RePOPE
收藏资源简介:
RePOPE数据集是针对POPE基准中存在的标签错误进行重新标注的结果。POPE数据集是用于视觉语言模型中物体幻觉的标准基准,主要基于MSCOCO图像数据集。RePOPE数据集通过重新标注MSCOCO中的500张图像,对标签错误进行了纠正,并为每个图像创建了新的标签集。该数据集旨在评估标签质量对模型结果的影响,并提供了更准确的数据集,以供研究人员在视觉语言模型中评估物体幻觉的能力。
The RePOPE dataset is a re-annotated result targeting the label errors present in the POPE benchmark. The POPE dataset is a standard benchmark for object hallucination in vision-language models, primarily based on the MSCOCO image dataset. The RePOPE dataset corrects label errors by re-annotating 500 images from MSCOCO and creates a new label set for each image. This dataset aims to evaluate the impact of label quality on model results, and provides a more accurate dataset for researchers to assess the object hallucination capability of vision-language models.




