Follow this repository to produce SMPL version of HumanML3D.

conda env create -f environment.yaml
conda activate torch_render

body-visualizer==1.1.0
configer==1.4.1
psbody-mesh==0.4

pip install git+https://github.com/nghorbani/body_visualizer.git
pip install git+https://github.com/MPI-IS/configer
pip install git+https://github.com/MPI-IS/mesh.git

Download SMPL+H mode from SMPL+H (choose Extended SMPL+H model used in AMASS project) and DMPL model from DMPL (choose DMPLs compatible with SMPL). Then place all the models under "./body_models/".

HumanML3D is a 3D human motion-language dataset that originates from a combination of HumanAct12 and Amass dataset.

• ACCD (ACCD)
• HDM05 (MPI_HDM05)
• TCDHands (TCD_handMocap)
• SFU (SFU)
• BMLmovi (BMLmovi)
• CMU (CMU)
• Mosh (MPI_mosh)
• EKUT (EKUT)
• KIT (KIT)
• Eyes_Janpan_Dataset (Eyes_Janpan_Dataset)
• BMLhandball (BMLhandball)
• Transitions (Transitions_mocap)
• PosePrior (MPI_Limits)
• HumanEva (HumanEva)
• SSM (SSM_synced)
• DFaust (DFaust_67)
• TotalCapture (TotalCapture)
• BMLrub (BioMotionLab_NTroje)

In the bracket we give the name of the unzipped file folder.

Unzip all datasets. You could use tools/unzip_amass.py.

Place all files under the directory ./amass_data/.

We follow the original HumanML3D to process the data (framerate, segment, mirror).

First unzip 'humanact12' in './pose_data'.

The following code will run SMPLify to get SMPL parameters from 3D joins.

python smplify_humanact12.py
# You can accelerate the process by running the same script in parallel simultaneously and utilizing multiple GPUs.

Move the generated './humanact12' in root to './pose_data/'.

Run the following to process the data.

# Downsample AMASS data to 20 fps
python s1_framrate.py

# Segment, Mirror, and Relocate
python s2_seg_augmentation.py

# generate 6D rotation representation
python s3_process_init.py

# Compute the mean, std
python s4_cal_mean_std.py

In the end, you should find the data you need at './HumanML3D/smpl/'.

# SMPL-H representation
{
    'bdata_poses': (frame_num, 52*3)
    'bdata_trans': (frame_num, 3) # global position
    'betas':        (16, )
    'gender':       'male'/'female'
    'pose_6d':     (frame_num, 52*6) # 6d rotation representation, better for training
}