Datasetpublic

wellesley-easel/medieval_decoration_summaries

Medieval Decoration Summaries This dataset contains per-page pixel summaries and semantic segmentation maps for 60,411 pages from 175 medieval manuscripts held by 29 archives and indexed in Lilian Randall's Images in the Margins of Gothic Manuscripts (1966). The segmentation maps were produced by a computer vision pipeline with foreground mIoU of 60%, and therefore should not be treated as ground truth. Each row represents one manuscript page. Fields… See the full description on the dataset page: https://huggingface.co/datasets/wellesley-easel/medieval_decoration_summaries.

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes61downloads

Dataset Card

Medieval Decoration Summaries

This dataset contains per-page pixel summaries and semantic segmentation maps for 60,411 pages from 175 medieval manuscripts held by 29 archives and indexed in Lilian Randall's Images in the Margins of Gothic Manuscripts (1966).

The segmentation maps were produced by a computer vision pipeline with foreground mIoU of 60%, and therefore should not be treated as ground truth.

Each row represents one manuscript page.

Fields

—manuscript: normalized dataset manuscript identifier
—archive: holding archive
—page: folio/page label
—image_width, image_height, total_pixels: segmentation-map dimensions and total pixel count
—*_pixel_count: exact number of pixels assigned to each semantic category
—*_percent: percentage of all image pixels assigned to each semantic category
—segmentation_map: the embedded RGB semantic segmentation PNG

The semantic categories are:

—background
—line_filler
—marginal_images
—painted_initials
—miniatures
—borders
—pen_flourishing
—pen_flourished_texts

The class colors encoded in segmentation_map are:

Category	RGB
background	0, 0, 0
line filler	255, 0, 0
marginal images	0, 255, 0
painted initials	0, 0, 255
miniatures	255, 255, 0
borders	255, 0, 255
pen flourishing	0, 255, 255
pen flourished texts	255, 128, 0

The class pixel counts sum to total_pixels in every row. Percentage values are stored to six decimal places and equal 100 * class_pixel_count / total_pixels.

Model Reliability

The segmentation maps were produced by a computer vision pipeline with foreground mIoU of 60%, and therefore should not be treated as ground truth.

Category	Balanced F1
background	0.98
line filler	0.82
marginal images	0.68
painted initials	0.72
miniatures	0.79
borders	0.76
pen flourishing	0.76
pen flourished texts	0.70

Privacy and source paths

No local filenames, source-image paths, segmentation paths, or other filesystem locations are included. Segmentation PNGs are stored as embedded image bytes.

Archive and reuse note

The archive value records the archive/provider assignment used by the source project. It is descriptive metadata and does not assert a uniform image reuse license. This release contains model-produced segmentation maps and aggregate pixel counts, not the source manuscript photographs. Users should consult the holding archive's current record and terms for any use involving source images.