---
license: cc-by-sa-4.0
pretty_name: '"ImgMT: Wikipedia-based Dataset For Evaluating Text-Image Translation"'
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: dev
    path:
    - data/dev/metadata.jsonl
    - data/dev/*/png/**
  - split: train
    path:
    - data/train/metadata.jsonl
    - data/train/*/png/**
  - split: test
    path:
    - data/test/metadata.jsonl
    - data/test/*/png/**
dataset_info:
  features:
  - name: source_language
    dtype: string
  - name: target_language
    dtype: string
  - name: source_image_file_name
    dtype: image
  - name: target_image_file_name
    dtype: image
  - name: texts
    sequence: string
  - name: translated_texts
    sequence: string
  - name: bounding_boxes
    sequence:
      struct:
      - name: x
        dtype: float64
      - name: 'y'
        dtype: float64
      - name: w
        dtype: float64
      - name: h
        dtype: float64
tags:
- image
- text
task_categories:
- image-text-to-text
- image-text-to-image
- translation
---

# Dataset Card for ImgMT: Wikipedia-based Dataset For Evaluating Text-in-Image Translation

## Dataset Summary

The **ImgMT** dataset is a multilingual multimodal benchmark consisting of infographics, maps, scientific diagrams, and educational illustrations collected from Wikimedia Commons. In this dataset, visually identical vector graphics differ exclusively in the language of the embedded textual labels. Bounding boxes and localized texts appearing at equivalent spatial positions across language versions are aligned as mutual translation pairs.

ImgMT is created to evaluate end-to-end **Text-in-Image Translation (TIT)** systems, optical character recognition (OCR) pipelines, and localized multimodal translation models.

The complete dataset contains **129 unique languages** across **more than 20 distinct writing systems (scripts)**, supporting evaluation across thousands of 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 source language
- A **target image** with the translated text in target language
- Ground-truth spatial annotations (**bounding boxes**) defining text locations
- Aligned strings pairing the extracted source texts with reference gold translations

### Dataset characteristics

- Images contain text embedded in graphics – e.g., diagrams, maps, educational illustrations (no natural scene text).
- Images in a group should be visually identical except for the language of the text.
- Text elements are mapped across language variants using a multi-stage pipeline combining SVG `<switch>` parameters, XML `id` tracking, and spatial bounding box overlap (IoU).

### Dataset Sources

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


## Uses

The dataset is suitable for:
- **Text-in-Image Translation (TIT):** End-to-end evaluation of systems reading source text visually and translating it into target text.
- **Visual OCR & Text Detection Benchmarking:** Evaluating text detectors and character recognition engines across diverse typographic scripts (Latin, Cyrillic, Arabic, Devanagari, Dravidian, Chinese, etc.).
- **Multimodal Alignment Analysis:** Investigating the impact of graphical and visual contexts on translation disambiguation.
- **Localized Image-to-Image Machine Translation (IIMT)**

## Dataset Structure

### Dataset Organization

The dataset is divided into three splits (approximately 80:10:10 ratio):

- `train/` — train set for fine-tuning
- `dev/` — development set used for hyperparameter 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), SVG files (in `svg/` folder), and reference JSON files (`{sourceLangCode}-{targetLangCode}.json`).

```text
data/
├── train/
│   ├── {languageGroupNumber}/
│   │   ├── svg/
│   │   │   ├── {langCode}.svg
│   │   ├── png/
│   │   │   ├── {langCode}.png
│   │   ├── {sourceLangCode}-{targetLangCode}.json
├── dev/
│   └── (Same structure as train)
└── test/
    └── (Same structure as train)
```


### Data Format

Each translation pair is stored in **JSON format** with the following structure:

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

  "source_PNG": {
    "size": {
      "width": px,
      "height": px
    },
    "wikimedia_url": "url",
    "licensing": {
        "license": "license_name",
        "original_license": "license_name",
        "author": "user_name"
    }
  },

  "source_texts": ["string"],

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

  "target_language": "language code",

  "target_PNG": {
    "size": {
      "width": px,
      "height": px
    },
    "wikimedia_url": "url",
    "licensing": {
        "license": "license_name",
        "original_license": "license_name",
        "author": "user_name"
    },
  },

