💡 Welcome to the Open-Sora-Dataset project! As part of the Open-Sora-Plan project, we specifically talk about the collection and processing of data sets. To build a high-quality video dataset for the open-source world, we started this project. 💪
We warmly welcome you to join us! Let's contribute to the open-source world together! Thank you for your support and contribution. ❤️
💡 欢迎来到Open-Sora-Dataset项目！我们作为Open-Sora-Plan项目的一部分，详细阐述数据集的收集和处理。为给开源世界构建一个高质量的视频数据，我们发起了这个项目。💪
我们非常欢迎您的加入！让我们共同为开源的世界贡献力量！感谢您的支持和贡献。 ❤️
we crawled 40258 videos from open-source websites under the CC0 license. All videos are of high quality without watermarks and All videos are of high quality without watermarks, and about 60% of them are landscape data. The total duration is about 274h 05m 13sThe main sources of data are divided into three parts：
- mixkit: The total number of videos we collected is 1234, the total duration is about 6h 19m 32s, and the total number of frames is 570815. The resolution and aspect ratio distribution histogram of the video is as follows (the ones that account for less than 1% are not listed)：
- pexels: The total number of videos we collected is 7408 the total duration is about 48h 49m 24s and the total number of frames is 5038641. The resolution and aspect ratio distribution histogram of the video is as follows (the ones that account for less than 1% are not listed)：
- pixabay: The total number of videos we collected is 31616 the total duration is about 218h 56m 17s and the total number of frames is 23508970. The resolution and aspect ratio distribution histogram of the video is as follows (the ones that account for less than 1% are not listed):
it is challenging to directly crawl a large quantity of high-quality dense captions from the internet. Therefore, we utilize a mature Image-captioner model to obtain high-quality dense captions. We conducted ablation experiments on two multimodal large models: ShareGPT4V-Captioner-7B and LLaVA-1.6-34B. The former is specifically designed for caption generation, while the latter is a general-purpose multimodal large model. After conducting our ablation experiments, we found that they are comparable in performance. However, there is a significant difference in their inference speed on the A800 GPU: 40s/it of batch size of 12 for ShareGPT4V-Captioner-7B, 15s/it of batch size of 1 for LLaVA-1.6-34B. We open-source all annotations here. We show some statistics here, and we set the maximum length of the model to 300, which covers almost 99% of the samples.
| Name | Avg length | Max | Std |
|---|---|---|---|
| ShareGPT4V-Captioner-7B | 170.0827524529121 | 467 | 53.689967539537776 |
| LLaVA-1.6-34B | 141.75851073472666 | 472 | 48.52492072346965 |
Use panda-70m to split transition video
- Clone this repository and navigate to Open-Sora-Plan folder
git clone https://github.com/PKU-YuanGroup/Open-Sora-Plan
cd Open-Sora-Plan
- Install the required packages
conda create -n opensora python=3.8 -y
conda activate opensora
pip install -e .
- Split video script
git clone https://github.com/PKU-YuanGroup/Open-Sora-Dataset
python split/no_transition.py --video_json_file /path/to/your_video /path/to/save
If you want to know more, check out Requirements and Installation
cd ocr_process
python ocr_process.py -i /path/to/video
Qingdao Weiyi Network Technology Co., Ltd.: Thank you very much for providing us with valuable data