EgoPlay数据集
收藏资源简介:
EgoPlay数据集是由Snap公司与阿卜杜拉国王科技大学联合构建的大规模事件触发视频编辑数据集,主要用于训练端到端的自我中心视频编辑模型。该数据集包含10.6万个事件触发剪辑-提示对,数据主要源自公开的Ego4D自我中心视频素材,并辅以少量人工标注的通用领域子集以增强多样性,涵盖正向触发、构造性负向触发及多事件提示等多种编辑模式。数据集通过多阶段视觉语言模型流水线构建,首先生成事件描述和编辑指令,随后在事件后片段应用编辑并融合过渡帧以确保时间连续性。该数据集的核心应用在于推动增强现实和可穿戴设备领域的自主视频编辑技术,旨在实现根据视觉事件自动触发编辑操作,解决传统方法需要手动标注时间边界和掩模的局限性。
The EgoPlay Dataset is a large-scale event-triggered video editing dataset jointly constructed by Snap Inc. and King Abdullah University of Science and Technology (KAUST), primarily used for training end-to-end egocentric video editing models. It contains 106,000 event-triggered clip-prompt pairs, which are mainly sourced from publicly available Ego4D egocentric video materials, supplemented by a small number of manually annotated general-domain subsets to enhance diversity, covering multiple editing modes including positive triggers, constructive negative triggers, and multi-event prompts. The dataset is built via a multi-stage vision-language model pipeline: first, event descriptions and editing instructions are generated, then edits are applied to post-event segments and transition frames are fused to ensure temporal consistency. The core applications of this dataset lie in advancing autonomous video editing technologies in the fields of augmented reality (AR) and wearable devices, aiming to enable automatic triggering of editing operations based on visual events and address the limitations of traditional methods that require manual annotation of temporal boundaries and masks.




