DECAPBENCH
收藏资源简介:
DECAPBENCH是一个针对现代视觉语言模型的新型细粒度图像字幕评价基准,由字节跳动研究院发布。该数据集包含了400条高质量、由人工审核的详细图像字幕,旨在克服传统图像字幕数据集的局限性,如缺乏详细信息和适当的评价标准。DECAPBENCH结合了人类标注者和机器生成的种子,以产生无幻觉的超级详细图像描述,为细粒度图像字幕任务提供了一种新的评价方法。
DECAPBENCH is a novel fine-grained image captioning evaluation benchmark tailored for modern vision-language models, released by ByteDance Research. This dataset comprises 400 high-quality, manually-reviewed detailed image captions, designed to address the limitations of conventional image captioning datasets, including insufficient detailed information and lack of appropriate evaluation criteria. DECAPBENCH integrates human annotators and machine-generated seeds to produce hallucination-free ultra-detailed image descriptions, thereby offering a novel evaluation approach for the fine-grained image captioning task.




