Historical City Maps for Semantic Segmentation

Rémi Petitpierre's dataset from École Polytechnique Fédérale de Lausanne contains 635 annotated image patches from historical city maps. It is designed for semantic segmentation into five classes: building blocks, non-built areas, water, road network, and background frame. The data includes 330 patches from Paris and 305 from a balanced corpus of city maps representing 90 countries worldwide.

Use Cases

• Training segmentation models to identify building blocks in historical maps.
• Analyzing the spatial distribution of road networks across different global cities.
• Comparing urban features like water bodies between Paris and other international cities.

Strengths

• Contains 635 manually annotated image patches, providing a labeled foundation for model training.
• Offers geographic diversity with samples from Paris and a corpus representing 90 countries.
• Defines a clear, structured annotation ontology with five distinct semantic classes.

Limitations

• The dataset is relatively small with 635 patches, which may limit model generalization.
• Key metadata such as column names, file size, and specific update dates are not provided in the sources.
• The temporal range of the historical maps is not specified, limiting time-based analysis.

Provenance

Source

Rémi Petitpierre, École Polytechnique Fédérale de Lausanne

Collection Method

Patches were taken from historical city maps; detailed methodology is described in an associated README file.

Geography

Paris, France, and a corpus of cities from 90 countries worldwide.