darknight054/indicstr12-full

IndicSTR12 (Full Images)

This document describes the full-image portion of the IndicSTR12 real dataset.

Source

This dataset was downloaded from the CVIT IndicSTR12 Project.

Paper: IndicSTR12: A Dataset for Indic Scene Text Recognition
Authors: Harsh Lunia, Ajoy Mondal, C V Jawahar
Conference: ICDAR 2023

Note: This repository contains only the Real Dataset. The Synthetic Dataset is not included.

Structure

raw_data/
├── assamese/
│   ├── [image_id].jpeg
│   └── [image_id]_gt.txt
├── bengali/
├── gujarati/
├── hindi/
├── kannada/
├── malayalam/
├── marathi/
├── odia/
├── punjabi/
├── tamil/
├── telugu/
└── urdu/

Ground Truth (_gt.txt)

Each line is tab-separated:

word_id    x1  y1  x2  y2  x3  y3  x4  y4    text

• (x1,y1)...(x4,y4) are the four polygon corners of the word box in the image.
• text is the Unicode transcription for that word.

Bounding Boxes (Polygon)

In the converted dataset, bboxes are stored as the original polygon:

[x1, y1, x2, y2, x3, y3, x4, y4]

Citation

If you use this dataset in your research, please cite:

@inproceedings{lunia2023indicstr12,
  title={IndicSTR12: A Dataset for Indic Scene Text Recognition},
  author={Lunia, Harsh and Mondal, Ajoy and Jawahar, CV},
  booktitle={International Conference on Document Analysis and Recognition},
  pages={233--250},
  year={2023},
  organization={Springer}
}

Acknowledgements

This work is supported by MeitY, Government of India, through the NLTM-Bhashini project.

License

Please refer to the original dataset source for licensing information.