GTSRB

Overview

The German Traffic Sign Recognition Benchmark (GTSRB) is an image classification dataset with 43 traffic sign classes. The official final distribution has 39,209 training images and 12,630 test images, for 51,839 images in total. There is no validation split. Labels retain the benchmark’s integer class IDs, 0 through 42; the ClassLabel names are the corresponding numeric strings.

The original images are PPM files with variable width and height. They are decoded as RGB without resizing, cropping, padding, or normalization. The dataset’s cache stores them as PNG while preserving their pixels and dimensions.

The training ZIP includes per-class annotation CSVs, but labels are read from the image class directories. The test image ZIP includes annotations without labels; the builder matches test images to the separate ground-truth CSV by filename. Width, height, and sign-region coordinates remain in the source files and are not exposed as dataset columns.

Data Structure

Each example has these fields:

Key	Type	Description
image	PIL.Image.Image	Original-size RGB traffic sign image
label	int	Official class ID, 0 through 42

Usage Example

from stable_datasets.images import GTSRB

train = GTSRB(split="train")
test = GTSRB(split="test")

print(len(train))  # 39209
print(len(test))   # 12630

sample = train[0]
print(sample.keys())
print(sample["image"].size)
print(sample["label"])

Source and License

The maintained benchmark distribution links to the official ERDA/SID GTSRB archive. The archive provides GTSRB_Final_Training_Images.zip, GTSRB_Final_Test_Images.zip, and GTSRB_Final_Test_GT.zip. The benchmark downloads moved to external hosting in 2019. The builder tries the original ERDA URLs first, then raw Wayback Machine copies of the same ZIPs if needed. Downloaded files are checked against their SHA-256 hashes. The benchmark and archive do not state a license suitable for the dataset license field, so it is left empty.

Citation

@inproceedings{Stallkamp-IJCNN-2011,
  author = {Johannes Stallkamp and Marc Schlipsing and Jan Salmen and Christian Igel},
  title = {The {G}erman {T}raffic {S}ign {R}ecognition {B}enchmark: A multi-class classification competition},
  booktitle = {IEEE International Joint Conference on Neural Networks},
  pages = {1453--1460},
  year = {2011}
}