SciMMIR

SciMMIR是一个专为科学领域多模态信息检索设计的基准数据集，由多模态艺术投影研究社区创建。该数据集包含53万精心策划的图像-文本对，这些对从科学文档中的图表和表格中提取，并附有详细的说明文字。数据集进一步通过两级子集-子类别层次结构注释，以促进对基线的更全面评估。SciMMIR旨在解决科学领域中多模态信息检索的独特挑战，如从图表或表格图像中有效提取关键文本信息的困难。该数据集的应用领域包括科学文献的自动化处理和理解，以及提高多模态信息检索系统的性能。

SciMMIR is a benchmark dataset dedicated to multimodal information retrieval in the scientific domain, created by the multimodal art projection research community. This dataset contains 530,000 carefully curated image-text pairs, extracted from figures and tables in scientific documents and accompanied by detailed descriptive captions. The dataset is further annotated with a two-level subset-subcategory hierarchy to facilitate more comprehensive evaluation of baseline retrieval systems. SciMMIR aims to address the unique challenges of multimodal information retrieval in the scientific domain, such as the difficulty of effectively extracting key textual information from figure or table images. Application scenarios of this dataset include automated processing and understanding of scientific literature, as well as enhancing the performance of multimodal information retrieval systems.