  "target_texts": ["string"],

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

```

| Field                        | Description                                     |
| ---------------------------- | ----------------------------------------------- |
| `alignment_method`           | Method used for pairing source and target texts |
| `source_language`            | Language code of the source text                |
| `source_PNG`                 | Information about the source image              |
| `source_texts`               | Source texts extracted from the image           |
| `source_text_bounding_boxes` | Bounding boxes of source text regions           |
| `target_language`            | Language code of the translated text            |
| `target_PNG`                 | Information about the translated image          |
| `target_texts`               | Translated texts corresponding to the source    |
| `target_text_bounding_boxes` | Bounding boxes of target text regions           |

Note: Elements across `source_texts` and `source_text_bounding_boxes`, `target_texts` and `target_text_bounding_boxes`, `source_texts` and `target_texts` are aligned — items at the same index correspond to the same text region.


## Dataset Statistics

- **Total Translation Pairs (JSON instances):** 35,612
- **Total Unique Languages:** 129
- **Unique Typographic Scripts:** 20+ *(Latin, Cyrillic, Arabic, Devanagari, Dravidian, Greek, Japan, etc.)*
- **Number of Language Groups:** ~1,855 unique graphic families
- **Language Group Sizes:** 2 to 70+ language variants per graphic family


## Evaluation Framework

We evaluate Text-in-Image Translation in two separate steps: text recognition and translation quality. For text recognition, we measure Character Error Rate (CER) and bounding box overlap (F1). For translation, we evaluate the generated text using chrF and COMET scores. If a model does not predict bounding box coordinates, our evaluation script automatically pairs the texts using bipartite matching based on text similarity and reading order.  


## 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.
- Text strings included in XML attributes `<path>` (represented as curves, not as text) could not be processed by our extracting pipeline, therefore they were not included in the final JSON metadata.


## Licensing

### Dataset Licensing

This dataset is distributed under a CC BY-SA 4.0 license. This license applies specifically to the annotations (bounding boxes, text alignments and structured JSON metadata) created for this project.


### Image Licensing

Please note that the images included in this dataset originate from Wikimedia Commons. While the dataset as a whole is provided under CC BY-SA 4.0, individual source images are subject to their own respective licenses as specified by their original authors (e.g., various versions of CC BY, CC BY-SA, or Public Domain). Detailed attribution for every image, including the specific author and original license, can be found in the corresponding JSON files.


## Citation

```
@inproceedings{uramova-etal-2026-imgmt,
    title = "{I}mg{MT}: {W}ikipedia-based Dataset For Evaluating Text-in-Image Translation",
    author = "Uramov{\'a}, {\v{S}}{\'a}rka  and
      Petrov, Petar  and
      Jon, Josef  and
      Bojar, Ond{\v{r}}ej  and
      Nov{\'a}k, Michal",
    editor = "Haddow, Barry  and
      Kocmi, Tom  and
      Koehn, Philipp  and
      Monz, Christof",
    booktitle = "Proceedings of the Eleventh Conference on Machine Translation",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.wmt-main.43/",
    doi = "10.18653/2026.wmt-main.43",
    pages = "765--781",
    ISBN = "979-8-89176-474-3",
    abstract = "Text-in-image translation aims to translate all textual content embedded in an image. In Text Image Translation (TIT), the output is translated text; In-Image Machine Translation (IIMT) additionally renders the translation back into the image. We introduce ImgMT, a massively multilingual dataset and benchmark mined from localized SVG illustrations on Wikimedia Commons, focusing on diagrams, maps, and infographics. ImgMT contains 35,612 aligned image pairs grouped into 1,855 image sets, spanning 129 languages, 20+ writing systems, and 4,886 language pairs.\par We propose a multi-level evaluation protocol that measures source-side text-region detection and OCR accuracy alongside target-side translation quality. We benchmark several multimodal large language models and a dedicated TIT system in English-to-X and X-to-English settings, and find that the source language and writing system substantially affect TIT quality. Finally, we show that ImgMT is not well suited for isolating the contribution of visual information in multimodal translation."
}
```

## Credits

We would like to acknowledge the community of creators on Wikimedia Commons whose work made this dataset possible.
For a complete list of image titles and their respective authors/contributors, please refer to the credits.md file included in this repository. This file provides a mapping between the dataset filenames and the original creators to ensure full transparency and compliance with original image licenses.