MCG-NJU/TimeLens2-93K
收藏资源简介:
TimeLens2-93K是一个大规模、长视频时间定位数据集。该版本包含23,793个视频和93,232个文本-时间间隔对,其中包括12,091个多跨度对。视频时长从短片段到近100分钟不等,覆盖广泛的网络领域,如娱乐、教育、体育、新闻、科学技术、游戏、旅行、车辆、音乐和日常生活。TimeLens2-93K提供了规模、长上下文覆盖、多跨度监督和显式标签验证的罕见组合。我们希望它能为训练通用视频模型奠定坚实基础,这些模型能够搜索长时间轴、返回可审计证据,并在不同的定位指令之间转移,而不仅仅是记忆一种时间戳格式。
TimeLens2-93K is a large-scale, long-video temporal grounding dataset. This release contains 23,793 videos and 93,232 text–temporal interval pairs, including 12,091 multi-span pairs. The videos range from short clips to nearly 100 minutes and cover broad web domains such as entertainment, education, sports, news, science and technology, gaming, travel, vehicles, music, and daily life. TimeLens2-93K offers a rare combination of scale, long-context coverage, multi-span supervision, and explicit label verification. We hope it serves as a strong foundation for training generalist video models that can search long timelines, return auditable evidence, and transfer across grounding instructions instead of merely memorizing one timestamp format.




