YOLO26s_RDD_Base

A YOLO26s (small, ~9M params) road-damage detector trained on ground-truth only — no distillation. It is the supervised baseline for a larger-student capacity study: does a bigger student (s) extract more from teacher pseudo-labels than the nano student did?

4-class CRDDC: D00 longitudinal, D10 transverse, D20 alligator, D40 pothole.

Training

• Data: Unified Road Defect Dataset (RDD-2022 + UAV-PDD2023 + RoadDamageVision), ground-truth labels only, 25,677 train imgs.
• 100 epochs, imgsz 640, batch auto — identical config to the YOLO26n baseline, so the cross-size comparison is fair.

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

Class	mAP@50	mAP@50-95	Precision	Recall	F1
Longitudinal (D00)	0.630	0.351	0.699	0.564	0.625
Transverse (D10)	0.669	0.352	0.710	0.583	0.640
Alligator (D20)	0.715	0.396	0.742	0.640	0.687
Pothole (D40)	0.733	0.388	0.774	0.649	0.706
Overall	0.687	0.372	0.731	0.609	0.665

Capacity vs the nano baseline (same GT-only data, same recipe)

Metric	YOLO26n_RDD_Base (2.4M)	YOLO26s_RDD_Base (9M)	Δ
mAP@50	0.635	0.687	+0.052
mAP@50-95	0.334	0.372	+0.038
F1	0.621	0.665	+0.044

The larger student lifts absolute accuracy substantially. Whether it also extracts a larger distillation gain is the open question, tested by the companion model YOLO26s_RDD_FRDC_Distilled_v2 (same s-model trained on the two-teacher pseudo-labeled set).

Evaluated against the original RDD ground truth, which has known missing annotations, so absolute precision/recall are conservative (the comparison is fair — every model uses the identical held-out val and is never trained on it).

Usage

from ultralytics import YOLO
model = YOLO("YOLO26s_RDD_Base.pt")
results = model("road.jpg")

Credits

• Datasets: RDD-2022 (Arya et al.) · UAV-PDD2023 · RoadDamageVision (Silva Zendron & Leithardt, CC BY 4.0).
• Model: YOLO26s, AGPL-3.0 (Ultralytics).