Datasets:
Tasks:
Video-Text-to-Text
Modalities:
Text
Formats:
webdataset
Languages:
English
Size:
1K - 10K
ArXiv:
License:
| 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 | |
| π [**Paper**](https://arxiv.org/abs/2512.14698) | π» [**Code**](https://github.com/TencentARC/TimeLens) | π [**Project Page**](https://timelens-arc-lab.github.io/) | π€ [**Model & Data**](https://huggingface.co/collections/TencentARC/timelens) | π [**TimeLens-Bench Leaderboard**](https://timelens-arc-lab.github.io/#leaderboard) | |
| ## β¨ 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%"> | |
| ### π 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> | |
| ## π 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). | |
| ## π 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} | |
| } | |
| ``` |