Panoramax/nl_subsign_for_detection

nl-subsigns-dataset

YOLO-format dataset of Dutch sub-signs (onderbord, OB-series, RVV 1990) for training and evaluating object detection models. Images are crops around detected main road signs sourced from the Panoramax platform using Dutch street-level imagery.

Dataset structure

dataset/
├── images/
│   ├── train/      # 560 images
│   └── val/        # 140 images
├── labels/
│   ├── train/      # YOLO format (.txt), one file per image
│   └── val/
├── data.yaml
test/
├── images/         # 1 000 images (independent test set)
├── labels/
└── data.yaml

Total: 1 700 images — 700 train/val + 1 000 independent test images.

Classes

ID	Name	Description
0	onderbord	Supplementary sign (OB-series, RVV 1990) below a main road sign

One class only — OB type or code classification is out of scope.

Split details

Split	Images	Positives (with sub-sign)	Negatives
Train	560	~50%	~50%
Val	140	~50%	~50%
Test	1 000	independent sample	independent sample

Train/val is an 80/20 random split. The test set was fetched separately and is disjoint from train/val by Panoramax picture ID — it is a genuinely independent benchmark.

Source

Images fetched from the Panoramax STAC API, filtered on osm|traffic_sign semantics with Dutch (NL:) values. Each image is a crop around a detected main traffic sign. Crops are sized to include the area directly below the main sign bounding box.

Annotations were created manually in Label Studio.

Label format

Standard YOLO bounding box format per image:

<class_id> <x_center> <y_center> <width> <height>

All values normalised to [0, 1]. Negative images have an empty label file.

Related model

Panoramax/detect_nl_subsigns — trained on this dataset.

License

Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) — inherited from the Panoramax source imagery.