Skip to content

Latest commit

 

History

History
107 lines (78 loc) · 4.91 KB

README.md

File metadata and controls

107 lines (78 loc) · 4.91 KB
Raw

MVHumanNet: A Large-scale Dataset of Multi-view Daily
Dressing Human Captures (CVPR 2024)

Project Page | Paper (Arxiv) | Dataset

by Zhangyang Xiong#, Chenghong Li#, Kenkun Liu#, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, Shuguang Cui and Xiaoguang Han* from GAP-Lab.

Introduction

MVHumanNet contains 4,500 human identities, 9,000 daily outfits, 60,000 motion sequences, 645 million with extensive annotations, including human masks, camera parameters , 2D and 3D keypoints, SMPL/SMPLX parameters, and corresponding textual descriptions.

Updates (MVHumanNet_Part1 and MVHumanNet_Part2 are available now!)

  • 2024.06.21: MVHumanNet_Part2 is released! 🔥🔥🔥🔥🔥🔥🔥

    We provide links to download MVHumanNet data. Please fill this form to get the download all links (you don't need to fill the previous forms).

    Currently, MVHumanNet_Part1 and MVHumanNet_Part2 together contain approximately 4000 IDs and 8000 outfits. The remaining data will be updated to the same links at a later time.

  • 2024.05.29: Script for downloading MVHumanNet

    We provide the script to download all the contents of the dataset. Before using it, please make sure you have filled out our form and obtained the download links.

  • 2024.05.24: Textual descriptions of MVHumanNet are released! Textual descriptions🔥🔥🔥🔥🔥🔥🔥

  • 2024.05.07: MVHumanNet_Part1 is released! 🔥🔥🔥🔥🔥🔥🔥

    MVHumanNet_Part1 contains about 2500 IDs and 4800 outfits. We provide links to download the MVHumanNet_Part1. Please fill out the form to get the download links.

  • 2023.12.20 Samples of MVHumanNet are available now！！！

    These samples contain 100 outfits, with 6+1 motions sequences for each. We provide a link to download the samples. Please fill out this form to get the download link.

Folder structure

|-- ROOT
    |-- outfits_ID # 100001
        |-- images    # Considering the limitation of storage space, we scaled the image to half the original size and masked some background.
            |-- camera_name
                |-- images  
            |-- camera_name
                |-- images
            ....
        |-- fmask   # corresponding masks.
            |-- camera_name
                |-- mask images 
            |-- camera_name
                |-- mask images 
            ....
        |-- annots # 2D image annotations by openpose.
            |-- camera_name
                |-- annotations  # json files
            |-- camera_name
                |-- annotations  # json files
            ....
        |-- openpose
            |-- camera_name
                |-- 2D keypoints  # json files
            |-- camera_name
                |-- 2D keypoints  # json files
            ....
        |-- smpl_param  #  optimizes from multi-view images
            |-- PKL files
        |-- smplx  #  optimizes from multi-view images
            |-- 3D keypoints
                |-- json files # 3D keypoints
            |-- smpl
                |-- json files 
            |-- smplx_mesh  
                |-- obj files  # smplx meshs
        |-- camera_extrinsics.json   # extrinsics of all cameras
        |-- camera_intrinsics.json   # intrinsics of all cameras
        |-- camera_scale.pkl

Tip

The camera extrinsics from camera_extrinsics.json represent world-to-camera matrix in OpenCV coordinate system. The translation should be multiplied by the camera scale from camera_scale.pkl to correct the scene scale.

If you find our work useful in your research, please consider citing:

@inproceedings{xiong2024mvhumannet,
  title={MVHumanNet: A Large-scale Dataset of Multi-view Daily Dressing Human Captures},
  author={Xiong, Zhangyang and Li, Chenghong and Liu, Kenkun and Liao, Hongjie and Hu, Jianqiao and Zhu, Junyi and Ning, Shuliang and Qiu, Lingteng and Wang, Chongjie and Wang, Shijie and others},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={19801--19811},
  year={2024}
}

License

The data is released under the MVHumanNet Terms of Use, and the code is released under the Attribution-NonCommercial 4.0 International License.

Copyright (c) 2024