Aerial imagery dataset

原始航拍数据来自新西兰土地信息服务网站。我们手动编辑了基督城的建筑矢量数据，大约有22,000座独立建筑。图像的原始地面分辨率为0.075m。由于航空影像的大小 (约25G)。
我们还对航拍图像的大部分部分 (包括18，7000座建筑物) 进行了降采样，以0.3m的地面分辨率，并将其裁剪为具有512 × 512像素的8,189图块。shapefile也被光栅化。即用样本分为三部分: 训练集 (130,500建筑物) 、验证集 (14,500建筑物) 和测试集 (42,000建筑物)。根据裁剪的顺序，您可以从图块中恢复原始图像。训练区域 (蓝色框): 64 × 74 = 4736瓷砖 (宽 × 高) 评估区域 (橙色框): 14 × 74 = 1036瓷砖最佳区域 (两个红色框): 17 × 42 23 × 74 = 2416瓷砖 (part1 part2)

The raw aerial imagery was sourced from the Land Information New Zealand (LINZ) website. We manually curated the building vector data for Christchurch, which contains approximately 22,000 individual buildings. The original ground sampling distance (GSD) of the imagery is 0.075 meters, and the total size of the raw aerial dataset is approximately 25 GB. We also downsampled most segments of the aerial imagery (covering 18,7000 buildings) to a GSD of 0.3 meters, and cropped them into 8,189 tiles with a resolution of 512 × 512 pixels. The corresponding shapefiles were also rasterized. The dataset is split into three subsets: training set (130,500 buildings), validation set (14,500 buildings), and test set (42,000 buildings). The original image can be reconstructed from the tiles based on the cropping order. The spatial split details are as follows:
- Training area (blue bounding box): 64 × 74 = 4,736 tiles (width × height)
- Evaluation area (orange bounding box): 14 × 74 = 1,036 tiles
- Optimal area (two red bounding boxes): 17 × 42 + 23 × 74 = 2,416 tiles (part 1, part 2)

搜集汇总

背景与挑战

背景概述

该数据集基于新西兰土地信息服务的原始航拍数据构建，由武汉大学于2018年发布，包含约22,000座建筑的矢量标注。图像经过降采样至0.3米分辨率，并裁剪为8,189个512×512像素的图块，划分为训练、验证和测试集，用于建筑识别任务。