# ISLR Dataset for Icelandic Sign Language (ÍTM)

This ISLR dataset for Icelandic Sign Language (ÍTM) was compiled in 2026 by Finnur Ágúst Ingimundarson (FÁI), as part of a master's thesis in computational linguistics at the University of Zurich, in collaboration with The Communication Centre for the Deaf and Hard of Hearing in Iceland ([SHH](https://island.is/en/o/shh)). 

The dataset was created by drawing on available bilingual dictionary data for ÍTM and Icelandic, namely from the dictionary component of [SignWiki Iceland](https://is.signwiki.org/index.php/Forsíða).

The data was provided directly by SHH and was consequently processed, reviewed, and filtered to create the dataset, which was used to evaluate two open-source frameworks for ISLR:

1. [OpenHands](https://github.com/AI4Bharat/OpenHands)
    - An adapted version of the framework for the ÍTM Dataset can be found [here](https://github.com/ZurichNLP/openhands-itm).
2. [SPOTER](https://github.com/maty-bohacek/spoter)
    - An adapted version of the framework that supports the pose format (and the ÍTM dataset) can be found [here](https://github.com/ZurichNLP/spoter-pose).

See Publications here below for information about the thesis and a paper, where the results of these experiments are described.

## Contents

The data is published in the following formats:

1) **Video:**
    - Videos in MP4 format and 30 fps.
2. **Poses in .pkl-format**
    - This format matches the one used in [OpenHands](https://github.com/AI4Bharat/OpenHands)
3. **Poses in [.pose-format](https://github.com/sign-language-processing/pose) from three different pose estimators:**
    - [AlphaPose](https://github.com/MVIG-SJTU/AlphaPose)
    - [MediaPipe](https://github.com/google-ai-edge/mediapipe)
    - [SDPose](https://github.com/T-S-Liang/SDPose-OOD)

The files all follow the same naming scheme (except for the extension) and in the accompanying JSON files, all file names are included without the file extension.

## Dataset Statistics

### Overview

| | Count / Length |
|---|---|
| Videos | 1,845 |
| Length | 1:44:29 |
| Classes | 849 |
| Signers | 33 |

### Samples Per Class

| Samples | Count | Proportion |
|---|---|---|
| 2 | 732 | 86.22% |
| 3 | 95 | 11.19% |
| 4 | 18 | 2.12% |
| 5 | 2 | 0.24% |
| 6 | 1 | 0.12% |
| 8 | 1 | 0.12% |

### Splits

| Split | Size |
|---|---|
| Train | 917 |
| Val | 79 |
| Test | 849 |
| **Total** | 1,845 |

### Dataset versions

Two other versions of the dataset are also provided:

- **Trimmed**: Classes with three or more samples (117)
    - `itm_data_trimmed.json`
- **Minimal**: Classes with four or more samples (22)
    - `itm_data_minimal.json`

These two subsets are better balanced than the full dataset, which has only 79 out of 849 classes, and the validation metrics are therefore more reliable. 

## Structure

**Example of JSON structure:**

```json
[
  {
    "word_label": "heild",
    "clean_text": "heild",
    "no_hands": 2,
    "fingerspelled": null,
    "instances": [
        {
            "video_id": "heild_s4_v1",
            "signer_id": 4,
            "status": "Deaf",
            "hand": null,
            "split": "test",
            "duration": 3.33,
            "fps": 30.0
        },
    ]
  }
]
```

## Publications

The results are reported in the following publications:

1. **Ingimundarson's master's thesis:**

```bibtex
@mastersthesis{ingimundarson-islr-itm,
    author = {Finnur {\'A}g{\'u}st Ingimundarson},
    title = {{Isolated Sign Language Recognition for Icelandic Sign Language (ITM): Experiments in a Low-resource Setting}},
    school = {University of Zurich},
    year = {2026},
    url= {https://doi.org/10.5167/uzh-436015}
}
```

An ÍTM translation of the thesis' abstract can be found [here](https://www.zora.uzh.ch/bitstreams/5395c2be-c8e9-4b29-9d72-ebdbd8a5163f/download)

2. **A condensed version of the thesis, presented at WLSP, 2nd Workshop on Sign Language Processing, at EMLNP 2026:**

```bibtex
@misc{ingimundarson2026isolatedsignlanguagerecognition,
      title={Isolated Sign Language Recognition for Icelandic Sign Language: Experiments in a Low-resource Setting}, 
      author={Finnur Ágúst Ingimundarson and Guðný Björk Þorvaldsdóttir and Mathias Müller and Sarah Ebling},
      year={2026},
      eprint={2609.25862},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.25862},
}
```

## License

The data is issed under a CC-BY-NC license that bans commercial use.