Cross-Dataset Transfer in Open-Set Writer Identification: From Hand-Drawn Circles to Handwritten Pages

Official code, open-set data splits, and reproduction scripts for our WML 2026 paper.

Md Raihan, Thomas Gorges, Lukas Hüttner, Vincent Christlein Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany Accepted at the 6th ICDAR Workshop on Machine Learning (WML 2026) — Vienna, Austria, 3 September 2026.

Overview

Writer identification is usually studied as a closed-set retrieval problem. This work tackles the harder, more realistic open-set setting — query pages may come from writers never seen at enrollment — under a tight budget of only 3 enrollment pages per writer, and asks how well a single recipe transfers across datasets: from the hand-drawn circles of the ICDAR CircleID competition to full handwritten pages in CVL (modern) and Historical-WI (historical). CircleID poses writer identification from a single hand-drawn circle — a minimal, content-free primitive — a deliberately stringent starting point for testing whether one open-set recipe generalises to information-rich handwriting.

The method is a frozen DINOv2 ViT-B/14 backbone with LoRA, NetVLAD aggregation, and a multi-prototype ArcFace head with hard-negative mining, followed by one of three open-set rejection rules — (A) per-writer multi-prototype Mahalanobis, (B) cluster K-sweep, (C) joint train+test clustering. Every dataset is cast into the same three-group protocol of Known / Pseudo-unknown / Unknown writers.

Headline results

Handwritten pages — open-set Top-1, three-seed ensemble (paper Table 3):

Dataset	Best rejection rule	OS-Top1	Known-Top1	AUROC
CVL	Joint train+test clustering	74.67%	63.20%	85.3%
Historical-WI	Per-writer MP-Mahalanobis	54.13%	13.06%	57.3%

Hand-drawn circles — CircleID private leaderboard (paper Table 2): 65.17%, up from the 47.02% competition-day submission.

Repository structure

Folder	Contents
cvl-hwi-writerid-method/	The open-set writer-ID method for CVL and Historical-WI: DINOv2 + LoRA + NetVLAD + multi-prototype ArcFace, plus the three rejection rules. Training, embedding extraction, post-hoc refinement, and evaluation. See its README and REPRODUCE.
splits/	The open-set split protocol: deterministic builders (build_cvl_splits.py, build_hwi_splits.py), a convention-agnostic verifier (verify_splits.py), and the ready-to-use split manifests (cvl_splits/, hwi_splits/) that define the Known / Pseudo-unknown / Unknown pools.
circleid/	Reproduction archive for the CircleID (hand-drawn circles) half of the paper — writer identification from a single hand-drawn circle. Isolates the exact code behind each row of Table 2, tracing the private-leaderboard result from the 47.02% competition-day submission to the final 65.17% configuration (organised as precompetition/ and post_competition/ stages). See its README.

Quickstart

The datasets are not redistributed here (licensing) — download them from their official sources (links in the method README) and use the split manifests provided in splits/. Then:

cd cvl-hwi-writerid-method
pip install -r requirements.txt
# train + extract → refine → reject → evaluate; full commands in REPRODUCE.md

• Verify the open-set splits at any time: python splits/verify_splits.py splits/cvl_splits splits/hwi_splits
• Full step-by-step reproduction: cvl-hwi-writerid-method/REPRODUCE.md

Citation

If you use this code or the open-set CVL / Historical-WI protocols, please cite:

@inproceedings{raihan2026crossdataset,
  title     = {Cross-Dataset Transfer in Open-Set Writer Identification:
               From Hand-Drawn Circles to Handwritten Pages},
  author    = {Raihan, Md and Gorges, Thomas and H\"uttner, Lukas and Christlein, Vincent},
  booktitle = {6th ICDAR Workshop on Machine Learning (WML)},
  address   = {Vienna, Austria},
  year      = {2026}
}

A machine-readable CITATION.cff is also provided.

License

Released under the MIT License.