视觉-语言推理分割的功能级建筑轮廓（BUFF）数据集

BUFF数据集是一个用于功能级建筑轮廓分割的基准数据集。它基于政府测绘档案，包含2022年深圳市约50万栋建筑的多边形轮廓与主导功能标签，并配对了0.5米分辨率的谷歌地球影像。数据经专家视觉核查，按国家标准划分为住宅、商业、教育等10个功能类别。该数据集共包含12,940张512×512的图像，按7:1:2比例划分为训练、验证和测试集，旨在为基于视觉与语言推理的建筑功能提取研究提供可靠基准，推动该领域的发展。The BUFF dataset is a benchmark dataset for functional-level building footprint segmentation. Based on government surveying and mapping archives, it contains polygonal footprints and dominant function labels of approximately 500,000 buildings in Shenzhen in 2022, paired with 0.5-meter-resolution Google Earth imagery. The data has undergone expert visual verification and is categorized into 10 functional categories including residential, commercial, educational and others in accordance with national standards. The dataset consists of 12,940 512×512 images, which are split into training, validation and test sets at a ratio of 7:1:2. It aims to provide a reliable benchmark for research on building function extraction based on visual and linguistic reasoning, and promote the development of this field.