Zihao Huang¹  Shoukang Hu²  Guangcong Wang³  Tianqi Liu¹
Yuhang Zang⁴  Zhiguo Cao¹  Wei Li²  Ziwei Liu²

>**TL;DR**: WildAvatar is a large-scale dataset from YouTube with 10,000+ human subjects, designed to address the limitations of existing laboratory datasets for avatar creation. ## 🔨 Environments ```bash conda create -n wildavatar python=3.9 conda activate wildavatar pip install -r requirements.txt pip install pyopengl==3.1.4 ``` ## 📦 Prepare Dataset 1. Download [WildAvatar.zip](#) 2. Put the **WildAvatar.zip** under [./data/WildAvatar/](./data/WildAvatar/). 3. Unzip **WildAvatar.zip** 4. Install [yt-dlp](https://github.com/yt-dlp/yt-dlp) 5. Run the following scripts + **If you need key frames** (RGB+MASK+SMPL, needed for [SMPL Visualization](https://github.com/wildavatar/WildAvatar_Toolbox/tree/main?tab=readme-ov-file#-smpl-visualization) and [Creating Wild Avatars](https://github.com/wildavatar/WildAvatar_Toolbox/tree/main?tab=readme-ov-file#-creating-wild-avatars) below), + please download and extract images from YouTube on your own, by running ```bash python prepare_data.py --ytdl ${PATH_TO_YT-DLP}$ ``` then you will find images downloaded in [./data/WildAvatar-videos](./data/WildAvatar/xxx/images). + **If you need video clips**, + please download video clips from YouTube on your own, by running ```bash python download_video.py --ytdl ${PATH_TO_YT-DLP}$ --output_root "./data/WildAvatar-videos" ``` then you will find video clips in [./data/WildAvatar-videos](./data/WildAvatar-videos). + **If you need raw videos** (the original user-updated videos), + please download video clips from YouTube on your own, by running ```bash python download_video.py --ytdl ${PATH_TO_YT-DLP}$ --output_root "./data/WildAvatar-videos-raw" --raw ``` then you will find video clips in [./data/WildAvatar-videos-raw](./data/WildAvatar-videos-raw). ## 📊 SMPL Visualization 1. Put the [SMPL_NEUTRAL.pkl](https://smpl.is.tue.mpg.de/) under [./assets/](./assets/). 2. Run the following script to visualize the smpl overlay of the human subject of ${youtube_ID} ```bash python vis_smpl.py --subject "${youtube_ID}" ``` 3. The SMPL mask and overlay visualization can be found in [data/WildAvatar/\${youtube_ID}/smpl](data/WildAvatar/${youtube_ID}/smpl) and [data/WildAvatar/\${youtube_ID}/smpl_masks](data/WildAvatar/${youtube_ID}/smpl_masks) For example, if you run ```bash python vis_smpl.py --subject "__-ChmS-8m8" ``` The SMPL mask and overlay visualization can be found in [data/WildAvatar/__-ChmS-8m8/smpl](data/WildAvatar/__-ChmS-8m8/smpl) and [data/WildAvatar/__-ChmS-8m8/smpl_masks](data/WildAvatar/__-ChmS-8m8/smpl_masks) ## 🎯 Creating Wild Avatars For training and testing on WildAvatar, we currently provide the adapted code for [HumanNeRF](./lib/humannerf) and [GauHuman](./lib/gauhuman). ## 📝 Citation If you find our work useful for your research, please cite our paper. ``` @article{huang2024wildavatar, title={WildAvatar: Web-scale In-the-wild Video Dataset for 3D Avatar Creation}, author={Huang, Zihao and Hu, ShouKang and Wang, Guangcong and Liu, Tianqi and Zang, Yuhang and Cao, Zhiguo and Li, Wei and Liu, Ziwei}, journal={arXiv preprint arXiv:2407.02165}, year={2024} } ``` ## 😃 Acknowledgement This project is built on source codes shared by [GauHuman](https://github.com/skhu101/GauHuman), [HumanNeRF](https://github.com/chungyiweng/humannerf), and [CLIFF](https://github.com/haofanwang/CLIFF). Many thanks for their excellent contributions! ## 📧 Contact If you have any questions, please feel free to contact Zihao Huang (zihaohuang at hust.edu.cn).