YOLO26n_RDD_FRDC_Distilled

YOLO26n · Distilled v1 (1 teacher: Co-DETR)

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

Method: Semi-supervised distillation, single teacher. Co-DETR (Swin-L) pseudo-labeled the unlabeled RDD-2022 test set; student trained on GT + pseudo-labels.

Metrics — RDD held-out validation (4,509 images, 11,470 boxes)

Full model comparison (same held-out val)

All five models in this study, evaluated identically (imgsz 640):

Reading it:

• Distillation helps — every distilled model beats its GT-only baseline.
• Capacity helps most — the YOLO26s models (+~0.05 mAP@50) clearly outperform the nano ones on this val.
• One vs two teachers is a near-tie at nano size; the two-teacher set's edge is small.
• Cross-domain check (independent RDDC2024-ID, Indonesia, 8,901 imgs): distilled models generalized better than baselines, while the larger GT-only model generalized worse — evidence the distillation's added data improves robustness, not just in-domain fit.

RDD ground truth has known missing annotations, so absolute precision/recall are conservative for all models. The comparison is fair — every model uses the identical held-out val, never trained on.

Usage

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

Models in this suite

• YOLO26n_RDD_Base · YOLO26n_RDD_FRDC_Distilled (v1) · YOLO26n_RDD_FRDC_Distilled_v2
• YOLO26s_RDD_Base · YOLO26s_RDD_FRDC_Distilled_v2

Datasets

• Base: TamAko783/Unified_Road_Defect_Dataset
• Distillation sets: v1 · v2 (two-teacher)

Credits

• Datasets: RDD-2022 (Arya et al.) · UAV-PDD2023 · RoadDamageVision (Silva Zendron & Leithardt, CC BY 4.0).
• Distillation teachers: Co-DETR Swin-L + RTMDet-x — FRDC (Wang Fangjun et al.), ORDDC'2024 winner.
• Student: YOLO26n, AGPL-3.0 (Ultralytics).