The DLA-RMR dataset: Annotated subset of RMR notebooks for CVC development

2026-07-15T15:55:18Z https://www.fdr.uni-hamburg.de/oai2d

oai:fdr.uni-hamburg.de:18089 2026-05-21T16:04:21Z user-uhh user-csmc

true 3.1 UHH.FDR 10.25592/uhhfdm.18089 Hussein Mohammed 0000-0001-5020-3592 Universität Hamburg Quang-Vinh Dang 0000-0002-6715-7112 Universität Hamburg The DLA-RMR dataset: Annotated subset of RMR notebooks for CVC development Universität Hamburg 2025 page detection word detection colour recognition recognition of writing implement visual navigation computational visual cataloguing 2025-09-05 en https://www.fdr.uni-hamburg.de/record/oai:fdr.uni-hamburg.de:18089 10.25592/uhhfdm.17809 10.25592/uhhfdm.17613 10.25592/uhhfdm.17615 10.25592/uhhfdm.17931 0.5 Creative Commons Attribution 4.0 International Open Access What’s new in this version: Some annotations were missing one of the visual attributes (orientation or writing implement). All missing attributes have been added in this version. This dataset is structured into four components, each serving a distinct role in the development of a document analysis system.  <ol> <li> Word-level annotations are provided in the file <code>word_annotations_for_cropped_images.json</code>. These annotations describe the images contained in the <code>cropped_images</code> folder. Each entry specifies the location of a word as a polygon, together with its orientation (horizontal, vertical, or tilted) and the type of writing implement used (ink or pencil). Additional metadata, such as bounding boxes and segmentation areas, is also included. </li> <li> Cropped images are stored in the <code>cropped_images</code> folder. This set comprises 50 images, each containing only the primary page extracted from the corresponding full notebook scans. </li> <li> Full images are located in the <code>full_images</code> folder. This collection also contains 50 items, representing the complete notebook scans in which the primary page appears alongside other material. </li> <li> Page-level annotations are contained in the <code>page_annotations</code> folder. These are provided in YOLO format, with a single class (<code>page</code>) defined in <code>classes.txt</code>. Each annotation file specifies the bounding box of the primary page within the corresponding image in the <code>full_images</code> folder. </li> </ol> Examples illustrate the annotation structure. In the JSON file, a typical word annotation records polygon coordinates, the attribute <code>"orientation": "horizontal"</code>, and <code>"writing_tool": "pencil"</code>. In the YOLO annotations, a sample entry such as <code>0 0.499023 0.500776 0.777344 0.816912</code> denotes the normalised coordinates of the primary page bounding box. Acknowledgement: The research for this work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy - EXC 2176 ‘Understanding Written Artefacts: Material, Interaction and Transmission in Manuscript Cultures’, project no. 390893796. The research was conducted within the scope of the Centre for the Study of Manuscript Cultures (CSMC) at Universität Hamburg. The images are taken from notebook pages of Rainer Maria Rilke, from the Deutsche Literaturarchiv Marbach (DLA), A:Rilke-Archiv Gernsbach.  We thank Hui Xu for her support in annotating the images.