The data used for VHM training and eval

VHM 基于一个大规模的遥感图像 - 文本数据集 VersaD 和一个包含事实性和欺骗性问题的诚实指令数据集 HnstD。VersaD 的注释内容丰富，提供了关于图像属性、对象属性和整体场景的详细信息，使 VHM 能够全面理解遥感图像并执行多样化的遥感任务。HnstD 则通过包含欺骗性问题，防止 VHM 对无意义的查询产生肯定回答，确保其诚实性。

VHM is built upon two datasets: a large-scale remote sensing image-text dataset named VersaD, and an honest instruction dataset HnstD that contains both factual and deceptive questions. VersaD is richly annotated with detailed information regarding image attributes, object attributes and the overall scene, which enables VHM to comprehensively understand remote sensing images and perform diverse remote sensing tasks. By incorporating deceptive questions, HnstD prevents VHM from generating affirmative responses to meaningless queries, thus ensuring its honesty.

搜集汇总

背景与挑战

背景概述

该数据集是用于VHM视觉语言模型训练和评估的遥感影像数据，包含预训练、微调和测评三部分数据，由多个研究机构联合发布，支持遥感影像分析研究。