Datasets:

Modalities:
Video
Size:
< 1K
Libraries:
Datasets
File size: 1,087 Bytes
412fae3
 
 
 
0ac11da
 
 
 
 
412fae3
 
 
0ac11da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
# Human whole body or hand movements

This dataset comprises SMPL-H body motion sequences reconstructed from third-person videos, as well as MANO hand motion sequences derived from ego-centric videos.

## Data-format

- whole_body:
  - "seq_name"/"file_name".npz
    - pose_body: (T, 63)
    - betas: (B)
    - root_orient: (T, 3)

    - trans: (T, 3)

    - "other parameters"

- ego_hand:
  - Sample ego-centric hand motion clips.
  - Each sample has one `.hdf5` annotation file and one same-name `.mp4` source video.
  - The `.hdf5` file contains camera intrinsics, camera transforms, hand/body transforms, and hand confidence values.
  - Example pair:
    - `ego_hand/95125_30577_34946.hdf5`
    - `ego_hand/95125_30577_34946.mp4`

## Usage

- whole_body
  - usage pip to install smplx, and use smplx.SMPLH to convert parameters into human mesh/joints
  - The converted human body mesh uses the negative Y-axis as the gravity axis.

- ego_hand
  - Keep the same-name `.hdf5` and `.mp4` files in one directory.
  - They can be visualized with the EgoDex 2D visualization script.