RarePlanes Dataset

RarePlanes 是来自 CosmiQ Works 和 AI.Reverie 的一个独特的开源机器学习数据集，它结合了真实和合成生成的卫星图像。 RarePlanes 数据集特别关注 AI.Reverie 合成数据的价值，以帮助计算机视觉算法自动检测飞机及其在卫星图像中的属性。尽管存在其他合成/真实组合数据集，但 RarePlanes 是最大的公开可用的超高分辨率数据集，旨在从开销的角度测试合成数据的价值。先前的研究表明，合成数据可以减少所需的实际训练数据量，并可能提高计算机视觉领域许多任务的性能。数据集的真实部分由 253 个 Maxar WorldView-3 卫星场景组成，跨越 112 个位置和 2,142 km^2，有 14,700 架人工标注的飞机。随附的合成数据集是通过 AI.Reverie 的新型模拟平台生成的，包含 50,000 张合成卫星图像和约 630,000 条飞机注释。真实和合成生成的飞机都具有 10 个细粒度属性，包括：飞机长度、翼展、机翼形状、机翼位置、翼展等级、推进力、发动机数量、垂直稳定器数量、鸭翼的存在和飞机角色。最后，我们进行了广泛的实验来评估真实和合成数据集并比较性能。通过这样做，我们展示了合成数据对于从头顶角度检测和分类飞机的任务的价值。

RarePlanes is a unique open-source machine learning dataset from CosmiQ Works and AI.Reverie that combines real and synthetically generated satellite imagery. The RarePlanes dataset specifically focuses on the value of AI.Reverie’s synthetic data to help computer vision algorithms automatically detect aircraft and their attributes in satellite imagery. While other synthetic/real combined datasets exist, RarePlanes is the largest publicly available ultra-high-resolution dataset designed to test the value of synthetic data from an overhead perspective. Prior research has shown that synthetic data can reduce the amount of real-world training data required and potentially improve performance across many tasks in the field of computer vision. The real portion of the dataset consists of 253 Maxar WorldView-3 satellite scenes spanning 112 locations and 2,142 km², with 14,700 manually annotated aircraft. The accompanying synthetic dataset, generated via AI.Reverie’s novel simulation platform, contains 50,000 synthetic satellite images and approximately 630,000 aircraft annotations. Both real and synthetically generated aircraft feature 10 fine-grained attributes, including: aircraft length, wingspan, wing shape, wing position, wingspan class, propulsion type, number of engines, number of vertical stabilizers, presence of canards, and aircraft role. Finally, we conducted extensive experiments to evaluate both the real and synthetic datasets and compare their performance. In doing so, we demonstrate the value of synthetic data for the task of detecting and classifying aircraft from an overhead perspective.

搜集汇总

背景与挑战

背景概述

RarePlanes Dataset是一个结合真实与合成卫星图像的开源数据集，包含253个真实卫星场景和50,000张合成图像，用于飞机检测和属性分类。数据集提供了14,700架真实飞机和630,000条合成飞机的细粒度标注，并通过实验验证了合成数据在计算机视觉任务中的价值。