en-wholebody.md

Whole-body data preprocessing

In whole-body data processing operation,

1. We generate SMPL-X 322-dim parameters first. Please run the following command for each dataset. The DATASET should be ['KIT', 'H3D', 'BABEL'].

   python smplx_extractor.py --data {DATASET}

   Here, for the BABEL subset, we only support the all mode (include both seg and seq modes).

   We think this is the best of both worlds. We will support both seg and seq modes in the near future. Welcome community contributions!

   Copy all text files to each subset file:

   cp -r ./babel_new_text ./whole-body-motion/BABEL/texts
   cp -r ./humanml3d_new_text ./whole-body-motion/H3D/texts
   cp -r ./kit_new_text ./whole-body-motion/KIT/texts

   After the first step, your file tree should be like:

   ./whole-body-motion
   ├── BABEL
   │   ├── joints           # *.npy for each subfile
   │   └── texts            # *.txt for each subfile
   ├── H3D
   │   ├── joints           # *.npy for each subfile
   │   └── texts            # *.txt for each subfile
   └── KIT
       ├── joints           # *.npy for each subfile
       └── texts            # *.txt for each subfile

2. Finally, we will unify the UniMoCap dataset into HumanML3D format.

   1. Please generate the mirrored texts at first.

      python diff.py --data whole-body-motion

   2. Merge three subsets.

      python split.py --motion_type whole-body-motion

   OK! After all these processing, we will get the SMPL-X-format UniMocap dataset following the structure like:

   ./whole-body-motion/UniMocap
   ├── smplx_322        # *.npy for each subfile
   ├── test.txt
   ├── texts            # *.txt for each subfile
   ├── train.txt
   └── val.txt

For this repo, we only generate the SMPL-X parameters, not a H3D-format like motion representation. We do not support the mean, variance calculation so far, because the the motion representation has not been unified. However, you can merge the current format into Motion-X as a subset. We will work on this to support a motion representation for SMPL-X in the following weeks.