BullingerDB: A Dataset for Handwritten Text Recognition and Writer Retrieval

2026-05-28 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Advanced, quick

Summary

BullingerDB is a new large-scale benchmark dataset designed for historical document analysis, focusing on handwritten text recognition (HTR) and writer retrieval. Derived from the correspondence of Heinrich Bullinger (1504-1575), the dataset encompasses 20,898 pages and 499,222 text lines from 796 writers over six decades. It features significant stylistic variation, multilingual content primarily in Latin and Early New High German, and includes meta-information like writer identity and time. Evaluations on BullingerDB show TrOCR achieving a Character Error Rate (CER) of 9.1% for text recognition. For writer retrieval, a new temporal nDCG metric was introduced, with mAP scores reaching 78.3%, highlighting challenges from long-term stylistic changes. This dataset aims to establish a new benchmark for multilingual historical text recognition and temporally-aware writer analysis.

Key takeaway

For Machine Learning Engineers developing historical document analysis systems, BullingerDB offers a critical benchmark. You should consider integrating this large, multilingual dataset to train and evaluate your handwritten text recognition and writer retrieval models, especially when dealing with stylistic variation over time. Apply the introduced temporal nDCG metric to assess your writer retrieval solutions more accurately, ensuring they account for chronological changes in handwriting.

Key insights

BullingerDB is a large, multilingual dataset for historical HTR and writer retrieval, introducing temporal metrics and highlighting challenges from long-term stylistic variation.

Principles

• Historical HTR benefits from multilingual, time-aware data.
• Long-term stylistic shifts complicate writer retrieval.
• Temporal metrics are vital for historical writer analysis.

Method

The study introduces BullingerDB, a dataset for historical document analysis, and proposes a temporal nDCG metric to assess time-aware writer retrieval performance, complementing standard mAP scores.

In practice

• Train HTR models on multilingual historical texts.
• Evaluate writer retrieval with temporal nDCG.
• Develop models robust to long-term style changes.

Topics

• BullingerDB
• Handwritten Text Recognition
• Writer Retrieval
• Historical Document Analysis
• Multilingual Text
• Temporal Metrics

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Curation of a Palaeohispanic Dataset for Machine Learning — A new structured dataset enables machine learning analysis of under-resourced Palaeohispanic languages.
• RARE: Redundancy-Aware Retrieval Evaluation Framework for High-Similarity Corpora — Existing RAG benchmarks misrepresent real-world performance due to high document redundancy and similarity in enterprise corpora.
• Automatic Metrical Scansion of Galician Poetry: First Results — A new system provides automatic metrical scansion for Galician poetry with 88% accuracy.
• Tokenization vs. Augmentation: A Systematic Study of Writer Variance in IMU-Based Online Handwriting Recognition — Optimal OnHWR strategies depend on whether the system addresses inter-writer style variability or intra-writer data sparsity.
• SomaliWeb v1: A Quality-Filtered Somali Web Corpus with a Matched Tokenizer and a Public Language-Identification Benchmark — SomaliWeb v1 provides a high-quality Somali corpus, tokenizer, and language-ID benchmark, addressing critical resource gaps.
• CLEF HIPE-2026: Evaluating Accurate and Efficient Person-Place Relation Extraction from Multilingual Historical Texts — HIPE-2026 evaluates person-place relation extraction in historical texts, emphasizing accuracy, efficiency, and generalization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.