Skip to content

Latest commit

 

History

History
139 lines (135 loc) · 6.13 KB

hierprob3d_preparation.md

File metadata and controls

139 lines (135 loc) · 6.13 KB
Raw

HierProb3D Data Preparation

Setup Data for Inference

  1. Create a directory that should contain the data for HierProb3D (from now on referred to as <hierprob3d_data_root>)
  2. Extract the downloaded SMPL male, female and neutral model *.pkl files to <hierprob3d_data_root>
  3. Download the SMPL-X model and extract the smplx folder into <hierprob3d_data_root>
  4. Download the SMPL <-> SMPL-X model correspondences and extract them to <hierprob3d_data_root>/transfer
  5. Run 
    garmentor/setup_scripts/setup_hierprob3d_inference.sh /path/to/garmentor/directory <hierprob3d_data_root>
    • This converts chumpy objects in these models to numpy arrays and copies necessary files to your <hierprob3d_data_root> directory
  6. Remove the original SMPL model *.pkl files from <hierprob3d_data_root>
  7. Download pre-trained checkpoints from here and extract them to <hierprob3d_data_root>

Your <hierprob3d_data_root> should now look like this:

<hierprob3d_data_root>/
├── cocoplus_regressor.npy
├── J_regressor_extra.npy
├── J_regressor_h36m.npy
├── pose_hrnet_w48_384x288.pth
├── poseMF_shapeGaussian_net_weights_female.tar
├── poseMF_shapeGaussian_net_weights_male.tar
├── poseMF_shapeGaussian_net_weights.tar
├── smpl
│   ├── SMPL_FEMALE.pkl
│   ├── SMPL_MALE.pkl
│   └── SMPL_NEUTRAL.pkl
├── smplx
│   ├── SMPLX_FEMALE.npz
│   ├── SMPLX_FEMALE.pkl
│   ├── SMPLX_MALE.npz
│   ├── SMPLX_MALE.pkl
│   ├── SMPLX_NEUTRAL.npz
│   ├── SMPLX_NEUTRAL.pkl
│   ├── smplx_npz.zip
│   └── version.txt
├── transfer
│   ├── smpl2smplh_def_transfer.pkl
│   ├── smpl2smplx_deftrafo_setup.pkl
│   ├── smplh2smpl_def_transfer.pkl
│   ├── smplh2smplx_deftrafo_setup.pkl
│   ├── smplx_mask_ids.npy
│   ├── smplx_to_smpl.pkl
│   ├── smplx2smpl_deftrafo_setup.pkl
│   └── smplx2smplh_deftrafo_setup.pkl
└── UV_Processed.mat

Setup Data for Training

Please note: You will need to download roughly 168 GB of data.

  1. Create the directory <hierprob3d_data_root>/training
  2. Download train and validation body poses and textures from here and move them to <hierprob3d_data_root>/training
  3. In the following, these data locations will be referenced:
    • <lsun_zipped>: Directory where the zipped LSUN scenes are located
    • <lsun_unzipped>: Directory where the unzipped LSUN databases are located
    • <lsun_images>: Directory where the extracted LSUN images are located
  4. Download the LSUN datasets to <lsun_zipped>
    • Option 1:
      • Create a new (or use an existing) python environment and install (e.g. with pip) the following packages:
        • numpy
        • opencv-python
        • lmdb
        • tqdm
      • From within garmentor/lsun_for_garmentor, run 
        python3 download.py -o <lsun_zipped>
    • Option 2:
      • Manually download the dataset from this website to <lsun_zipped>
  5. Extract the downloaded *.zip files to <lsun_unzipped>
    • Do not extract the test images
  6. To extract the images from the unzipped databases, you can
    • Option 1:
      • Use the previously created Python environment and run, from within garmentor/lsun_for_garmentor: 
        python data.py export <list all extracted folders from inside <lsun_unzipped> > --out_dir <lsun_images> [--max_extract X]
        • The --max_extract X option allows you to only extract a maximum of X images from each database
    • Option 2:
      • Run 
        garmentor/setup_scripts/setup_hierprob3d_data.sh /path/to/garmentor/directory <lsun_unzipped> <lsun_images> [X]
        • Specifying X is optional and lets you set the maximum number of images that should be extracted from each database
    • lsun_images should now look like this: 
      <lsun_images>/
├── bedroom_train_lmdb_images
├── bedroom_val_lmdb_images
├── bridge_train_lmdb_images
├── bridge_val_lmdb_images
├── church_outdoor_train_lmdb_images
├── church_outdoor_val_lmdb_images
├── classroom_train_lmdb_images
├── classroom_val_lmdb_images
├── conference_room_train_lmdb_images
├── conference_room_val_lmdb_images
├── dining_room_train_lmdb_images
├── dining_room_val_lmdb_images
├── kitchen_train_lmdb_images
├── kitchen_val_lmdb_images
├── living_room_train_lmdb_images
├── living_room_val_lmdb_images
├── restaurant_train_lmdb_images
├── restaurant_val_lmdb_images
├── tower_train_lmdb_images
└── tower_val_lmdb_images
  7. Run 
    garmentor/setup_scripts/setup_hierprob3d_training.sh /path/to/garmentor/directory <hierprob3d_data_root> <lsun_images>

Your <hierprob3d_data_root> directory should now look like this:

<hierprob3d_data_root>/
├── cocoplus_regressor.npy
├── J_regressor_extra.npy
├── J_regressor_h36m.npy
├── pose_hrnet_w48_384x288.pth
├── poseMF_shapeGaussian_net_weights_female.tar
├── poseMF_shapeGaussian_net_weights_male.tar
├── poseMF_shapeGaussian_net_weights.tar
├── smpl
│   ├── SMPL_FEMALE.pkl
│   ├── SMPL_MALE.pkl
│   └── SMPL_NEUTRAL.pkl
├── training
│   ├── lsun_backgrounds
│   │   ├── train
│   │   └── val
│   ├── smpl_train_poses.npz
│   ├── smpl_train_textures.npz
│   ├── smpl_val_poses.npz
│   └── smpl_val_textures.npz
└── UV_Processed.mat