Motion Capture from Internet Videos

Motion Capture from Internet Videos
Junting Dong*, Qing Shuai*, Yuanqing Zhang, Xian Liu, Xiaowei Zhou, Hujun Bao
ECCV 2020 Project Page

Datasets

Internet video dataset

Download

Modified Human3.6M dataset

You can download our modified Human3.6M dataset here.

Create your own synthetic data

First, we split the origin videos into different folders, and store the 3D annotations as follows.

<path_to_data>
├── data_2d_h36m_cpn_ft_h36m_dbb.npz
├── joints3d
│   ├── S9_Directions 1.mat
│   ├── S9_Directions.mat
│   ├── ...
│   ├── ...
│   ├── ...
│   ├── S9_WalkTogether 1.mat
│   └── S9_WalkTogether.mat
└── S9
    ├── Directions
    │   ├── Directions.54138969.mp4
    │   ├── Directions.55011271.mp4
    │   ├── Directions.58860488.mp4
    │   └── Directions.60457274.mp4
    ├── Directions1
    │   ├── Directions1.54138969.mp4
    │   ├── Directions1.55011271.mp4
    │   ├── Directions1.58860488.mp4
    │   └── Directions1.60457274.mp4
    |   ......
    ├── WalkTogether
    │   ├── WalkTogether.54138969.mp4
    │   ├── WalkTogether.55011271.mp4
    │   ├── WalkTogether.58860488.mp4
    │   └── WalkTogether.60457274.mp4
    └── WalkTogether1
        ├── ......

We use finetune cpn output as our 2D pose from videopose3d

wget https://dl.fbaipublicfiles.com/video-pose-3d/data_2d_h36m_cpn_ft_h36m_dbb.npz

After all, you can generate the synthetic data. More details can be found in the file script/dataset/sample_h36m.py.

python3 script/dataset/sample_h36m.py --video_path <path_to_data>/S9

Quantitative evaluation

Our quantitative evaluation includes two parts: match and reconstruction. We provide the evaluation scripts as example.

zju3dv / iMoCap