FACTUAL

FACTUAL是一个用于文本场景图解析的大型基准数据集，由莫纳什大学和武汉大学共同创建。该数据集包含40,369个并行示例，通过对Visual Genome数据集中的描述进行重新标注，使用名为FACTUAL-MR的新中间表示来确保场景图的忠实性和一致性。FACTUAL数据集的创建过程涉及严格的定义和质量控制，旨在解决现有场景图解析器在忠实性和一致性方面的挑战。该数据集广泛应用于图像标题评估和零样本图像检索等视觉语言任务，显著提高了这些任务的性能。FACTUAL is a large-scale benchmark dataset for textual scene graph parsing, co-developed by Monash University and Wuhan University. This dataset contains 40,369 parallel samples, which are re-annotated based on the captions from the Visual Genome dataset, utilizing a novel intermediate representation named FACTUAL-MR to ensure the faithfulness and consistency of scene graphs. The construction process of the FACTUAL dataset involves strict definitions and quality control, aiming to address the challenges faced by existing scene graph parsers in terms of faithfulness and consistency. This dataset is widely applied to vision-language tasks such as image caption evaluation and zero-shot image retrieval, and has significantly improved the performance of these tasks.

搜集汇总

背景与挑战

背景概述

FACTUAL是一个用于文本场景图解析的大型基准数据集，由莫纳什大学和武汉大学共同创建，包含40,369个并行示例，基于Visual Genome数据重新标注并使用FACTUAL-MR中间表示来确保忠实性和一致性。该数据集通过严格质量控制解决现有解析器的挑战，广泛应用于图像标题评估和零样本图像检索等视觉语言任务，显著提升了性能。