RunwayGuard RT-DETR-L (FOD-A grouped split)

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).

Files

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

Checksum

SHA256(best.pt) = C2D2A418069D9AC8658CA85B8359A9E1AB740CE96826C5138178B080F9243269

Grouped-test (frozen)

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.

Usage

# 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

Scope / limits

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

Downloads last month
33
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Space using DarkKnight1217/RunwayGuard-rtdetr-l 1

Paper for DarkKnight1217/RunwayGuard-rtdetr-l