Visual-Language Reasoning Segmentation of Function-level Building Footprint (BUFF) dataset

We developed a benchmark dataset for function-level building footprint segmentation, named the Building Footprint Function (BUFF) dataset. The building function samples in this dataset were collected from government survey records, which provide information on approximately 500,000 buildings in Shenzhen, China. It delineates every building in Shenzhen at 2022, in the form of polygon data, and records the name and function of each building, such as Huangpu Yayuan (a residential), Xinshi Village (an urban village), Shenzhen Yaohua Experimental Middle School (a school), the Shenzhen Stock Exchange (a business tower), and the Shenzhen Municipal Government.

Following the Chinese Standard of Land Use Classification, we reclassified the function property of each building polygon into 10 categories with specified definition, i.e., urban village, business, commercial, residential, factory, government, medical service, hotel, institution, education, and others. For example, business refers to the financial, internet or insurance office building that is tall and usually for working, while commercial refers to the shopping centers or sports centers, etc., that has large roof area. It worth noting that the building function defined in this study refers to the primary function of the building. For example, if the ground floor of a residential building includes a kindergarten (school) or shop (commercial), but the upper floors are residential, the primary function of the building is still a residential building in our annotation. This strategy helps ensure consistency and data quality in annotations while better reflecting the main attributes of the buildings.

To ensure the reliability of the function samples, we further overlap the building polygon onto the 0.5-meter resolution Google Earth imagery from 2022 and conduct a visual inspection of the samples. For buildings that are difficult to identify, we further leverage the Google Street View images for assistance. For inconsistent location, we manually filtered out or replenish polygons, ensuring a one-to-one correspondence between the buildings in the images and their polygon annotations. It is worth noting that there is currently no publicly available building-footprint-scale function dataset, as the annotation process is highly labor-intensive and difficult to ensure accuracy. By contrast, the BUFF dataset in this study is built upon government survey records from urban planning related departments, and has been cross-verified by experienced survey experts through visual interpretation, which can substantially ensure its reliability. 

After then, we converted the building polygon to raster format, and paired it with 0.5-meter resolution Google Earth images, to construct the Building Footprint Function (BUFF) dataset. In the end, we cropped a total of 12,940 images at 512 × 512 pixels, covering circa 500,000 buildings. The dataset was split into a 7:1:2 ratio, with 9056 images for training, 1304 for validation, and 2580 for testing.

The corresponding paper for this dataset can be assessed in https://doi.org/10.1016/j.scs.2025.106439.

The corresponding code for Visual-Language Reasoning Segmentation of Function-level Building Footprint (BUFF) dataset can be assessed in https://github.com/whheda/LaRSE.

We hope it can be used as a benchmark to promote further development of function-level building footprint extraction research.