Noor-Khan/traffic-sign-eval-split

Traffic Sign Eval Split (GTSRB held-out validation)

A labelled evaluation split of 7,841 German traffic-sign images in 43 classes, prepared by Noorullah Zamindar to evaluate his traffic-sign classifier Noor-Khan/traffic-sign-classification (validation accuracy 0.9719 on this exact split).

What is in this dataset

• Split: one split, validation — 7,841 images.
• Columns: image (the original sign photo), label (0–42, with the 43 GTSRB class names), filename (original class folder + file name, e.g. 00026/00001_00003.ppm).
• Images: original .ppm files, untouched (sizes vary; the model resizes them to 50×50 and scales pixels to [0, 1] before predicting).
• File: a single Parquet file, data/validation-00000-of-00001.parquet (1.2 MB), loads directly with the datasets library — no loading script.

from datasets import load_dataset

ds = load_dataset("Noor-Khan/traffic-sign-eval-split")
print(ds)  # {'validation': Dataset(7841 rows)}

How the split was made

This is the held-out 20% validation split of the GTSRB training images, reproduced exactly from the training script of the model (train.py in the Traffic-Sign-Classification repo): all 39,209 training images enumerated class by class (folders 00000–00042, filenames sorted), indices shuffled with numpy.RandomState(42), and the first 20% (7,841 images) taken as the validation split. The build script is kept in this repo as build_dataset.py, so the split is fully reproducible.

Per-class image counts range from 38 (class 0, speed limit 20 km/h) to 496 (class 1, speed limit 30 km/h); see build_info.json for the full table.

Why not the official GTSRB test set?

The official GTSRB final test set (12,630 images) ships without public labels — its ground-truth file GT-final_test.test.csv contains only filename, size and ROI columns, no ClassId. A labelled eval set therefore cannot be built from it. This validation split is labelled, and it is the exact split the companion model was measured on.

Source and license — please read

Noorullah Zamindar did not collect these images. They come from the German Traffic Sign Recognition Benchmark (GTSRB), created by the Institute for Neuroinformatics, Ruhr-Universität Bochum, and published for the IJCNN 2011 competition. Source: https://benchmark.ini.rub.de/?section=gtsrb&subsection=dataset

GTSRB is free to use for research, when citing the publication. This repository only re-packages a subset of the public training images as an eval split; the images remain the work of the GTSRB creators.

If you use this dataset, please cite:

@inproceedings{Stallkamp-IJCNN-2011,
  author    = {Johannes Stallkamp and Marc Schlipsing and Jan Salmen and Christian Igel},
  title     = {The German Traffic Sign Recognition Benchmark: A multi-class classification competition},
  booktitle = {IEEE International Joint Conference on Neural Networks},
  year      = {2011},
  pages     = {1453--1460},
}

and:

@article{stallkamp2012man,
  author  = {Stallkamp, Johannes and Schlipsing, Marc and Salmen, Jan and Igel, Christian},
  title   = {Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition},
  journal = {Neural Networks},
  volume  = {32},
  pages   = {323--332},
  year    = {2012},
}

The 43 classes

0. Speed limit (20km/h) · 1. Speed limit (30km/h) · 2. Speed limit (50km/h) ·
1. Speed limit (60km/h) · 4. Speed limit (70km/h) · 5. Speed limit (80km/h) ·
2. End of speed limit (80km/h) · 7. Speed limit (100km/h) ·
3. Speed limit (120km/h) · 9. No passing ·
4. No passing for vehicles over 3.5 metric tons ·
5. Right-of-way at the next intersection · 12. Priority road · 13. Yield ·
6. Stop · 15. No vehicles ·
7. Vehicles over 3.5 metric tons prohibited · 17. No entry ·
8. General caution · 19. Dangerous curve to the left ·
9. Dangerous curve to the right · 21. Double curve · 22. Bumpy road ·
10. Slippery road · 24. Road narrows on the right · 25. Road work ·
11. Traffic signals · 27. Pedestrians · 28. Children crossing ·
12. Bicycles crossing · 30. Beware of ice/snow · 31. Wild animals crossing ·
13. End of all speed and passing limits · 33. Turn right ahead ·
14. Turn left ahead · 35. Ahead only · 36. Go straight or right ·
15. Go straight or left · 38. Keep right · 39. Keep left ·
16. Roundabout mandatory · 41. End of no passing ·
17. End of no passing by vehicles over 3.5 metric tons

Limitations

• Signs are German; the set is not representative of signs in other countries.
• Class sizes are imbalanced (38–496 images per class), as in the original data.
• Images are small, cropped sign photos — not full street scenes.
• Do not use this for safety-critical decisions without extensive validation.