Yuvrajg2107/pranavx-gtsrb-reliability

GTSRB for PranavX AI Reliability Experiments

This repository repackages the German Traffic Sign Recognition Benchmark (GTSRB) for a traffic-sign classification reliability experiment. It preserves the original PPM image bytes and GTSRB class IDs in Parquet. No image is resized, cropped further, enhanced, or synthetically corrupted here. This is a project-specific derivative split, not a new official GTSRB release or an official benchmark leaderboard split.

The uploader reports receiving written permission for public rehosting of the GTSRB images. The permission text has not been supplied for independent review, so this card does not claim a general redistribution license or permission for downstream users to republish the images. The original public distribution pages do not specify a machine-readable redistribution license. No CC0, MIT, or other open license is claimed for the source images. Consult the original archive and obtain any rights you need before further redistribution.

Purpose and experimental protocol

GTSRB contains cropped images with one traffic sign per image. The original dataset has 43 class IDs. For this project's known-versus-unseen-class experiment, IDs 0–34 are known to the classifier and IDs 35–42 are withheld from classifier training. This 35/8 choice is an experimental protocol, not a property or requirement of the original benchmark. A published traffic-sign OOD study has used a similar GTSRB class-withholding setup: Iyengar et al..

Split	What it contains	Permitted role in this project
train	Original training images, IDs 0–34	Fit the image classifier.
validation	Different physical-sign training tracks, IDs 0–34	Choose the architecture and checkpoint; tune hyperparameters here.
calibration	Different physical-sign training tracks, IDs 0–34	Fit temperature scaling and freeze known-class reliability thresholds.
unused_ood_train	Original training images, IDs 35–42	Archive completeness only. Do not use to fit the classifier or choose OOD thresholds in the claimed unseen-class experiment.
test_known	Original official test images, IDs 0–34	Final evaluation of clean recognition and reliability only.
test_ood	Original official test images, IDs 35–42	Final held-out-class OOD evaluation only.

Verified row counts are shown in this table:

Split	Images
train	23,940
validation	5,100
calibration	5,159
unused_ood_train	5,010
test_known	11,040
test_ood	1,590

Together these are all 39,209 original training images and all 12,630 original test images. The exact per-class counts and source ZIP SHA-256 values are in manifest.json.

The train, validation, and calibration assignments are deterministic and grouped by GTSRB physical-sign track within each known class. The same sign's near-duplicate frames never cross these three splits. Tracks are shuffled per class with the seed and algorithm recorded in manifest.json. The official test set is never used to train, calibrate, or select thresholds. The test images are split into known and unseen classes only after the protocol is fixed.

Withholding eight classes does not make this a complete real-world OOD benchmark. The unseen signs still come from GTSRB's acquisition setting. Live phone images, full road scenes, non-German sign systems, and unrelated objects are outside this benchmark. A full-scene upload needs a detector or manual crop before this classifier can be evaluated fairly.

Data fields

Each row has:

• image: the exact original PPM file bytes, exposed as a Hugging Face Image feature.
• class_id: original integer label from GTSRB, unchanged.
• original_path: path within the corresponding official ZIP.
• source_split: official_train or official_test.
• track_id: <class>/<physical-sign-track> for original training images. It is null for ordinary official test images because their ground-truth CSV does not provide the same track key.
• width, height: original dimensions.
• roi_x1, roi_y1, roi_x2, roi_y2: original annotated sign region coordinates.

The files are sharded Parquet with embedded images. manifest.json records exact row and track counts by split, source archive SHA-256 values, class counts, and the split seed. The preparation script, prepare.py, validates archive records, image dimensions, and track separation before writing.

Loading and using it

from datasets import load_dataset

ds = load_dataset("Yuvrajg2107/pranavx-gtsrb-reliability")
known_train = ds["train"]
calibration = ds["calibration"]
ood_test = ds["test_ood"]

sample = known_train[0]
print(sample["class_id"], sample["image"].size)

The dataset contains unprocessed images with varying sizes. Resize and normalize inside the model pipeline. Use the same deterministic preprocessing for validation, calibration, testing and single-image demonstrations. Train-time augmentation must apply only to train. Do not merge unused_ood_train into train, and do not retune thresholds after inspecting test_known or test_ood outcomes.

To reproduce the packaged data from the three original ZIPs:

python prepare.py --source "C:\path\to\official_gtsrb_zips" --output "C:\path\to\empty_output_folder"

Required Python packages: datasets, pyarrow, and Pillow. The source files are GTSRB_Final_Training_Images.zip, GTSRB_Final_Test_Images.zip, and GTSRB_Final_Test_GT.zip from the official archive. The output directory must be new or empty. The script does not modify those ZIPs.

Source and limitations

Original dataset: Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel, The German Traffic Sign Recognition Benchmark: A Multi-class Classification Competition, IJCNN 2011, pp. 1453–1460. The maintainers' benchmark page asks users to cite this paper. The original archive identifies the official training and test files.

This collection was designed for traffic-sign recognition of an already cropped sign, not whole-scene detection or deployment in a vehicle. Its images vary in resolution and lighting and may not represent current roads, countries outside Germany, or all sign designs. A model's confidence on a sign does not certify safe driving. Reliability labels in the associated project are research outputs, not operational safety approvals.

Project relationship

This repository is the data input for the PranavX AI Reliability Engine project. The project's own evaluation will report calibrated confidence, uncertainty, OOD scores, selective prediction decisions, error analysis, and failure cases. This dataset card does not claim those results already exist.