ICDAR 2019 cBAD: 3,021 Annotated Historical Document Images for Baseline Detection

3,021 annotated document page images form the training, evaluation, and test sets for the ICDAR 2019 Competition on Baseline Detection (cBAD). The dataset was created by Markus Diem of TU Wien and consists of real-world images collected from seven European archives, with all baselines manually annotated. The training and evaluation sets contain PAGE XML files with annotated text regions and baselines.

Use Cases

• Training baseline detection models based on manually annotated baselines in document images.
• Evaluating document analysis algorithms using the provided competition test set.
• Researching historical document layout and text region segmentation based on PAGE XML annotations.
• Benchmarking performance against the ICDAR 2019 competition results.

Strengths

• 3,021 annotated document page images provide a substantial corpus for model training and evaluation.
• All baselines were manually annotated, which likely ensures high-quality ground truth.
• Data is sourced from seven European archives, offering real-world variety in document styles.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.
• The ground truth for the test set was published after a competition deadline in May 2019, which may affect immediate reproducibility for that period.

Provenance

Source

Markus Diem, TU Wien

Collection Method

Images collected from seven European archives; baselines manually annotated.

Geography

European archives