遥感图像军用飞机目标识别数据集（MAR20)

遥感图像军用飞机目标识别是对遥感图像中的军用飞机进行定位和细粒度分类，其在侦察预警、情报分析等领域起着至关重要的作用。但是，由于数据集匮乏，遥感图像军用飞机目标识别发展相对缓慢。为推动该领域的研究进展，本文构建了公开的遥感图像军用飞机目标识别数据集MAR20（Military Aircraft Recognition）。该数据集具有以下特点：（1）MAR20是目前规模最大的遥感图像军用飞机目标识别数据集，包含3842张图像、20种军用飞机型号以及22341个实例，并且每个目标实例具有水平边界框和有向边界框两种标注方式；（2）由于所有的细粒度类别均隶属于飞机大类，因此不同型号的飞机往往具有相似的特征，导致不同型号目标具有较高的相似性；（3）由于遥感图像采集过程中受到气候、季节、光照、遮挡、乃至大气散射等因素的影响，相同型号的目标存在较大的类内差异性。最后，为建立遥感图像军用飞机目标识别基准，本文在MAR20数据集上评估了7种常用的水平框目标识别方法和8种有向框目标识别方法。数据来源于：https://gcheng-nwpu.github.io/#Datasets

Military aircraft target recognition from remote sensing images involves locating military aircraft and performing fine-grained classification on them, which plays a critical role in scenarios such as reconnaissance and early warning, intelligence analysis. However, the development of this field is relatively slow due to the shortage of available datasets. To promote research progress in this area, this paper constructs a public remote sensing image military aircraft target recognition dataset named MAR20 (Military Aircraft Recognition). This dataset has the following characteristics: (1) MAR20 is currently the largest military aircraft target recognition dataset for remote sensing images, containing 3,842 images, 20 types of military aircraft models, and 22,341 instances, with each target instance annotated using both horizontal bounding boxes and oriented bounding boxes; (2) Since all fine-grained categories belong to the general aircraft category, aircraft of different models often exhibit similar features, leading to high similarity between targets of different models; (3) Targets of the same model have significant intra-class variability, as the remote sensing image acquisition process is affected by factors such as climate, season, lighting, occlusion, and even atmospheric scattering. Finally, to establish a benchmark for military aircraft target recognition from remote sensing images, this paper evaluates 7 commonly used horizontal bounding box-based object recognition methods and 8 oriented bounding box-based object recognition methods on the MAR20 dataset. The dataset is available at: https://gcheng-nwpu.github.io/#Datasets

搜集汇总

背景与挑战

背景概述

MAR20是当前最大的遥感图像军用飞机目标识别数据集，包含3842张图像、20种飞机模型和22341个实例，提供水平和定向两种边界框标注方式。该数据集因飞机模型相似性高和同一模型受环境因素影响大而具有挑战性，适用于军事侦察和情报分析领域的研究。