TimeLens-Bench

📑 Paper | 💻 Code | 🏠 Project Page | 🤗 Model & Data | 🏆 TimeLens-Bench Leaderboard

✨ Dataset Description

TimeLens-Bench is a comprehensive, high-quality evaluation benchmark for video temporal grounding, proposed in our paper TimeLens: Rethinking Video Temporal Grounding with Multimodal LLMs.

During our annotation process, we identified critical quality issues within existing datasets and performed extensive manual corrections. We observed a dramatic re-ranking of models on TimeLens-Bench compared to legacy benchmarks, demonstrating that TimeLens-Bench provides more reliable evaluation for video temporal grounding. (See more details in our paper and project page.)

📊 Dataset Statistics

The benchmark consists of manually refined versions of three widely used evaluation datasets for video temporal grounding:

* To reduce the high evaluation cost from the excessively large ActivityNet Captions, we sampled videos uniformly across duration bins to curate ActivityNet-TimeLens.

🚀 Usage

To download and use the benchmark for evaluation, please refer to the instructions in our GitHub Repository.

📝 Citation

If you find our work helpful for your research and applications, please cite our paper:

@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}
}