MVP-Bench
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/GuanzhenLi/MVP-Bench
下载链接
链接失效反馈官方服务：
资源简介：
该数据集名为MVP-Bench，包含了530组自然图像和经过处理的图像对，每组图像对都附带有不同感知层次的问题，旨在诊断大型视觉-语言模型（LVLMs）的视觉感知能力。MVP-Bench数据集不仅能够评估模型在高级和低级视觉感知方面的能力，而且还重点关注了处理过的内容如何影响模型的感知。该数据集的规模为530组图像对，其任务是对自然图像和经过处理的图像上大型视觉-语言模型的多层次视觉感知能力进行评估。
The dataset named MVP-Bench contains 530 pairs of natural and processed images, with questions at different perceptual levels attached to each pair. It is designed to diagnose the visual perception capabilities of large vision-language models (LVLMs). MVP-Bench not only evaluates models' abilities in both high-level and low-level visual perception, but also focuses on how processed image content affects model perception. This benchmark provides a targeted assessment of multi-level visual perception capabilities of LVLMs on both natural and processed image inputs.
提供机构：
Guanzhen Li et al.
