Datasets:

Bekhouche
/

Union14M-STR

language: en
license: mit
task_categories:
  - image-to-text
  - text-recognition
  - self-supervised-learning
tags:
  - scene-text-recognition
  - str
  - ocr
  - computer-vision
  - multimodal
  - self-supervised
size_categories:
  - 10M<n<100M

Union14M-STR: Combined Scene Text Recognition Dataset

Dataset Description

Union14M-STR is a combined dataset containing both labeled and unlabeled images for comprehensive Scene Text Recognition (STR) training. It combines Union14M-L (4M labeled) and Union14M-U (10M unlabeled) datasets.

Key Features

4M labeled images from 14 public datasets
10M unlabeled images for self-supervised learning
Combined training supporting both supervised and self-supervised approaches
5 difficulty levels for labeled data
Multiple data sources for comprehensive coverage

Dataset Structure

{
    "image": PIL.Image,
    "text": str or null,
    "source": str,
    "subset": str,
    "has_label": bool
}

Splits

train: Combined labeled and unlabeled training data
valid: Labeled validation data only

Usage

from datasets import load_dataset

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

# Access different splits
train_data = dataset["train"]
valid_data = dataset["valid"]

# Filter by label availability
labeled_data = train_data.filter(lambda x: x["has_label"] == True)
unlabeled_data = train_data.filter(lambda x: x["has_label"] == False)

# Example usage
for sample in train_data:
    image = sample["image"]
    text = sample["text"]  # May be None for unlabeled data
    has_label = sample["has_label"]
    
    if has_label:
        # Supervised training
        pass
    else:
        # Self-supervised training
        pass

Citation

If you use this dataset, please cite the original Union14M paper and acknowledge the source datasets.