TencentARC/TimeLens-Bench

TimeLens-Bench是一个全面的、高质量的视频时间定位评估基准，旨在提供更可靠的评估。该基准包含三个广泛使用的视频时间定位评估数据集的精修版本：Charades-TimeLens、ActivityNet-TimeLens和QVHighlights-TimeLens。这些数据集经过手动修正，解决了现有数据集中的关键质量问题。统计数据包括视频数量、平均持续时间和注释数量。具体来说，Charades-TimeLens包含1313个视频，平均时长为29.6秒，3363个注释；ActivityNet-TimeLens包含1455个视频，平均时长为134.9秒，4500个注释；QVHighlights-TimeLens包含1511个视频，平均时长为149.6秒，1541个注释。

TimeLens-Bench is a comprehensive, high-quality evaluation benchmark for video temporal grounding, designed to provide more reliable evaluation. The benchmark consists of manually refined versions of three widely used evaluation datasets for video temporal grounding: Charades-TimeLens, ActivityNet-TimeLens, and QVHighlights-TimeLens. These datasets have undergone extensive manual corrections to address critical quality issues in existing datasets. Statistics include the number of videos, average duration, and number of annotations. Specifically, Charades-TimeLens contains 1,313 videos with an average duration of 29.6 seconds and 3,363 annotations; ActivityNet-TimeLens contains 1,455 videos with an average duration of 134.9 seconds and 4,500 annotations; and QVHighlights-TimeLens contains 1,511 videos with an average duration of 149.6 seconds and 1,541 annotations.