ADV-FORMS

You have reached the ADV-FORMS dataset... We present ADV-FORMS, an openly accessible dataset of historical, form-based documents with layout annotations and transcriptions of handwritten and printed text. The collection comprises 1,200 questionnaires from the early 1930s spanning multiple layout templates, completed with handwritten, typewritten, and stamped entries, alongside non-textual elements such as underlines, drawings, and selection marks. The handwriting includes both German Kurrent and Latin scripts. ADV-FORMS provides manual annotations with bounding boxes for 17 recurring, visually distinct layout categories, as well as manual transcriptions of the handwritten content and all typewritten or stamped text. Overview The repository contains the following elements: ADV-FORMS_images.zip: 1200 scanned questionnaires in JPG format. ADV-FORMS_ls.json: annotations and transcriptions for each image in a single JSON file following the Label Studio schema. split.zip: suggested split for machine learning applications into training, validation, and test sets (identical to the split used in the ICDAR 2026 paper). ADV-FORMS_annotation_guidelines.pdf: guidelines used during dataset assembly; can be used to understand the layout categories. JSON-Schema Documentation Top-level structure Each top-level item represents a single image task and includes: data: image metadata, including the image path. annotations: a list of annotation records. created_at: task creation timestamp. updated_at: task update timestamp. total_annotations: number of annotation records. Example { "data": { "image": "ADV-FORMS_images/q003_003_1881_R.jpg" }, "annotations": [ { "id": 86, "completed_by": 3, "result": [ ... ] } ], "created_at": "2026-06-23T07:45:00Z", "updated_at": "2026-06-23T07:45:00Z", "total_annotations": 1 } annotations Each annotations entry contains a result array with multiple Label Studio result items. In this dataset, the two primary item types are rectanglelabels and textarea, which typically appear together for the same labeled region. rectanglelabels rectanglelabels results define labeled bounding boxes on the image. A typical rectanglelabels item looks like: { "id": "2726ba22-eec2-4121-bcda-8b592b6d7cda", "type": "rectanglelabels", "from_name": "label", "to_name": "image", "value": { "x": 47.67, "y": 3.20, "width": 7.37, "height": 5.95, "rotation": 0, "rectanglelabels": ["location_ID_part1"] } } Important value fields: x, y: top-left corner coordinates as percentages of the image size. width, height: box size as percentages of the image dimensions. rotation: box rotation, usually 0. rectanglelabels: class label array for the region. rectanglelabels entries identify where on the image a labeled field appears and what label category it belongs to. textarea textarea results contain transcribed text for a labeled region. A typical textarea item looks like: { "id": "2726ba22-eec2-4121-bcda-8b592b6d7cda", "type": "textarea", "from_name": "transcription", "to_name": "image", "value": { "text": ["153"], "x": 47.67, "y": 3.20, "width": 7.37, "height": 5.95, "rotation": 0 } } Important value fields: text: array of transcribed strings. x, y, width, height: optional geometry, matching the related rectanglelabels box. textarea entries capture the transcribed text content for each labeled region. How rectanglelabels and textarea work together In this dataset, annotation results are often paired by the same id value: rectanglelabels marks the labeled region and class. textarea provides the text transcription for that same region. Pairing by id allows combining region labels with transcription content. Notes for users Coordinates are expressed as percentages, not pixels. Use original_width and original_height when converting relative coordinates to absolute pixels. Annotation records may include other Label Studio metadata, but rectanglelabels and textarea are the primary fields for extraction. Purpose This README is intended for users who need to: understand the cleaned Label Studio schema used by ADV-FORMS_ls.json extract bounding boxes and transcriptions convert annotations to other formats (PAGE XML, COCO, VOC, etc.) Direct Label Studio Import The JSON schema of the dataset can be directly imported into Label Studio. Adjust the image paths in the JSON to match your setup and copy the labeling interface template below into your Label Studio project. Adapting image paths for your Label Studio setup The data.image field in this JSON contains ADV-FORMS_images/ .jpg. When importing the data into Label Studio, you must adjust this path to match your concrete setup. Label Studio supports multiple storage backends. Depending on your configuration, you may need to: Local file storage: use relative or absolute filesystem paths. S3 or cloud storage: use URLs (e.g., https://s3.example.com/bucket/...). Label Studio local files: use the special /data/local-files/?d=path format. Example adjustments: Original: ADV-FORMS_images/q003_003_1881_R.jpg For S3: https://my-bucket.s3.amazonaws.com/ADV-FORMS_images/q003_003_1881_R.jpg For Label Studio local files: /data/local-files/?d=ADV-FORMS/images/q003_003_1881_R.jpg For absolute filesystem path: /home/user/data/ADV-FORMS_images/q003_003_1881_R.jpg Before importing, adjust all data.image paths in your JSON to match your storage setup. Labeling interface template Use the following Label Studio task template to set up the labeling interface: References For more information on Label Studio and annotation workflows, refer to: Importing Annotations: Label Studio — Import Predictions and Annotations — Learn how to import pre-existing annotations into a Label Studio project. Defining Labeling Interface: Label Studio — Labeling Configuration — Reference for creating custom task templates with XML-based labeling controls. Image Configuration: Label Studio — Image Control — Details on configuring image display and storage in Label Studio tasks. Bounding Boxes: Label Studio — Rectangle Labels — Documentation on RectangleLabels for region annotation. Text Areas: Label Studio — Text Areas — Reference for text transcription and input controls.