MC-Bench
收藏资源简介:
MC-Bench是由浙江大学认知计算与学习实验室创建的多上下文视觉定位基准数据集,旨在评估多模态大语言模型(MLLMs)在多图像场景中的视觉定位能力。该数据集包含2000个高质量、手工标注的样本,涵盖了多种领域和学科。数据集内容包括成对的图像、实例级别的标注和相应的文本提示,文本提示分为三种风格:指称、比较和推理。数据集的创建过程包括从多个数据源收集多样化的图像,并精心组织和标注。MC-Bench的应用领域主要集中在多图像场景下的实例级视觉定位任务,旨在解决现有MLLMs在复杂文本描述和跨图像上下文理解方面的不足。
MC-Bench is a multi-context visual grounding benchmark dataset developed by the Cognitive Computing and Learning Lab at Zhejiang University, which is designed to evaluate the visual grounding capabilities of multimodal large language models (MLLMs) in multi-image scenarios. It comprises 2000 high-quality, manually annotated samples spanning multiple domains and disciplines. The dataset includes paired images, instance-level annotations, and corresponding text prompts, which are categorized into three styles: referential, comparative, and inferential. The construction of MC-Bench involves collecting diverse images from multiple data sources, followed by rigorous organization and annotation. Its primary application focuses on instance-level visual grounding tasks within multi-image scenarios, aiming to address the shortcomings of existing MLLMs in complex text description and cross-image context comprehension.
MC-Bench: A Benchmark for Multi-Context Visual Grounding in the Era of MLLMs
数据集概述
- 名称: MC-Bench
- 类型: 多上下文视觉定位基准
- 领域: 视觉语言理解
- 目标: 评估多模态大语言模型(MLLMs)在多图像实例级视觉语言问题上的能力
数据集详情
- 样本数量: 2,000个高质量手动标注样本
- 样本组成: 包含实例级标注的图像对和相应的文本提示
- 文本提示类型:
- 引用型(Referring)
- 比较型(Comparison)
- 推理型(Reasoning)
- 涵盖技能: 超过10种实用技能,如多跳推理、常识推理、多视图推理、时间理解等
数据集链接
- 下载地址: Google Drive
相关论文
- arXiv链接: arXiv (Coming soon)
作者信息
- 作者:
- Yunqiu Xu
- Linchao Zhu
- Yi Yang
- 机构: ReLER Lab, CCAI, Zhejiang University
- 状态: 提交中




