Datasets:

Modalities:
Text
Formats:
webdataset
Languages:
English
ArXiv:
Libraries:
Datasets
WebDataset
License:
TimeLens-Bench / README.md
JungleGym's picture
Update README.md
5fc78c4 verified
---
license: other
license_name: bsd-3-clause
license_link: https://github.com/TencentARC/TimeLens/blob/main/LICENSE
language:
- en
task_categories:
- video-text-to-text
pretty_name: TimeLens
size_categories:
- 10K<n<100K
---
# TimeLens-Bench
πŸ“‘ [**Paper**](https://arxiv.org/abs/2512.14698) | πŸ’» [**Code**](https://github.com/TencentARC/TimeLens) | 🏠 [**Project Page**](https://timelens-arc-lab.github.io/) | πŸ€— [**Model & Data**](https://huggingface.co/collections/TencentARC/timelens) | πŸ† [**TimeLens-Bench Leaderboard**](https://timelens-arc-lab.github.io/#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](TODO).
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](TODO) and [project page](https://timelens-arc-lab.github.io/).)
<img src="https://cdn-uploads.huggingface.co/production/uploads/65372e922c6ef949b22c26d9/31s82GO6S5LKlW0-kcIFU.png" alt="performance_comparison_charades-1" width="35%">
### πŸ“Š Dataset Statistics
The benchmark consists of manually refined versions of **three** widely used evaluation datasets for video temporal grounding:
| Refined Dataset | # Videos | Avg. Duration | # Annotations | Source Dataset | Source Dataset Link |
| :--- | :---: | :---: | :---: | :--- | :--- |
| **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 |
<small>* To reduce the high evaluation cost from the excessively large ActivityNet Captions, we sampled videos uniformly across duration bins to curate ActivityNet-TimeLens.</small>
## πŸš€ Usage
To download and use the benchmark for evaluation, please refer to the instructions in our [**GitHub Repository**](https://github.com/TencentARC/TimeLens#-evaluation-on-timelens-bench).
## πŸ“ Citation
If you find our work helpful for your research and applications, please cite our paper:
```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}
}
```