FIT-RS

该数据集是一个大规模的指令调整数据集，包含了180万851条指令样本，旨在提升远程感知大型多模态模型（RSLMMs）的细粒度理解能力。它涵盖了48个重要目标类别和58个高价值语义关系类别，并在场景图上提供了详细的注释。此外，该数据集包含了超过40万个三元组和超过21万个带有旋转边界框的对象，这些数据跨越了1273幅极高分辨率（VHR）的遥感图像。该数据集的任务是针对遥感图像理解进行细粒度的指令调整，包括关系推理和图像级别的场景图生成。

This is a large-scale instruction tuning dataset consisting of 1,800,851 instruction samples, designed to enhance the fine-grained understanding capabilities of remote sensing large multimodal models (RSLMMs). It covers 48 important object categories and 58 high-value semantic relation categories, with detailed annotations provided on scene graphs. Additionally, the dataset contains over 400,000 triples and over 210,000 objects with rotated bounding boxes, spanning 1,273 very high-resolution (VHR) remote sensing images. The task of this dataset focuses on fine-grained instruction tuning for remote sensing image understanding, including relational reasoning and image-level scene graph generation.

搜集汇总

背景与挑战

背景概述

FIT-RS是一个专为遥感图像细粒度理解设计的大规模指令数据集，包含180多万条样本和1273幅高分辨率图像，提供48类目标、58类关系及40万+三元组的详细标注，主要支持关系推理和场景图生成任务。