You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

detect_nl_subsigns

YOLOv8 model for detecting sub-signs (Dutch: onderbord) below main road signs in the Netherlands. Fine-tuned from Panoramax/detect_fr_road_signs_subsigns on Dutch street-level imagery from the Panoramax platform.

What is an onderbord?

An onderbord is a supplementary sign from the Dutch OB-series (RVV 1990) mounted directly below a main road sign. This model detects their presence as a single class — it does not classify the OB type or code.

Model details

Property	Value
Architecture	YOLOv8 (Ultralytics)
Base model	`Panoramax/detect_fr_road_signs_subsigns` (`panonceau` class)
Fine-tuning epochs	500
Input	Cropped image around a detected main sign
Output class	`onderbord` (1 class)
Training data	Dutch Panoramax imagery — see dataset card

Performance

Validation set (sign crops, 80/20 split)

Metric	Value
mAP50	0.993
Precision	0.977
Recall	0.962

Live evaluation on 727 real Dutch road signs (Panoramax DataFusion, 2026-07-13)

Evaluated per main sign using the /v1/annotate/ API endpoint. ~13% of signs had a real sub-sign.

Confidence threshold	Precision	Recall	F1
0.25 (default)	68.3%	99.0%	80.9%
0.60 (balanced)	79.5%	96.9%	87.3%
0.90 (best F1)	90.3%	87.5%	88.9%

Note on precision at conf=0.25: the low precision reflects a class imbalance in the live dataset (13% positive rate) and a per-box counting method. The model misses only 1 in 96 real sub-signs at this threshold. For production use, conf=0.60–0.90 is recommended depending on your tolerance for false positives.

Usage

from ultralytics import YOLO

model = YOLO("Panoramax/detect_nl_subsigns")

results = model.predict("sign_crop.jpg", conf=0.60)

The model expects a crop around a detected main sign — not a full street photo. Use a bounding box expanded slightly below the main sign to ensure the sub-sign is visible in the crop.

Training

Fine-tuned with the following hyperparameters:

Parameter	Value
Epochs	500
Batch size	16
Frozen backbone layers	10
Initial learning rate	0.01
Framework	Ultralytics YOLOv8

Training data sourced from Panoramax via STAC, filtered on osm\|traffic_sign semantics with Dutch (NL:) values. Images were annotated in Label Studio.

License

Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)

Downloads last month: -

Model tree for Panoramax/detect_nl_subsigns

Base model

Panoramax/detect_fr_road_signs_subsigns

Finetuned

(1)

this model