YOLO26n_RDD_Base

YOLO26n road-damage detector (4-class CRDDC: D00 longitudinal, D10 transverse, D20 alligator, D40 pothole). Trained on the Unified Road Defect Dataset (GT labels only (baseline).)

Metrics (best epoch 100, our val split)

mAP50	mAP50-95	precision	recall	F1
0.6346	0.3339	0.6988	0.5588	0.621

Usage

from ultralytics import YOLO
model = YOLO("YOLO26n_RDD_Base.pt")
results = model("road.jpg")

Training data & distillation

• Dataset: TamAko783/Unified_Road_Defect_Dataset (RDD-2022 + UAV-PDD2023 + RoadDamageVision, 30,186 images, 4 classes).
• Distillation teacher (for the _FRDC_Distilled variant): Co-DETR (Swin-L) from WangFangjun/FRDC-RDD (FRDC, ORDDC'2024 winner) — used offline to pseudo-label the training set.

Credits

RDD-2022 (Arya et al.), UAV-PDD2023, RoadDamageVision (Silva Zendron & Leithardt, CC BY 4.0), FRDC Co-DETR teacher (Apache-2.0). Student model under AGPL-3.0 (Ultralytics).