---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: annotations
    struct:
    - name: polygons
      list:
        list: int32
    - name: texts
      list: string
  splits:
  - name: train
    num_bytes: 15232741472.28
    num_examples: 1544
  download_size: 15214355688
  dataset_size: 15232741472.28
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: mit
task_categories:
- object-detection
language:
- km
tags:
- TrorYongOCR
pretty_name: KhmerST
---

# Disclaimer: This is not my dataset. I put it here to ease its use for **Khmer Scene Text Detection and Recognition** research.

# KhmerST

This repository provides scene-text images from the **KhmerST** benchmark dataset.

---

## ✅ Credits / Citation

If you use this dataset, please **cite the original KhmerST paper**:

**KhmerST: A Low-Resource Khmer Scene Text Detection and Recognition Benchmark**  
Vannkinh Nom, Souhail Bakkali, Muhammad Muzzamil Luqman, Mickaël Coustaty, Jean-Marc Ogier  

📄 Paper: https://arxiv.org/pdf/2410.18277

You can use the bibtex below:
```bibtex
@inproceedings{nom2024khmerst,
  title={KhmerST: a low-resource khmer scene text detection and recognition benchmark},
  author={Nom, Vannkinh and Bakkali, Souhail and Luqman, Muhammad Muzzamil and Coustaty, Micka{\"e}l and Ogier, Jean-Marc},
  booktitle={Proceedings of the Asian Conference on Computer Vision},
  pages={1777--1792},
  year={2024}
}
```

Original dataset source:  
https://gitlab.com/vannkinhnom123/khmerst  

---

## Dataset Description

KhmerST is the **first Khmer scene-text dataset** consisting of:

- 1,544 annotated images 
  - 997 indoor scenes
  - 547 outdoor scenes
- Diverse conditions:
  - flat and raised text
  - low illumination
  - distant and partially occluded text
- Line-level text annotations
- Polygon bounding boxes

---

## Dataset Format

Each sample contains the following columns:

| Column     | Type   | Description |
|------------|--------|-------------|
| `image`    | PIL Image | PIL Image object in RGB |
| `annotations` | dict | Dictionary of keys, `polygons`, and `texts` |

`polygons` is a list of lists of `8` integers, each list of `8` integers, `[x1, y1, x2, y2, x3, y3, x4, y4]`, corresponds to a polygon detected in the image.

`texts` is a list of texts, each is enclosed in a polygon.

The order of elements in `polygons` and `texts` are strict: the first text in `texts` is enclosed in the first polygon in `polygons`, and so on.

Example:

```json
{'image': <PIL.Image.Image image mode=RGB size=2500x2445>,
 'annotations': {
    'polygons': [ # in format [x1, y1, x2, y2, x3, y3, x4, y4]
      [104, 538, 827, 628, 824, 749, 97, 670],
      [1227, 735, 1433, 733, 1435, 837, 1231, 843]
    ],
    'texts': ['បាយស្រូបបឋម មីស៊ុបបឋម', 'អេស៊ីលីដា']
  }
}
```

## Usage

You can load the dataset with:
```python
from datasets import load_dataset

ds = load_dataset("KrorngAI/KhmerST")
print(ds["train"][0])
```

## Acknowledgment
All credit goes to the KhmerST dataset creators
``` bash
Vannkinh Nom, Souhail Bakkali, Muhammad Muzzamil Luqman,
Mickaël Coustaty, and Jean-Marc Ogier.

```