TimeLens-Bench
收藏资源简介:
TimeLens-Bench是一个用于视频时间定位的综合、高质量评估基准,由论文《TimeLens: Rethinking Video Temporal Grounding with Multimodal LLMs》提出。在标注过程中,发现了现有数据集中的关键质量问题,并进行了大量手动修正。与旧有基准相比,TimeLens-Bench显著改变了模型的排名,证明了其提供了更可靠的视频时间定位评估。数据集包含三个经过手动精炼的评估数据集:Charades-TimeLens、ActivityNet-TimeLens和QVHighlights-TimeLens,分别来自Charades-STA、ActivityNet-Captions和QVHighlights源数据集。
TimeLens-Bench is a comprehensive, high-quality evaluation benchmark for video temporal grounding, proposed in the paper "TimeLens: Rethinking Video Temporal Grounding with Multimodal LLMs". During the annotation process, critical quality issues in existing datasets were identified and extensively manually corrected. Compared with prior benchmarks, TimeLens-Bench significantly alters the ranking of models, proving that it provides more reliable evaluations for video temporal grounding. The dataset includes three manually refined evaluation datasets: Charades-TimeLens, ActivityNet-TimeLens, and QVHighlights-TimeLens, which are derived from the source datasets Charades-STA, ActivityNet-Captions, and QVHighlights respectively.
TimeLens-Bench 数据集概述
数据集基本信息
- 数据集名称:TimeLens-Bench
- 许可证:bsd-3-clause
- 许可证链接:https://github.com/TencentARC/TimeLens/blob/main/LICENSE
- 主要语言:英语 (en)
- 任务类别:视频文本到文本 (video-text-to-text)
- 数据规模:10K<n<100K
数据集描述
TimeLens-Bench 是一个用于视频时序定位的综合、高质量评估基准。该数据集在标注过程中识别并手动修正了现有数据集中的关键质量问题。与旧有基准相比,模型在 TimeLens-Bench 上的排名发生了显著变化,表明该基准能为视频时序定位提供更可靠的评估。
数据集统计
该基准包含三个广泛使用的视频时序定位评估数据集的精炼版本。
| 精炼数据集 | 视频数量 | 平均时长 (秒) | 标注数量 | 源数据集 | 源数据集链接 |
|---|---|---|---|---|---|
| Charades-TimeLens | 1313 | 29.6 | 3363 | Charades-STA | https://github.com/jiyanggao/TALL |
| ActivityNet-TimeLens | 1455* | 134.9 | 4500 | ActivityNet-Captions | https://cs.stanford.edu/people/ranjaykrishna/densevid/ |
| QVHighlights-TimeLens | 1511 | 149.6 | 1541 | QVHighlights | https://github.com/jayleicn/moment_detr |
注:为降低 ActivityNet Captions 过高的评估成本,ActivityNet-TimeLens 通过在不同时长区间内均匀采样视频进行构建。
相关资源
- 论文:https://arxiv.org/abs/2512.14698
- 代码仓库:https://github.com/TencentARC/TimeLens
- 项目主页:https://timelens-arc-lab.github.io/
- 模型与数据集合:https://huggingface.co/collections/TencentARC/timelens
- 排行榜:https://timelens-arc-lab.github.io/#leaderboard
- 使用说明:请参考 GitHub 仓库中的指南 (https://github.com/TencentARC/TimeLens#-evaluation-on-timelens-bench)。
引用信息
如果本工作对您的研究和应用有所帮助,请引用我们的论文: bibtex @article{zhang2025timelens, title={TimeLens: Rethinking Video Temporal Grounding with Multimodal LLMs}, author={Zhang, Jun and Wang, Teng and Ge, Yuying and Ge, Yixiao and Li, Xinhao and Shan, Ying and Wang, Limin}, journal={arXiv preprint arXiv:2512.14698}, year={2025} }




