Union14M-L-STR: Labeled Scene Text Recognition Dataset

Dataset Description

Union14M-L-STR contains 4M labeled images collected from 14 public available datasets for Scene Text Recognition (STR). This dataset has been refined through several strategies including cropping and de-duplication.

Key Features

• 4M labeled images from 14 public datasets
• Cropped images using minimal axis-aligned bounding boxes
• De-duplicated to remove duplicate images
• 5 difficulty levels: easy, medium, hard, challenging, normal
• Benchmark splits for 9 different challenges

Dataset Structure

{
    "image": PIL.Image,
    "text": str,
    "difficulty": str,
    "source_dataset": str,
    "original_filename": str
}

Splits

• train: Training data with different difficulty levels
• valid: Validation data
• test: Test data (same as validation for now)
• benchmark_*: Various benchmark categories (artistic, curve, etc.)

Source Datasets

The dataset combines images from 14 public datasets including:

• art_curve, art_scene, COCOTextV2, hier_curve, hier_scene
• IIIT-ILST, KAIST, LSVT, MLT19, MTWI, neocr_dataset
• OpenVINO, ReCTS, RCTW, TextOCR, Uber

Usage

from datasets import load_dataset

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

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

# Example usage
for sample in train_data:
    image = sample["image"]
    text = sample["text"]
    difficulty = sample["difficulty"]
    # Process your data...

Citation

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