ICDAR 2025 Competition on Glyph Detection in 15th-Century European Printed Documents
This competition is focused on the detection of glyphs in early European printed documents from the 15th century. The primary aim is to build an extensive corpus of glyphs by accurately extracting a large number of characters, with an emphasis on high precision rather than complete coverage. We will provide dataset that contains multiple historical printed documents with varying image quality.
Example of some glyphs, and their variations inside a single document.
Tasks
There are two main tasks:
- Glyph detection and localization
For this first tasks, participants have to produce tight bounding boxes around the glyphs. - Glyph classification
For this second task, participants have to identify which glyph is contained by the bounding boxes produced in the first task.
Registration
Participation is open to anybody without prior registration. However, registered participants will be informed by e-mail of any update, such as the publication of the test data.
Registration can be made either by filling the following form: https://forms.gle/kiwTuNUtTNzfmUtRA
or by sending an e-mail to the organizers (see “Contact” below).
Tracks
As in our previous competition (ICDAR2024 Competition on Multi Font Group Recognition and OCR), two tracks are planned:
Track 1: provided data only
The participants can use only the provided data for model optimization and validation. Using networks pre-trained on ImageNet is allowed.
Track 2: data alchemist
Besides data from Track 1, participants can use any additional data they already have or can find, and use it in any way they want (e.g., for SSL-based pre-training).
Timeline
- Soon: publication of the training data
- Ongoing: registration is open.
- 30th of March: participants receive the test set (without ground truth), and a link to a competition platform where they can submit their results
- 20th of April: deadline for the participants to submit:
- Their results on the competition platform,
- A short description of their method, their team name, and names of team members to the e-mail addresses given in the “Contact” section below.
Data
To be uploaded soon. Registered participants will be informed by e-mail.
Evaluation
Two main metrics will be considered:
- The mean detection precision with IoU thresholds between 0.5 and 0.95 with step size of 0.05,
- The area under the receiver operating characteristic curve (AUROC)
The participants will be ranked based on the harmonic mean of these two measures.
Competition platform
The link to the competition platform will be sent by e-mail to registered participants on the 30th of March, and added here.
What to submit
Results on the test set must have the same structure as the training ground truth, i.e., a COCO .json file.
The method description, which has to be submitted to the organizers, should ideally be one to few paragraphs long. It should ideally contain:
- Team name as it should be published,
- Names and affiliation of the team members,
- Methodology, including how you trained your method,
- Necessary citations (e.g., if your approach is based on some previous work).
Contact
The competition organizers can be contacted by e-mail at the following addresses: X@fau.de, where X is “mathias.seuret” and “vincent.christlein”. Please send e-mails to both, and start the e-mail title with “Competition: “.
