Panoramax/nl_subsign_for_classification

nl-subsigns-classification-dataset

Image classification dataset of Dutch sub-signs (onderbord, OB-series, RVV 1990). Images are crops around detected sub-signs sourced from the Panoramax platform.

Used to train Panoramax/classify_nl_subsigns.

Dataset structure

Standard ImageFolder layout (compatible with PyTorch ImageFolder, Ultralytics YOLOv8 classify):

dataset/
├── train/          # 2,892 images
│   ├── AMSCAM/
│   ├── OB02/
│   ├── OB11/
│   └── ...         # one folder per class
└── val/            # 725 images
    ├── AMSCAM/
    ├── OB02/
    ├── OB11/
    └── ...

Classes (24)

Class	Train	Val	Description
AMSCAM	44	11	Amsterdam/camera zone sign
OB02	7	2	Distance indication
OB11	12	3	Except cyclists
OB12	3	1	Except mopeds
OB303	76	19	Direction arrow left
OB304	84	21	Direction arrow right
OB501l	98	25	Lane indication left
OB501r	60	15	Lane indication right
OB502	155	39	Lane indication both
OB503	111	28	Lane indication ahead
OB503OB02	43	11	Combination OB503 + OB02
OB503OB04	22	6	Combination OB503 + OB04
OB504	205	51	Lane indication ahead + both
OB504l	18	4	Lane indication ahead + left
OB504r	14	4	Lane indication ahead + right
OB505	216	54	Lane indication all directions
OB52	151	38	Time restriction
OB54	131	33	Day restriction
OB62	10	3	Vehicle type exception
OB64	40	10	Vehicle type exception variant
OB711	10	2	Zone start
OB712	43	11	Zone repeat
OB713	21	5	Zone end
tekstbord	1,318	329	Text-based sign (→ OCR for exact type)
Total	2,892	725

Design choice: tekstbord

Text-based sub-signs (e.g. OB0, OB20, OB104 and variants) are visually ambiguous at low resolution. All text-based variants are merged into a single tekstbord class. An OCR module handles exact type identification from the text content. This avoids training a classifier to distinguish signs that are only differentiable by the characters printed on them.

Source

Images sourced from Panoramax via an iterative bootstrap process:

1. Initial crops were manually labelled in Label Studio.
2. A model was trained and used to predict new crops from Panoramax imagery.
3. Predictions were manually validated and added to the dataset.
4. Steps 2–3 were repeated until the dataset was large enough for a stable model.

This is dataset version 2026-07-16 — the final version used to train classify_nl_subsigns.

Related model

Panoramax/classify_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.