Download

Getting Started

1. Create folders that store body models, datasets, and logs.

   export REPO_DIR=$PWD
   mkdir -p $REPO_DIR/body_models  # body models
   mkdir -p $REPO_DIR/datasets  # datasets
   mkdir -p $REPO_DIR/logs  # logs
   

2. Download body models.

   Get yourself registered and download models_smplx_v1_1.zip from SMPL-X, and place it at ${REPO_DIR}/body_models/smplx. The data structure should follow the hierarchy below.

   ${REPO_DIR}
   |-- body_models
   |   |-- smplx
   |   |   |-- SMPLX_NEUTRAL.pkl
   |   |   |-- SMPLX_NEUTRAL.npz
   |   |   |-- ...
   |-- datasets
   |-- logs
   |-- README.md
   |-- ...
   

3. Download motion retrieval datasets.

   • Register and download all the available SMPL+H sequences from AMASS (This would take long time).
   • Register and download babel_v1-0_release.zip from BABEL. Note that BABEL provides the annotations for the AMASS dataset.

   Please place the datasets at ${REPO_DIR}/datasets.

   You also need to download our manually preprocessed BABEL raw labels,

   cd ${REPO_DIR}/datasets 
   wget https://www.dropbox.com/s/6p69k1tl1gk4lom/clip-actor_preprocess.zip
   unzip clip-actor_preprocess.zip
   

   Finally, you need to download featp_2_fps.json from BABEL github.

   The data structure should follow the hierarchy below.

    ${REPO_DIR}
    |-- body_models
    |-- datasets
    |   |-- amass
    |   |   |-- ACCAD
    |   |   |-- BMLhandball
    |   |   |-- ...
    |   |   |-- ...
    |   |-- babel
    |   |   |-- train.json
    |   |   |-- val.json
    |   |   |-- ...
    |   |   |-- raw_label.npy 
    |   |   |-- encoded_raw_label.npy
    |   |   |-- featp_2_fps.json
    |-- logs
    |-- README.md
    |-- ...