icdar2021-mapseg

# ICDAR 2021 MapSeg — historical map segmentation Map sheets from nine atlases of the City of Paris produced between 1894 and 1937 by the city's Map Service (*Service du plan*), annotated for three segmentation tasks. Built for the [ICDAR 2021 Competition on Historical Map Segmentation](https://icdar21-mapseg.github.io/) and deposited on Zenodo ([10.5281/zenodo.4817662](https://doi.org/10.5281/zenodo.4817662)). This repository is a format conversion of that deposit — same images, same ground truth, repackaged so it loads with `datasets` instead of a 1.9 GB tarball. The atlases were drawn at 1/5000 scale using trigonometric triangulation, roughly 20 sheets per year, for urban management and planning. Sheets are large: up to 10632 × 7711 pixels. ## Three tasks, three configs The tasks have different target shapes, so they are separate configs rather than one dataset with a task column. | Config | Task | Target | train / val / test | |---|---|---|---| | `detect_building_blocks` | find building blocks within map content | binary mask | 1 / 1 / 3 | | `segment_map_area` | separate map content from title, legend, scale | binary mask | 26 / 6 / 95 | | `locate_graticule_intersections` | locate coordinate grid crossings | list of `[x, y]` points | 26 / 6 / 95 | `detect_building_blocks` has five images in total. It was the competition's flagship task, but it is not a training set in any ordinary sense — treat it as a benchmark, not something to fit on. The split sizes are inverted from the usual shape: 95 test sheets against 26 training sheets. That is how the competition was designed, and it is preserved here. ## Structure All three configs share `sheet_id` (the upstream three-digit identifier — `1NN` train, `2NN` validation, `3NN` test) and `image`. | Config | Additional fields | |---|---| | `detect_building_blocks` | `label` (binary GT mask), `area_mask` (valid region; discard predictions outside it) | | `segment_map_area` | `label` (binary GT mask) | | `locate_graticule_intersections` | `points`, a list of `[x, y]` float coordinates | Masks are mode `L` with values strictly `{0, 255}`, and match their image dimensions exactly. ## Caveats **Task 3 shares Task 2's images.** The upstream archive does not duplicate them — the Task 3 folder contains only ground-truth CSVs. The conversion pulls the JPEGs across so the config is self-contained, which means the images appear twice across the repository and `locate_graticule_intersections` costs 1.77 GB of mostly redundant bytes. Load the config you need rather than the whole repo. **Points are float32.** The upstream CSVs give sub-pixel coordinates to one decimal place; float32 storage rounds them, so `2373.2` reads back as `2373.19995`. The error is around 5×10⁻⁵ px and irrelevant for evaluation, but exact string comparison against the source CSVs will not match. **Images are above Pillow's decompression-bomb threshold.** At ~8×10⁷ pixels a sheet exceeds the default `MAX_IMAGE_PIXELS` limit and Pillow will refuse to decode it. Set `PIL.Image.MAX_IMAGE_PIXELS = None` before loading, as in the snippet below. **Do not score with your own metric.** The competition's evaluation tools are released separately at [icdar21-mapseg-eval](https://github.com/icdar21-mapseg/icdar21-mapseg-eval), and published results used version 1.0.3. Numbers computed any other way are not comparable to the competition leaderboard. **One city, one map service, forty years.** Every sheet comes from the same cartographic tradition and the same drawing conventions. Models fitted here should not be assumed to transfer to other map series. ## Load ```python import PIL.Image PIL.Image.MAX_IMAGE_PIXELS = None # sheets run to ~8e7 px, over Pillow's default limit from datasets import load_dataset # stream — the test split alone is ~1.3 GB and sheets are very large ds = load_dataset("biglam/icdar2021-mapseg", "segment_map_area", split="test", streaming=True) ex = next(iter(ds)) ex["image"].size, ex["label"].size ``` ## Licence CC BY 4.0, following the upstream deposit. The underlying atlases are reproduced with permission granted by the City of Paris to the competition organisers. That permission covers this material as distributed; it is not a general release of the City of Paris map holdings. Original documents: *Atlas municipal des vingt arrondissements de Paris*, 1894–1937, Bibliothèque de l'Hôtel de Ville, City of Paris. ## Credit Data created by Joseph Chazalon, Edwin Carlinet and Thierry Géraud (EPITA Research and Development Laboratory), Yizi Chen, Julien Perret and Clément Mallet (Univ. Gustave Eiffel, IGN-ENSG, LaSTIG), and Bertrand Duménieu (LaDéHiS, CRH, EHESS). Partially funded by the French National Research Agency, project SoDuCo, grant [ANR-18-CE38-0013](https://anr.fr/Projet-ANR-18-CE38-0013). Images courtesy of the City of Paris. Converted and repackaged for the Hub by [Daniel van Strien](https://huggingface.co/davanstrien). ```bibtex @misc{chazalon.21.mapseg.data, author = {Chazalon, Joseph and Carlinet, Edwin and Chen, Yizi and Perret, Julien and Mallet, Cl{\'e}ment and Dum{\'e}nieu, Bertrand and G{\'e}raud, Thierry}, title = {{ICDAR 2021 Competition on Historical Map Segmentation --- Dataset}}, howpublished = {online dataset}, year = {2021}, doi = {10.5281/zenodo.4817662} } ```