ubiMap-l: A Benchmark for Crowdsourced Thematic Map Layout Retrieval and Embedding-based Analysis

The ubiMap dataset is comprised of 3,530 map images collected from the Bing image search service (1,730 maps) and Geo-Journal (1,800 maps). Each image has been manually labeled with 22 types of map elements, including their boundary shapes and category properties, resulting in an average of 5.92 elements per map. ubiMap-l is built uopon ubiMap by removing maps that contained only one element, which results a total of 3,515 maps for map layout retrieval test. We first opensourced 703 maps in ubiMap-l that we used for testing our map layout representation learning framework, MapLayNet. Besides 703 map images and their layout label data, embedding of MapLayNet and its baseline model is provided along with the python codes for embedding visualizaiton. Please cite the paper if you use the dataset. Yang, J., Chen, C., Jia, F., Xie, X., Fang, L., Wang, G., &amp; Meng, L. (2025). MapLayNet: map layout representation learning using weakly supervised structure-aware graph neural networks. <i>Cartography and Geographic Information Science</i>, 1–22. https://doi.org/10.1080/15230406.2025.2533316