Datasets:
dataset_info:
features:
- name: manuscript
dtype: string
- name: archive
dtype: string
- name: page
dtype: string
- name: segmentation_width
dtype: int32
- name: segmentation_height
dtype: int32
- name: segmentation_map
dtype: image
- name: source_image_credit
dtype: string
splits:
- name: train
num_bytes: 1494122305
num_examples: 49994
download_size: 1066351649
dataset_size: 1494122305
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
pretty_name: Randall Predicted Segmentations
license_name: predicted-annotations-license-unspecified-source-images-omitted
tags:
- image
- medieval-manuscripts
- semantic-segmentation
- model-predictions
license: other
task_categories:
- image-segmentation
Randall Predicted Segmentations
This dataset contains 49,994 predicted semantic segmentation maps for 148 medieval manuscripts indexed in Lilian Randall's Images in the Margins of Gothic Manuscripts (1966). Each predicted segmentation map is linked to manuscript, archive, and folio/page metadata.
These segmentations were generated by a model based on a multi-class adaptation of the MapSAM model proposed in Xia et al. (2025). Further details of the model training can be found in ADD IN.
Fields
manuscript: normalized public manuscript identifier.archive: holding institution or image provider.page: folio or page label from the manuscript inventory.segmentation_widthandsegmentation_height: prediction dimensions.segmentation_map: embedded RGB PNG prediction using the colors documented inid2label.json.source_image_credit: page-specific credit for the source image from which the prediction was derived.
The dataset does not contain source manuscript images.
Labels
| ID | Label |
|---|---|
| 0 | unlabeled/background |
| 1 | line filler |
| 2 | marginal images |
| 3 | painted initials |
| 4 | miniatures |
| 5 | borders |
| 6 | pen flourishing |
| 7 | pen flourished texts |
Attribution and rights
source_image_credit records provenance for an omitted source image; it is not
a license and does not mean attribution alone authorizes image reuse. The
represented institutions have mixed image-reuse terms, including open,
conditional, permission-based, and unresolved records. Generic source
identifiers remain explicitly generic rather than being presented as verified
shelfmarks. Users should review current source-record terms and determine
whether their intended use is permitted.
The predictions are model outputs rather than human-reviewed ground truth.
Citation
ADD IN