cBAD: ICDAR 2017 Competition on Baseline Detection in Archival Documents

ScriptNet's cBAD dataset contains the training and test set for the ICDAR 2017 Competition on Baseline Detection in Archival Documents. It comprises 2035 annotated document page images collected from 9 different archives, split into two tracks: one for simple handwritten paragraphs and another for complex documents with tables, marginalia, and noise. The training data includes PAGE XML files with manually annotated text regions and baselines.

Use Cases

• Training baseline detection models based on manually annotated PAGE XML ground truth.
• Benchmarking text line segmentation algorithms on simple handwritten paragraphs.
• Developing robust baseline detectors for complex archival documents containing tables and marginalia.
• Evaluating model performance on skewed text lines up to 180 degrees.
• Cross-archive analysis of historical document layouts and handwriting styles.

Strengths

• Contains 2035 annotated document page images from 9 different archives, providing diversity.
• Features manually annotated baselines and text regions in PAGE XML format for precise ground truth.
• Includes two distinct tracks (simple and complex documents) for testing different algorithmic capabilities.

Limitations

• Sources disagree on total page counts (750 vs. 1200 vs. 2035), creating uncertainty about dataset size.
• Column names and specific file formats are not detailed in the provided metadata.
• The last update date is unknown, so the dataset's freshness cannot be verified.

Provenance

Source

Markus Diem, TU Wien

Collection Method

Manually annotated images collected from 9 different archives and document collections.