ViTexQA

数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。 当前数据集卡片使用的是默认模版，数据集的贡献者未提供更加详细的数据集介绍，但是您可以通过如下GIT Clone命令，或者ModelScope SDK来下载数据集 <div align="center"> # ViTexQA: A Multi-Frame Temporal Perception Dataset for Video Text Question Answering []() []() [](https://Modelscope.co/datasets/Zhentao1/ViTexQA/tree/main) </div> --- ## 🔥 News - **2026/06** 🎉 ViTexQA has been accepted to **ECCV 2026**. - **2026/06** 🤗 ViTexQA dataset is released on Modelscope. - **Coming Soon** Training code for FrameThinker. --- ## 📖 Introduction ViTexQA is a large-scale benchmark for **multi-frame video text understanding**. Unlike existing video text QA datasets, where many questions can still be answered from a single frame, **every question in ViTexQA requires integrating textual information across multiple video frames.** <div align="center"> <img src="assets/0.png" width="95%"> </div> Our contributions include: - ✅ 5,147 videos - ✅ 6,864 QA pairs - ✅ Multi-frame dependency - ✅ Temporal Chain-of-Thought annotations - ✅ Diverse real-world scenarios (sports, news, driving, tutorials, etc.) - ✅ Synthetic rolling-text video generation pipeline --- ## 📊 Dataset Statistics <div align="center"> <img src="assets/1.png" width="95%"> </div> ViTexQA contains | Item | Value | |------|------:| | Videos | 5,147 | | QA pairs | 6,864 | | Categories | 30 | | Duration | 363 Hours | --- --- # 📝 Annotation Pipeline To ensure that every question requires **genuine multi-frame temporal perception**, we develop a rigorous three-round human annotation pipeline. <div align="center"> <img src="assets/2.png" width="95%"> </div> The annotation process consists of three stages: ### Round 1 · Initial Annotation - Two annotators independently create question-answer pairs. - Questions must require integrating textual information across multiple frames. - Single-frame answerable questions are rejected. ### Round 2 · Quality Evaluation Each sample is reviewed by senior evaluators according to: - Multi-frame dependency - Answer correctness - Question clarity Samples are assigned three quality levels: - **Score 1:** Reject - **Score 2:** Revision Required - **Score 3:** Accepted ### Round 3 · Revision Samples requiring revision are returned to expert annotators for refinement and re-evaluation until all annotations satisfy the highest quality standard. This iterative annotation strategy guarantees: - ✅ Multi-frame dependency - ✅ High-quality QA pairs - ✅ Accurate temporal reasoning annotations - ✅ Diverse and unambiguous questions # 📥 Dataset Download The complete dataset is hosted on Modelscope. Download includes: ``` ViTexQA/ │ ├── train.json ├── test.json │ └── videos/ ├── Real_video/ └── Synthetic_video/ ``` Each sample contains - index - video_name - question - answer - temporal CoT - duration - category --- # 🛠 Synthetic Video Generation We also release the synthetic rolling-text video generation pipeline used in the paper. ``` synthetic/ ├── ILSVRC2012/ ├── TextQA/ ├── font/ ├── textvqa.txt ├── words.txt ├── image2video_rolling.py └── image2video_textqa.py ``` The pipeline supports - Random text corpus - Random fonts - Random colors - Shadows - Transparency - Scrolling animation - Typewriter animation - Multiple transition effects - Automatic QA annotation generation Example ```bash python image2video_rolling.py python image2video_textqa.py ``` Generated videos can be directly used for training and evaluation. --- # 📷 Examples <div align="center"> <img src="assets/3.png" width="90%"> </div> --- # 📄 Paper If you find our work useful, please consider citing ```bibtex @article{guo2026vitexqa, title={ViTexQA: A Multi-Frame Temporal Perception Dataset for Video Text Question Answering}, author={Guo, Zhentao and Duan, Chen and Guan, Tongkun and Wang, Zining and Zhou, Kai and Yan, Pengfei}, journal={arXiv preprint arXiv:2606.24602}, year={2026} } ``` --- # 🙏 Acknowledgements We thank all annotators and the open-source community for making this project possible. --- # ⭐ Star If ViTexQA is useful for your research, please consider giving this repository a ⭐.