Mementos
收藏资源简介:
Mementos是一个专为评估多模态大型语言模型(MLLMs)在处理图像序列时的推理能力而设计的新基准。该数据集由马里兰大学帕克分校创建,包含4761个图像序列,涵盖日常生活、机器人任务和漫画风格的故事板等多种场景。每个序列都配有人工标注的主要对象及其行为描述。数据集的创建过程涉及从视频中提取图像并进行人工审核和调整,确保描述与实际图像序列相符。Mementos旨在解决现有MLLMs在理解动态世界方面的局限性,特别是在处理连续图像时的行为推断和逻辑关联。
Mementos is a novel benchmark designed to evaluate the reasoning capabilities of multimodal large language models (MLLMs) when processing image sequences. Developed by the University of Maryland, College Park, this dataset contains 4,761 image sequences spanning diverse scenarios including daily life, robotics tasks, and comic-style storyboards. Each sequence is paired with manually annotated descriptions of its primary objects and their behaviors. The dataset creation process involved extracting frames from videos followed by manual review and refinement to ensure that the descriptions align perfectly with the corresponding image sequences. Mementos aims to address the existing limitations of current MLLMs in understanding the dynamic world, particularly regarding behavior inference and logical association when handling continuous images.




