README.md

---
annotations_creators:
- crowd-generated
language:
- de
- en
- fr
license: cc-by-sa-4.0
task_categories:
- object-detection
- image-classification
pretty_name: 'ARTigo: Social Image Tagging'
tags:
- lam
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: hash_id
    dtype: string
  - name: titles
    struct:
    - name: id
      list: int64
    - name: name
      list: string
  - name: creators
    struct:
    - name: id
      list: int64
    - name: name
      list: string
  - name: location
    dtype: string
  - name: institution
    dtype: string
  - name: source
    struct:
    - name: id
      dtype: int64
    - name: name
      dtype: string
    - name: url
      dtype: string
  - name: path
    dtype: string
  - name: tags
    struct:
    - name: id
      list: int64
    - name: name
      list: string
    - name: language
      list: string
    - name: count
      list: int64
    - name: regions
      list:
      - name: x
        list: float64
      - name: y
        list: float64
      - name: width
        list: float64
      - name: height
        list: float64
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 5905082165
    num_examples: 60633
  download_size: 11451421751
  dataset_size: 5905082165
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# ARTigo: Social Image Tagging

60,633 digital reproductions of artworks with crowdsourced tags, from [ARTigo](https://www.artigo.org/) — a citizen-science project run since 2010 by the Institute of Art History and the Institute of Informatics at LMU Munich. Players are shown an image and type tags against a clock, scoring when a tag matches one their anonymous opponent gives or one recorded in an earlier session. The aggregate of those game rounds is this dataset. Built from the [v1.5 Zenodo deposit](https://doi.org/10.5281/zenodo.8202331) (1 August 2023), whose `data.jsonl` holds 60,981 records; the deposit is refreshed monthly, so current versions on Zenodo are larger than this build.

Measured on that deposit: 270,024 distinct tags, 4.9M tag–image pairs, and just under 10M individual taggings, at a median of 61 distinct tags per image. Images come from nine sources, named in `source.name`: Artemis (26,402), Koeln (21,055), Rijksmuseum (5,633), Mkg-Hamburg (3,965), Kunsthalle-Karlsruhe (2,145), Amherst-College (1,365), and small contributions from Albertina, Stadt-Land-Bild and Streetart.

## Fields

| field | notes |
| --- | --- |
| `id`, `hash_id` | integer id; content hash used as the image filename |
| `titles`, `creators` | lists of `{id, name}`; 8,769 distinct creator names |
| `location`, `institution` | holding place and institution, where known |
| `source` | `{id, name, url}` — one of the nine collections above |
| `path` | URL of the image on the ARTigo API |
| `tags` | list of `{id, name, language, count, regions}` |
| `image` | the reproduction itself |

Metadata coverage is partial: `creators` is empty on 12,182 records (20%), `institution` on 13,687 (22%), `location` on 17,795 (29%). Tags are unedited — the depositors deliberately left player spelling errors uncorrected rather than risk mangling specialist art-historical vocabulary.

```python
from datasets import load_dataset

ds = load_dataset("biglam/artigo", split="train")
row = ds[0]
row["image"]                 # the artwork
row["tags"]["name"]          # tags are stored column-wise, not as a list of dicts
row["tags"]["language"]      # tag language, mostly de
```

## The tags are overwhelmingly German

By tag–image pair: German 4,378,382 (89%), English 472,028 (9.6%), French 68,557 (1.4%). Anything trained on this without filtering learns a German tag vocabulary.

`language` is also a weaker signal than it looks. It records which language version of ARTigo the tag was collected on, not the language of the tag: the data paper notes that "English tags created in the German version of ARTigo are labelled as German".

## `regions` is almost always empty

Visual annotation — tagging a rectangle rather than the whole image — only started in 2022. In this deposit 521 records out of 60,981 (0.9%) carry any region at all, across 1,935 tag–image pairs. The `object-detection` task tag in the metadata is not supported at usable scale; in practice this is an image-tagging dataset. Where regions exist, the data paper describes them as normalised bounding boxes given by the x and y of the bottom-left corner plus width and height.

## Use

- Training or evaluating image tagging and text-to-image retrieval on art-historical images, in German.
- `count` is a per-image agreement signal — head tags were given by many players, the long tail by one. It is the obvious weight or threshold for filtering noise.
- Comparing player vocabulary against curatorial vocabulary, using `titles` and `creators` as the institutional description of the same object.

## Licence

The Zenodo record is CC BY-SA 4.0. That covers the deposit as a whole, including the aggregated annotations; the record does not break out rights per source institution for the reproductions themselves, which come from nine third-party collections. Check the source before redistributing images, and note that share-alike propagates to derivatives.

## Citation

```
@dataset{bry_et_al_artigo,
  author       = {Bry, François and
                  Kohle, Hubertus and
                  Krefeld, Thomas and
                  Riepl, Christian and
                  Schneider, Stefanie and
                  Schön, Gerhard and
                  Schulz, Klaus},
  title        = {{ARTigo}: Social Image Tagging (Aggregated Data)},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.8202331},
  url          = {https://doi.org/10.5281/zenodo.8202331}}
```

Data paper: Schneider, Stefanie (2024). "ARTigo: Data from Social Tagging with Art-historical Images", *Journal of Open Humanities Data* 10, [10.5334/johd.247](https://doi.org/10.5334/johd.247). Code: [arthist-lmu/artigo](https://github.com/arthist-lmu/artigo).