无人机图像语义分割数据集(Aeroscapes)

AeroScapes是一个新的空中语义分割数据集，包含3269张由无人机在5米至50米高度拍摄的720p图像，并提供了11个类别的密集标注。与现有数据集不同，AeroScapes专注于空中视角，场景结构和物体尺度差异显著。本文提出了一种简单有效的方法，通过从其他领域（如地面视角或室内场景）进行渐进式微调，将知识转移到空中分割任务中。我们训练了多个模型，并通过模型集成显著提升了性能，绝对性能提高了8.12%，优于标准基线方法。AeroScapes is a novel aerial semantic segmentation dataset consisting of 3269 720p images captured by drones at altitudes ranging from 5 meters to 50 meters, with dense annotations for 11 categories. Unlike existing datasets, AeroScapes focuses on aerial perspectives, where scene structures and object scales vary significantly. This work proposes a simple yet effective method to transfer knowledge to the aerial segmentation task via progressive fine-tuning from other domains, such as ground-level perspectives or indoor scenes. We train multiple models and achieve significant performance improvement through model ensemble, with an absolute performance gain of 8.12%, outperforming standard baseline methods.

搜集汇总

背景与挑战

背景概述

Aeroscapes是一个无人机图像语义分割数据集，包含3,269张在5-50米高度拍摄的图像，标注了11个类别，专注于空中视角，场景结构和物体尺度具有独特性。该数据集用于支持跨领域知识迁移研究，属于测绘科学与技术中的遥感信息工程领域，数据类型为栅格数据。