Semantic Segmentation Map Dataset (Semap)

The Semantic Segmentation Map Dataset (Semap) contains 1,439 manually annotated map samples. Specifically, the dataset includes 356 image patches from the Historical City Maps Semantic Segmentation Dataset (HCMSSD, [1]), 78 samples extracted from 19th-century European cadastres [2–4], three from Paris city atlases [5], and 1,002 newly annotated samples, drawn from the Aggregated Dataset on the History of Cartography (ADHOC Images, [6]). Additionally, it comprises 12,122 synthetically generated image samples and corresponding labels. Both datasets are part of R. Petitpierre's PhD thesis [7], and related paper [8]. Extensive details on annotation and synthetic generation procedures are provided in that work. Organization of the data The data is organized as shown below. The file structure differentiates real annotated images from synthetic data. project_root/ ├── images/ │ ├── real/ │ │ ├── berkeleylib_000004-0_2662_1421.jpg │ │ └── bnf_006136-1_1060_1608.jpg │ └── synthetic/ │ ├── 2_output_1.jpg │ └── 6_output_3.jpg ├── labels/ │ ├── real/ │ │ ├── berkeleylib_000004-0_2662_1421.png │ │ └── bnf_006136-1_1060_1608.png │ └── synthetic/ │ ├── 2_output_1.png │ └── 6_output_3.png ├── partitions/ │ ├── train.txt │ ├── val.txt │ └── test.txt └── model/ ├── best_mIoU_iter_138828.pth ├── mask2former_swin-l-in22k-384x384-pre_8xb2-160k_ade20k-640x640.py └── training.log Label syntax The labels are stored as single-channel PNG images. The pixel values indicate the semantic class: 0 background1 contours2 built3 non-built4 water5 road network Partitions The data is divided into training, validation, and test sets. Each partition file lists the raw file names of the corresponding images and labels, one per line. Example: davidrumsey_030108-0_6817_567 bnf_001250-7_6345_1736 ... 14_output_214 12_output_591 Synthetic samples are listed at the end of each file to facilitate their exclusion when finetuning on real data. Model weights A Mask2Former model with a Swin-L backbone, trained as described in [7,8], is provided for convenience and reproducibility. The implementation used is distributed by OpenMMLab [9]. The repository includes: best_mIoU_iter_138828.pth — model weights (Pytorch format) mask2former_swin-l-in22k-384x384-pre_8xb2-160k_ade20k-640x640.py — exact training configuration training.log — training log Descriptive statistics Number of semantic classes: 5 + backgroundNumber of manually annotated image samples: 1,439Number of synthetically-generated samples: Image sample size: min: 768 × 768 pixels max: 1000 × 1000 pixels Use and Citation For any mention of this dataset, please cite : @misc{semap_petitpierre_2025, author = {Petitpierre, R{\'{e}}mi and Gomez Donoso, Damien and Kriesel, Ben}, title = {{Semantic Segmentation Map Dataset (Semap)}}, year = {2025}, publisher = {EPFL}, url = {https://doi.org/10.5281/zenodo.16164781}}@misc{generalizable_segmentation_petitpierre_2026, author = {Petitpierre, R{\'{e}}mi}, title = {{Generalizable Multiscale Segmentation of Heterogeneous Map Collections}}, year = {2026}, publisher = {arXiv}, doi = {10.48550/arXiv.2603.05037}}@phdthesis{studying_maps_petitpierre_2025, author = {Petitpierre, R{\'{e}}mi}, title = {{Studying Maps at Scale: A Digital Investigation of Cartography and the Evolution of Figuration}}, year = {2025}, school = {EPFL}} Corresponding author Rémi PETITPIERRE - [e-mail] - ORCID - Github - Scholar - ResearchGate Work ethics 80% of the data were annotated by RP. The remainder were annotated by DGD and BK, two master's students from EPFL, Switzerland. The students were paid for their work using public funding, and were offered the possibility to be associated with the publication of the data. License This project is licensed under the CC BY 4.0 License. See the license_images file for details about the respective reuse policy of digitized map images. Liability We do not assume any liability for the use of this dataset. Version notes 1.0.0 – Stable release1.0.1 – Minor correction to filename consistency1.0.2 – Added separate README file for ease of download References Petitpierre, R. (2021). Historical City Maps Semantic Segmentation Dataset. V1.0. https://doi.org/10.5281/zenodo.5513639 di Lenardo I, Barman R, Pardini F, et al. (2021) Une approche computationnelle du cadastre napoléonien de Venise. Humanités numériques 3. https://doi.org/10.4000/revuehn.1786 Petitpierre R, Rappo L and di Lenardo I (2023) Recartographier l’espace napoléonien. In: Humanistica 2023, Genève, Switzerland, June 2023. Géographie. Association francophone des humanités numériques. Available at: https://hal.science/hal-04109214 Li S, Cerioni A, Herny C, et al. (2024) Vectorization of historical cadastral plans from the 1850s in the Canton of Geneva. Geneva, Switzerland: Swiss Territorial Data Lab. Available at: https://tech.stdl.ch/PROJ-CADMAP/ Chazalon J, Carlinet E, Chen Y, et al. (2021) ICDAR 2021 Competition on Historical Map Segmentation. arXiv:2105.13265 [cs]. Petitpierre, R. (2025). Aggregated Database on the History of Cartography (ADHOC). V1.0.0. https://doi.org/10.5281/zenodo.16277852 Petitpierre R (2025) Studying Maps at Scale: A Digital Investigation of Cartography and the Evolution of Figuration. PhD thesis. École Polytechnique Fédérale de Lausanne. https://doi.org/10.5075/epfl-thesis-11559 Petitpierre R (2026) Generalizable Multiscale Segmentation of Heterogeneous Map Collections. https://doi.org/10.48550/arXiv.2603.05037 OpenMMLab (2025) MMSegmentation. v1.2.2. https://github.com/open-mmlab/mmsegmentation