ICDAR 2026 Competition on Long-Term Handwriting Author Identification
The AnyScript Challenge
The first benchmark to evaluate writer identification systems across manuscripts separated by decades, incorporating the temporal dimension of handwriting evolution.
Introduction & Motivation
Traditional writer identification frameworks treat handwriting style as a static snapshot, overlooking the crucial dimension of temporal evolution. Handwriting naturally evolves due to aging, context, and changing writing materials, limiting the real-world applicability of current approaches.
The AnyScript competition introduces temporal transitivity to authorship identification, challenging systems to identify common authorship across manuscripts separated by decades. Our dataset from the Spanish Digital Library (Biblioteca Digital Hispánica) spans 500+ years of authentic autograph manuscripts.
Temporal Evolution
Account for progressive changes in writing style over time
Multi-Page Context
Process long contextual sequences efficiently
Historical Significance
18th-20th century literary and artistic works
Task Definition
The competition follows a query-by-example retrieval framework. Given a query document (page or book), systems must retrieve other documents written by the same author from the dataset.
Intra-Book Evaluation
Goal: Identify which books query pages belong to
Input: 1,000 individual pages excluded from training books
Output: Page-level retrieval rankings
Challenge: Match pages to their source books
Extra-Book Evaluation
Goal: Identify books by the same author across temporal spans
Input: 200 complete books excluded from training set
Output: Book-level retrieval rankings
Challenge: Long-term stylistic continuity across decades
Temporal Transitivity
Systems are expected to incorporate temporal reasoning. Since direct inference between temporally distant manuscripts (A ↔ C) is challenging, methods should leverage intermediate associations through temporally closer manuscripts (A ↔ B ↔ C).
Evaluation Metrics
Three complementary metrics evaluate retrieval performance:
Recall@K
Measures the proportion of relevant documents among the top K retrieved items (K = 1, 5, 10).
Mean Average Precision (mAP)
Summarizes ranking performance by averaging precision across all queries. Rewards systems that rank relevant items higher.
mAP = (1/Q) Σ AP(q)
Normalized Discounted Cumulative Gain (nDCG)
Uses normalized temporal distance as relevance score to promote historically coherent rankings.
nDCGq = DCGq / IDCGq
The AnyScript Dataset
Dataset Statistics
- 500,000+ images (colored + binarized)
- 2,286 books by 1,019 unique authors
- Temporal span: 500+ years (18th-20th centuries)
- Content: Novels, plays, music scores, literary works
- Language: Primarily Spanish
- Source: Spanish Digital Library autograph copies
Data Characteristics
The dataset contains verified autograph manuscripts - original works produced by their creators. Documents labeled as autographs but dating after authors' deaths have been removed to ensure authenticity.
Temporal coverage focuses on the late 19th and early 20th centuries, with some authors having continuous book representation spanning 4 decades, enabling study of long-term stylistic evolution.
Dataset Examples
The AnyScript dataset contains diverse historical documents including personal letters, theater scripts, music scores, and literary works spanning over 500 years. Below are representative examples of the types of manuscripts included in our collection.
Personal letter from 19th century
Historical handwritten correspondence demonstrating the evolution of personal writing styles and formal letter composition.
Music score manuscript
Handwritten musical notation showcasing the unique challenges of identifying authorship in specialized symbolic writing systems.
Literary work
Antique handwritten document illustrating the temporal evolution of handwriting styles across different literary contexts.
Temporal Challenge
These examples demonstrate the core challenge of our competition: identifying common authorship across documents that may span decades, accounting for natural evolution in handwriting style, different contexts (personal vs. formal), and varying writing materials.
Submission Format
Both evaluation tracks require CSV file submissions with the following structure:
Required CSV Columns
query_document_id,retrieved_document_id,similarity_score
Intra-Book Track
Submit page-to-page similarity scores for 1,000 query pages against all training pages.
query_page_001,train_page_1234,0.89query_page_001,train_page_5678,0.76
Extra-Book Track
Submit book-to-book similarity scores for 200 query books against all training books.
query_book_001,train_book_045,0.92query_book_001,train_book_128,0.68
Downloads
| Dataset Component | Description | Status | Download |
|---|---|---|---|
| Training Set | Full-resolution colored images from training books | Coming Soon | |
| Training Set (Binarized) | Binarized versions of training images | Coming Soon | |
| Training Labels | Author and date annotations for training data | Coming Soon | |
| Test Queries | Unlabeled test set for evaluation | Available March 2026 |
Competition Schedule
Competition Launch
Competition website and validation set live
Test Set Release
Test set available on evaluation server
Results Deadline
Competition results submission deadline
Report Submission
Initial competition report submission due
Camera-Ready Papers
Camera-ready paper submission due
Winners Announced
Competition winners announcement
Organizing Team
Adrià Molina Rodríguez
PhD Candidate
amolina@cvc.uab.cat
Cultural heritage, data governance, responsible AI, retrieval systems, data visualization, graphs
Carlos Boned Riera
PhD Student
cboned@cvc.uab.cat
Representation learning, graph-based methods, historical manuscripts, retrieval systems
Pau Torras
PhD Student
ptorras@cvc.uab.cat
Historical handwritten music scores recognition, digital humanities
Oriol Ramos Terrades
Associate Professor & Department Head
oriolrt@cvc.uab.cat
Probabilistic graphical models, document image analysis, symbol retrieval
Josep Lladós
Associate Professor & CVC Director
josep@cvc.uab.cat
Document analysis, structural pattern recognition, computer vision
Contact Information
Institution
Computer Vision Center and Universitat Autònoma de Barcelona
General Inquiries
Email: amolina@cvc.uab.cat