A Large Labeled Dataset for Complex Layout Analysis of Medieval Manuscripts

This dataset is designed for historical document layout analysis and semantic segmentation. It contains 1,800 pages of medieval manuscript images with corresponding ground truth annotations. The annotations are generated using a semi-automatic iterative process combining automatic segmentation and manual correction. The dataset is split into 1,200 training images, 300 validation images, and 300 test images. Each image is paired with a pixel-level annotation belonging to six non-overlapping classes: background, decoration, filler, text, main body, and drop caps. The dataset is intended for research in document analysis, computer vision, and deep learning.