X-LeBench
收藏资源简介:
X-LeBench是由布里斯托大学、曼彻斯特大学、剑桥大学等机构联合创建的一个用于极长自我中心视频理解的基准数据集。该数据集通过结合合成日常生活计划和真实视频片段生成,模拟了432个视频生活日志,时长从23分钟到16.4小时不等。数据集的内容涵盖了丰富的日常生活场景,数据来源于Ego4D大规模自我中心视频数据集。创建过程中,利用大语言模型(LLMs)生成与现实世界活动一致的合成计划,并通过动态优化模拟过程生成视频日志。该数据集旨在解决长时自我中心视频理解中的挑战,推动个性化智能助手、长期活动分析等领域的研究。
X-LeBench is a benchmark dataset for extremely long egocentric video understanding, jointly created by institutions including the University of Bristol, the University of Manchester, the University of Cambridge, and other academic organizations. This dataset is generated by combining synthetic daily life plans and real video clips, simulating 432 video lifelogs with durations ranging from 23 minutes to 16.4 hours. It covers a rich variety of daily life scenarios, with data sourced from the large-scale egocentric video dataset Ego4D. During the creation process, large language models (LLMs) are utilized to generate synthetic plans that align with real-world activities, and video lifelogs are produced through dynamic optimization-based simulation procedures. This benchmark aims to address the challenges in long-duration egocentric video understanding, and promote research in fields such as personalized intelligent assistants and long-term activity analysis.




