Union14M-U-STR: Unlabeled Scene Text Recognition Dataset

Dataset Description

Union14M-U-STR contains 10M unlabeled images collected from 3 large datasets for self-supervised learning in Scene Text Recognition (STR). Images are collected using IoU Voting method.

Key Features

• 10M unlabeled images for self-supervised learning
• 3 data sources: book32, common_crawl, openvino
• No text labels - designed for self-supervised approaches
• IoU Voting collection method

Dataset Structure

{
    "image": PIL.Image,
    "text": null,
    "source": str,
    "image_id": str,
    "has_label": false
}

Splits

• train: All unlabeled images (no validation/test splits for unlabeled data)

Data Sources

• book32: Book-related text images
• common_crawl: Web images from Common Crawl
• openvino: Images from OpenVINO dataset

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Bekhouche/Union14M-U-STR")

# Access training data
train_data = dataset["train"]

# Example usage for self-supervised learning
for sample in train_data:
    image = sample["image"]
    source = sample["source"]
    # Use for self-supervised STR training...

Citation

If you use this dataset, please cite the original Union14M paper.