FOD-A: A Dataset for Foreign Object Debris in Airports
Paper • 2110.03072 • Published
How to use DarkKnight1217/RunwayGuard-rtdetr-l with ultralytics:
# Couldn't find a valid YOLO version tag.
# Replace XX with the correct version.
from ultralytics import YOLOvXX
model = YOLOvXX.from_pretrained("DarkKnight1217/RunwayGuard-rtdetr-l")
source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
model.predict(source=source, save=True)Fine-tuned RT-DETR-L (Ultralytics 8.4.147) for visible airport foreign-object debris (FOD) on FOD-A v2.1, using a deterministic grouped train/val/test split (seed 42).
| File | Role |
|---|---|
best.pt |
Selected checkpoint (best val fitness during training) |
results.csv |
Per-epoch train/val curves |
test_metrics.json |
Held-out grouped-test metrics |
SHA256(best.pt) = C2D2A418069D9AC8658CA85B8359A9E1AB740CE96826C5138178B080F9243269
| Metric | Value |
|---|---|
| mAP50 | 0.755 |
| mAP50–95 | 0.636 |
| Library P / R | 0.734 / 0.812 |
Seven operational classes: fastener_hardware, hand_tool, flexible_debris, loose_metal, plastic_paper_debris, component_container, natural_debris.
# download
huggingface-cli download USER/RunwayGuard-rtdetr-l best.pt --local-dir ./weights
# or
curl -L -o weights/best.pt https://huggingface.co/USER/RunwayGuard-rtdetr-l/resolve/main/best.pt
Point the FastAPI app at the local file:
export MODEL_PATH=weights/best.pt
uvicorn app.main:app --host 0.0.0.0 --port 8000
Detects visible candidate debris in an image. Does not certify runway clearance, material, mass, or future aircraft damage. See the GitHub repo memo for split rationale and failure cases.
Code: https://github.com/riyaz-ahamed-07/RunwayGuard
Data: FOD-A (MIT) — cite arXiv:2110.03072