Preprocess PROX data for pose and motion generation in indoor scenes

February 7, 2023 ยท View on GitHub

Data preparation

  1. Download original PROX data, contaning scene meshes, camera data, and so on. And they are organized as follows:

    -| $YOUR_PATH/PROX/
---| body_segments/
---| cam2world/
---| scenes/
---| sdf/
---| preprocess_scenes/
    • preprocess_scenes/ are pre-processed scenes that are used for models. You can generate these with the following instruction or just use our pre-processed results.
    • Other folders are directly downloaded from PROX.
    • The path of PROX is specified by dataset.prox_dir in ./configs/task/pose_gen.yaml or ./configs/task/motion_gen.yaml.

  2. Download LEMO data, which will be used as ground truth of the pose and motion.

    • The path of LEMO is specified by dataset.data_dir in ./configs/task/pose_gen.yaml or ./configs/task/motion_gen.yaml.

  3. Download SMPL-X weights.

    • The path of SMPL-X is specified by dataset.smpl_dir in ./configs/task/pose_gen.yaml or ./configs/task/motion_gen.yaml.

  4. Download VPoser weights. (Optional)

    • VPoser is optional for optimization guidance. We don't use this in our default setting.
    • The path of VPoser is specified by dataset.vposer_dir in ./configs/task/pose_gen.yaml or ./configs/task/motion_gen.yaml.

Preprocess the scene to point cloud

  1. Change the folder path configuration in preprocessing/prox/prox_scene.py, i.e., scene_dir of original PROX data folder and preprocess_scenes_dir used to save processed scene point cloud.

    Notes: you also need to change the data path in configuration, i.e., *.yaml files, for training and evaluation.

  2. Execute to process prox scenes:

    cd preprocessing/prox
python prox_scene.py