BRENNEN

Industries / Media, arts and recreation / Museums

Digitizing

Photographing, scanning and describing objects, and putting the records online.

6

Milestones achieved, from 2011 to 2020.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2011

    Paintings at gigapixel size

    On 1 February 2011 Google launched the Art Project with 17 museums, including the National Gallery in London, the Metropolitan Museum of Art and the Uffizi, showing more than 1,000 artworks online and gallery tours made with Street View cameras. Each museum chose one work to be photographed at "gigapixel" size: the Frick Collection's release says each such image "contains up to 14 billion pixels", and the National Gallery says around 7 billion.

    nationalgallery.org.ukfrick.org

    Credit

    Google; 17 partner museums

  2. 2012

    Free downloads at the Rijksmuseum

    On 31 October 2012 the Rijksmuseum in Amsterdam launched Rijksstudio, where images of 125,000 works "can be freely downloaded, zoomed in on, shared, added to personal 'studios', or manipulated copyright-free".

    infodocket.com

    Credit

    Rijksmuseum

  3. 2015

    A shared standard for images

    On 16 June 2015 eleven institutions, among them the British Library, the University of Oxford, Stanford, Yale and the Bibliothèque nationale de France, formed the consortium behind the International Image Interoperability Framework (IIIF), "a set of open standards for delivering high-quality, attributed digital objects online at scale".

    iiif.io

    Credit

    the eleven founding members of the IIIF Consortium

  4. 2017

    The Met opens its images

    On 7 February 2017 the Metropolitan Museum of Art made 375,000 images of works in its collection free "to copy, remix, and distribute freely", by Hyperallergic's report.

    hyperallergic.com

    Credit

    Metropolitan Museum of Art

  5. 2020

    Smithsonian Open Access

    In February 2020 the Smithsonian removed copyright restrictions from 2.8 million 2D and 3D images and "nearly two centuries of data", using CC0, Creative Commons' tool for placing work in the public domain. The images come from all 19 Smithsonian museums, nine research centres, its libraries and archives, and the National Zoo.

    creativecommons.org

    Credit

    Smithsonian Institution

  6. 2020

    The Night Watch in close-up

    In May 2020 the Rijksmuseum put online a photograph of Rembrandt's The Night Watch of 44.8 gigapixels, stitched from 528 exposures, with 20 micrometres between pixels, close enough to see single brushstrokes.

    dutchnews.nl

    Credit

    Rijksmuseum

Watch for

  1. The whole British Museum online

    The British Museum says its Documentation and Digitisation Programme will make every object in its collection "accessible digitally" with records that "meet approved documentation standards", and estimates the work will take five years. It reports that records of "almost five million objects" can already be seen and downloaded, and that by 2025–26 90% of its Africa collection had been photographed and 70,000 coin and medal images added in the year.

    Achieved whenan outside auditor, or a follow-up independent review, confirms that every object in the collection has a record that meets the museum's standard.

    assets.publishing.service.gov.uk

    Claimed by

    British Museum, Annual Report and Accounts 2025–26, July 2026

Open

  1. 01

    Read the old registers

    A pipeline that reads scanned pages of a museum's handwritten and typed accession registers (the ledgers in which each object was entered on arrival), turns each entry into a structured record, and links it to the matching record in the collections database.

    The brief
    Credit · unclaimed
  2. 02

    Scoring drafted catalogue records

    A public test: take objects from open-access collections that already have full curator records, hide the records, have AI models draft them from the photographs and a short label, and score each field (object type, material, technique, date range, place, measurements) against the curators' version.

    The brief
    Credit · unclaimed

Open 01 · the brief

Read the old registers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A pipeline that reads scanned pages of a museum's handwritten and typed accession registers (the ledgers in which each object was entered on arrival), turns each entry into a structured record, and links it to the matching record in the collections database.

Why it is still open

In Birmingham Museums Trust's Documentation Detectives project, more than 3,400 volunteers typed up over 85,000 register records, dated 1912 to 2003, by hand on the Zooniverse platform; the last pages went up on 31 December 2025, and the next step is to import the transcriptions "exactly as the volunteers have typed them up". That finished set is a ready answer key, and no published test of machine reading against it was found.

birminghammuseums.org.ukzooniverse.org

Built with what exists

Handwriting recognition such as Transkribus, or an open vision-language model retrained on a sample of pages (see the Note "One GPU to fine-tune"); a document agent that splits entries into fields such as number, name, source and date (see the Note "The jobs agents repeat"); fuzzy matching on accession numbers to link entries to database records.

  1. First step

    Ask Birmingham Museums Trust for 200 scanned register pages with their volunteer transcriptions, and run an off-the-shelf handwriting model on them for a baseline.

  2. Done looks like

    On 5,000 register entries that volunteers have already typed up, the machine's text has a character error rate under 5% and gets at least 95% of accession numbers exactly right, and at least 90% of entries are linked to the right database record, as checked by Birmingham's staff.

Open 02 · the brief

Scoring drafted catalogue records

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test: take objects from open-access collections that already have full curator records, hide the records, have AI models draft them from the photographs and a short label, and score each field (object type, material, technique, date range, place, measurements) against the curators' version.

Why it is still open

Open collections such as the Smithsonian's 2.8 million CC0 images and data can serve as the answer key. The nearest published evaluation found, by Miguel Arana-Catania and Neil Jefferies at the University of Oxford's Bodleian Libraries in July 2026, compared GPT and Gemini models on 1,886 images from 87 dissertations, not museum objects, and found that "the plausibility of many of these variant results makes error detection difficult". No comparable test on museum object records was found.

arxiv.orgcreativecommons.org

Built with what exists

Open-access images and records from the Smithsonian, the Met and the Rijksmuseum; several commercial and open vision-language models, including small ones retrained on one graphics chip (see the Note "One GPU to fine-tune"); a scoring script with curators' rules for when two dates or materials count as the same.

  1. First step

    Download 500 public-domain object records with images from one museum's open data and write the scoring rules for five fields.

  2. Done looks like

    5,000 objects of at least five kinds (paintings, prints, coins, ceramics, textiles), with field-by-field accuracy published for at least four models; on 200 drafted records also judged blind by two curators, the curators' verdicts match the automatic score at least 90% of the time.