ORBITaL-Net Training Library for Building Extraction

The Oak Ridge Building Image and TrAining Label Net (ORBITaL-Net), is a training dataset designed to enable the learning of building detection deep learning models. It consists of over 128,000 individual samples drawn from thousands of separate high resolution satellite images (average resolution 0.47 m). Each sample is a 500x500 pixel patch with accompanying binary label raster with each pixel hand-annotated by expertly trained image analysts as either building or non-building. This dataset has a large degree of geographic and semantic variety, including samples from North America, South America, Africa, the Middle East, and Asia, as well as samples that include a variety of viewing angles, vernacular architecture styles, LU/LC contexts, and atmospheric conditions.

搜集汇总

背景与挑战

背景概述

ORBITaL-Net是一个用于建筑物提取的深度学习训练数据集，包含超过12.8万个500x500像素的高分辨率卫星图像样本（平均分辨率0.47米），每个样本均附带手工标注的建筑物/非建筑物二进制标签。该数据集具有广泛的地理和语义多样性，覆盖北美、南美、非洲、中东和亚洲等多个地区，并包含不同建筑风格、土地利用背景和大气条件，由橡树岭国家实验室发布。