README.md

---
license: cc-by-sa-4.0
task_categories:
- translation
language:
- en
- fr
- de
- it
- hu
- pl
- ru
- es
- cs
- tr
- sv
- nl
- fi
- 'no'
- ro
- la
- kn
- id
- zh
- ko
- ka
- ja
- vi
- sq
- be
- et
pretty_name: Multilingual Image Translation Groups
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: dev
    path:
    - dev/metadata.jsonl
  - split: test
    path:
    - test/metadata.jsonl
  features:
  - name: source_language
    dtype: string
  - name: target_language
    dtype: string
  - name: source_file_name
    dtype: image
  - name: target_file_name
    dtype: image
  - name: texts
    sequence: string
  - name: translated_texts
    sequence: string
tags:
- image
- text
---

# Dataset Card for HuggingFaceFW/Multilingual-Image-Translation-Groups

## Dataset Summary

The IMG-MT dataset is a multilingual dataset of images containing text, where visually identical images differ only in the language of the embedded text. The texts appearing at the same positions across images are translations of each other.

The dataset was created to support the evaluation of systems that perform **text detection, translation and rendering in images**, such as pipelines for image-based machine translation or OCR-based translation systems.

All images originate from Wikimedia Commons and are grouped into language groups consisting of visually identical images that differ only in the language of the text.

The dataset contains **26 languages** and supports evaluation of translation between multiple language pairs.


## Dataset Details

### Dataset Description

The dataset consists of pairs of images containing translated text. Each pair contains:

- a **source image** with text in one language
- a **target image** with the translated text in another language

Bounding boxes for text regions are provided, along with the corresponding source and translated text strings.

### Dataset characteristics

- Images contain text embedded in graphics (e.g., diagrams, maps, educational illustrations).
- Images in a group are visually identical except for the language of the text.
- Text blocks are aligned across images using bounding boxes.

### Dataset Sources

- **Source of images:** Images were collected from **Wikimedia Commons**


## Uses

The dataset is suitable for:

- Evaluation of **image translation pipelines**
- Evaluation of **OCR + machine translation systems**
- Benchmarking **text detection in images**
- Testing **text rendering after translation**
- Multilingual **visual-language research**


## Dataset Structure

### Dataset Organization

The dataset is divided into two parts:

- `dev/` — development set used for tuning and development
- `test/` — test set used for final evaluation

Each of these sets contains several numbered folders (`{languageGroupNumber}/`), each representing a language group.

These folders store images (in `png/` folder) and reference JSON files (`{sourceLangCode}-{targetLangCode}.json`).


| Folder/File | Description |
|---|---|
| `dev/` | Development dataset |
| `test/` | Test dataset |
| `{languageGroupNumber}/` | Group of images that differ only by language |
| `png/` | Images in different languages |
| `{sourceLangCode}-{targetLangCode}.json` | Reference translation data |


### Data Format

Reference translation data are stored in **JSON format** with the following structure:

```json
{
  "source_language": "language code",

  "source_PNG": {
    "size": {
      "width": px,
      "height": px
    },
    "path_to_image": "path",
    "wikimedia_url": "url"
  },

  "text_bounding_box": [
    {
      "x": float,
      "y": float,
      "w": float,
      "h": float
    }
  ],

  "texts": ["string"],

  "target_language": "language code",

  "translated_texts": ["string"],

  "target_PNG": {
    "size": {
      "width": px,
      "height": px
    },
    "path_to_image": "path",
    "wikimedia_url": "url"
  }
}

```

| Field               | Description                                  |
| ------------------- | -------------------------------------------- |
| `source_language`   | Language code of the source text             |
| `source_PNG`        | Information about the source image           |
| `text_bounding_box` | Bounding boxes of text regions               |
| `texts`             | Source texts extracted from the image        |
| `target_language`   | Language code of the translated text         |
| `translated_texts`  | Translated texts corresponding to the source |
| `target_PNG`        | Information about the translated image       |


The lists: `text_bounding_box`, `texts`, `translated_texts` are aligned — items at the same index correspond to the same text region.


## Dataset Statistics

The dataset contains 26 languages and multiple language groups of varying sizes.

Example language statistics:


| Language  | Code | Groups | Pairings |
| --------- | ---- | ------ | -------- |
| English   | en   | 435    | 1530     |
| French    | fr   | 204    | 796      |
| German    | de   | 160    | 668      |
| Italian   | it   | 120    | 604      |
| Hungarian | hu   | 79     | 368      |
| Polish    | pl   | 67     | 398      |
| Russian   | ru   | 67     | 296      |
| Spanish   | es   | 62     | 266      |
| Czech     | cs   | 59     | 314      |
| Turkish   | tr   | 50     | 260      |
| Swedish   | sv   | 49     | 284      |
| Dutch     | nl   | 37     | 200      |
| Finnish   | fi   | 31     | 184      |
| Norwegian | no   | 28     | 146      |
| Romanian  | ro   | 28     | 146      |
| Latin     | la   |  9     |  70      |

- `Groups` = number of image groups containing the given language
- `Pairings` = number of language pairs in the dataset where the given language is either the source or target language

Smaller language groups also include languages such as: Kannada, Indonesian, Chinese, Korean, Georgian, Japanese, Vietnamese, Albanian, Belarusian and Estonian.

Language groups range in size from 2 to 12 images.



## Dataset Creation

### Source Data

Images were collected from Wikimedia Commons.

#### Data Collection and Processing

The collection process involved:

- Searching for images that exist in multiple languages

- Selecting images that are visually identical except for the text language

- Grouping images into language groups

- Extracting text regions using bounding boxes

- Recording corresponding source and translated texts

Each language group contains images representing the same graphic with text translated into different languages.


## Bias, Risks and Limitations

Several limitations should be considered:

- Language distribution is imbalanced, with English and major European languages being overrepresented.

- Some languages appear only in a small number of examples.

- The dataset mainly contains diagram-like images rather than natural scene text.

- The dataset size is relatively small, which limits its use for large-scale training.