GeoMapSet: A POI-driven multimodal dataset for map image understanding

Here we present GeoMapSet, a POI-driven, automatically constructed multimodal dataset designed to bridge this gap. Using OpenStreetMap as a unified base source, we crawled and stitched 17,835 map images (768×768 or 1024×1024 pixels) and retrieved associated point-of-interest (POI) data via the Overpass API. A two-level category hierarchy (22 fine-grained icon classes mapped to 6 functional broad classes) and a parameterized three-tier description system (quantity-level, summary-level, and feature-level) were designed to convert discrete POI records into coherent textual descriptions in both English and Chinese. It is expected to serve as a foundational resource for geospatial multimodal learning, map interpretation, and the development of generalizable geographic artificial intelligence.