| # Polish Language Document Layout Detection — Test A |
|
|
| PolEval 2026 task: find the layout regions on scanned Polish document pages and label each with one of 11 |
| categories: Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text, |
| Title. Page images and the task description: https://github.com/poleval/2026-document-layout |
|
|
| The inputs list one page per line: |
|
|
| - `id`: the page identifier a submission must use. |
| - `image_id`: the page's COCO image id. |
| - `file_name`: the page image in the organisers' repository, relative to `poleval_doclaynet_coco_test_a/`. |
| - `width`, `height`: the coordinate frame boxes are given in. |
|
|
| Submissions are one JSONL line per page, `{"id": "<page id>", "output": [{"category", "bbox": [x, y, w, h], |
| "score"}]}`. They are scored by COCO mAP@[.50:.95], the mean of the 11 per-category APs. |
|
|
| The data is converted from the organisers' COCO files without changing any value. 37 pages are listed as 1×1, with |
| fractional box coordinates, exactly as the organisers published them. |
|
|
