YOLO26n_RDD_FRDC_Distilled

A YOLO26n road-damage detector (4-class CRDDC: D00 longitudinal, D10 transverse, D20 alligator, D40 pothole) trained with semi-supervised knowledge distillation from a Co-DETR teacher. It is the distilled student and outperforms the ground-truth-only baseline while remaining a fast, NMS-free nano model.

How it was trained (FRDC semi-supervised recipe)

1. Teacher — Co-DETR (Swin-L), the ORDDC'2024-winning model from WangFangjun/FRDC-RDD.
2. Pseudo-labeling — the teacher labeled the 9,035 unlabeled RDD-2022 test images (confidence ≥ 0.5); 4,719 images with confident detections were kept.
3. Student training — YOLO26n trained on ground-truth train (30,186 imgs) + 4,719 pseudo-labeled images (Unified_Road_Defect_FRDC_Pseudolabeled), 100 epochs, imgsz 640.

The original ground-truth labels were left untouched — the improvement comes from adding new, teacher-labeled images, not from relabeling existing data.

Metrics (held-out validation, 4,509 images)

Per-class

Class	mAP@50	mAP@50-95	Precision	Recall	F1
Longitudinal (D00)	0.590	0.324	0.676	0.519	0.587
Transverse (D10)	0.593	0.299	0.684	0.503	0.580
Alligator (D20)	0.687	0.376	0.712	0.616	0.660
Pothole (D40)	0.688	0.348	0.738	0.613	0.670
Overall	0.640	0.337	0.702	0.564	0.625

vs. ground-truth-only baseline (YOLO26n_RDD_Base)

Metric	Base	Distilled (this model)	Δ
mAP@50	0.635	0.640	+0.005
mAP@50-95	0.334	0.337	+0.003
Recall	0.559	0.564	+0.005
F1	0.621	0.625	+0.004

Distillation improves D00, D20, and D40 (pothole), with recall up across the board.

Evaluated against the original RDD ground truth, which has known missing annotations, so absolute recall is conservative for all models (the comparison is fair — both models use the same held-out val).

Usage

from ultralytics import YOLO
model = YOLO("YOLO26n_RDD_FRDC_Distilled.pt")
results = model("road.jpg")

Datasets

• Base: TamAko783/Unified_Road_Defect_Dataset (RDD-2022 + UAV-PDD2023 + RoadDamageVision, 30,186 images, 4 classes)
• Distillation set: TamAko783/Unified_Road_Defect_FRDC_Pseudolabeled

Credits

• RDD-2022 (Arya et al.) · UAV-PDD2023 · RoadDamageVision (Silva Zendron & Leithardt, CC BY 4.0).
• Distillation teacher: Co-DETR Swin-L — FRDC (Wang Fangjun et al.), ORDDC'2024 winner, Apache-2.0.
• Student model: AGPL-3.0 (Ultralytics).