calvadosdep/yolo26n-directional-signs

YOLO26n — French directional road sign detector (v4)

YOLO26n fine-tuned to detect French directional road signs in street-level imagery. Trained on human-reviewed detections (SignValidation) with confirmed false positives as hard negatives, plus manually added giratoire examples.

Classes

id	name
0	directionnel
1	giratoire
2	communes (panneaux d'entrée d'agglomération)

Training

• Base: yolo26n.pt, imgsz=1024, scale=0.9, early stop at epoch 161 (best 111, A10G)
• Data: calvadosdep/dir-signs-training — 955 train / 108 val images: fully reviewed Panoramax detections (validated boxes with reviewer-corrected classes + 330 confirmed-FP background images), 54 manually added giratoire photos, and the original pierrelortie dataset.

Validation (108 real-world images, human ground truth)

class	mAP50	mAP50-95
all	0.969	0.856
directionnel	0.928	0.783
giratoire	0.995	0.962*
communes	0.985	0.825

P 0.961 · R 0.940. *giratoire val images come from the same driving sequence as part of the train set — treat this figure as optimistic.

Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download("calvadosdep/yolo26n-directional-signs", "best.pt")
model = YOLO(weights)
results = model.predict("image.jpg", conf=0.25, imgsz=1024)

Note: YOLO26 is NMS-free but occasionally emits near-duplicate boxes — apply a same-class IoU>0.7 dedup keeping the highest confidence.

History

• v4 (current, root best.pt = v4/best.pt): full review coverage +54 manual giratoire images — mAP50-95 0.856, giratoire finally functional
• v3 (v3/): first retrain on reviewed data + hard negatives — mAP50-95 0.829, −84% confirmed false positives vs v2
• v2, v1: in git history