RSITMD

RSITMD数据集是为跨模态遥感图像检索任务而构建的精细粒度数据集。该数据集包含4743张遥感图像，对应提供了23715个描述性的标题，以及1至5个精细关键词，以增强对图像中目标对象的描述。与现有的遥感图像数据集相比，RSITMD数据集的标题更加多样化，重复性更低，更能体现遥感图像中目标对象的精细关系。此外，RSITMD数据集的构建考虑了减少类内相似性，增加类间差异性，使得数据集更适合于遥感图像检索任务。实验结果表明，RSITMD数据集在降低数据集相似性、提高检索精度方面具有明显优势。The RSITMD dataset is a fine-grained benchmark constructed for cross-modal remote sensing image retrieval tasks. It comprises 4743 remote sensing images, paired with a total of 23715 descriptive captions (averaging 5 captions per image) and 1 to 5 fine-grained keywords per image to elaborate on the target objects contained in the imagery. Compared with existing remote sensing image datasets, the captions of RSITMD exhibit greater diversity and lower repetition rates, which better capture the fine-grained relationships between target objects in remote sensing images. Furthermore, the construction of RSITMD prioritizes reducing intra-class similarity and enhancing inter-class discrepancy, rendering the dataset more suitable for cross-modal remote sensing image retrieval tasks. Experimental results demonstrate that the RSITMD dataset offers significant advantages in mitigating dataset similarity and boosting retrieval accuracy.

搜集汇总

背景与挑战

背景概述

RSITMD是一个多模态遥感影像数据集，专注于图像文本多模态任务，包括图像检索、语义检索和遥感影像分类，采用MIT开源许可证，但目前数据集文件暂不可用，可能影响实际应用。