hlwu/changechat-105k
收藏资源简介:
ChangeChat-105k是一个指令遵循数据集,用于支持遥感图像变化分析(RSICA),该范式结合变化检测和视觉问答,实现对双时相遥感图像的多轮指令引导探索。每个样本包含一个双时相图像对(image_A, image_B)以及覆盖六种交互类型的问答:变化描述、分类、量化、定位、开放问答和多轮对话。注释通过基于规则和GPT辅助的混合过程生成,基础图像来自LEVIR-CC数据集。本仓库仅提供注释,需从LEVIR-CC获取图像对。数据集包含约105k个指令样本,覆盖10,077个双时相图像对,包括训练和测试分割,样本格式为JSON,使用ShareGPT风格或LEVIR-CC风格。
ChangeChat-105k is an instruction-following dataset introduced to support remote-sensing image change analysis (RSICA) — a paradigm that combines change detection with visual question answering for multi-turn, instruction-guided exploration of bi-temporal remote-sensing imagery. Each sample pairs a bi-temporal image pair (image_A, image_B) with a question and answer covering six interaction types: change captioning, classification, quantification, localization, open-ended QA, and multi-turn dialogues. The annotations were generated through a hybrid rule-based and GPT-assisted process; the underlying bi-temporal images come from the LEVIR-CC dataset. This repository only ships the annotations, and the image pairs must be obtained from LEVIR-CC. The dataset contains approximately 105k instruction samples over 10,077 bi-temporal image pairs, including train and test splits, with sample formats in JSON using either ShareGPT-style or LEVIR-CC style schemas.




