPORT_MODE=oriented_rcnn_pre_nms
make export-onnx EXPERIMENT=runs/rotated_faster_rcnn/<timestamp> EXPORT_MODE=faster_rcnn_pre_nms
make export-onnx EXPERIMENT=runs/rotated_rtmdet/<timestamp> EXPORT_MODE=rotated_rtmdet_pre_nms

Equivalent CLI (placeholder paths):

odet export onnx \
  --config /path/to/oriented-det/runs/rotated_fcos/<timestamp>/config.json \
  --checkpoint /path/to/oriented-det/runs/rotated_fcos/<timestamp>/checkpoints/checkpoint_best.pth \
  --output /path/to/oriented-det/onnx_export/model.onnx

Artifacts (gitignored except the folder README) land in onnx_export/:

File	Role
`model.onnx`	Backbone + head + box decode; padded pre-NMS tensors
`model.export_meta.json`	Mean/std, class names, score floor, NMS IoU
`preprocess.py`, `nms.py`, `postprocess.py`, `runtime.py`, `ort_runtime.py`	Copies of the consumer stack
`demo.py`, `infer_onnx.py`	Copies of the infer scripts

ONNX input is NCHW RGB, already mean/std-normalized, typically [1, 3, 1024, 1024]. Do not feed raw JPEGs into the graph.

Infer¶

odet export infer --onnx ./onnx_export/model.onnx --smoke
odet export infer --onnx ./onnx_export/model.onnx --images /path/to/data/tiles --output ./onnx_export/predictions
make export-demo

Python:

from PIL import Image
from export.runtime import detect_image

dets = detect_image(Image.open("tile.jpg"), "onnx_export/model.onnx")
# each det: cx, cy, w, h, angle_rad, score, label, class_name  (original image pixels)

From a copied onnx_export/ folder only:

cd onnx_export
pip install -r requirements-runtime.txt
python demo.py

from runtime import detect_image
dets = detect_image("tile.jpg", "model.onnx")

Val-split predictions in the same JSON shape as odet preds (then make export-metrics / odet preds --metrics-from-json):

odet export preds \
  --config /path/to/oriented-det/runs/rotated_fcos/<timestamp>/config.json \
  --onnx /path/to/oriented-det/onnx_export/model.onnx

Sliding-window tiling is out of scope for the ONNX path — use pre-tiled images (same as PyTorch odet preds on DOTA tiles). Preprocess is bilinear resize to a fixed canvas (default 1024×1024). That matches DOTA tiles, not keep_ratio + pad (HRSC / SSDD / HRSID).

Modes¶

`--mode` / `EXPORT_MODE`	Detector
`rotated_fcos_pre_nms` (default)	Rotated FCOS
`rotated_rtmdet_pre_nms`	Rotated RTMDet
`oriented_rcnn_pre_nms`	Oriented R-CNN
`faster_rcnn_pre_nms`	Rotated Faster R-CNN

Decode runs inside ONNX. Final rotated NMS is numpy (export.nms; optional --nms-backend shapely). See export/PARITY.md.