Skip to content

GEN_DATA.md

Latest commit

 

History

History
67 lines (48 loc) · 2.41 KB

GEN_DATA.md

File metadata and controls

67 lines (48 loc) · 2.41 KB
Raw

Training Dataset

Data Organization

The structure of the training dataset we used is the following

data_dir
├── body_masks
|   └── 0004
│   └── 0016
├── images
|   └── 0004
│   └── 0016
├── labels
|   └── 0004
│   └── 0016
├── masks
│   └── 0004
│   └── 0016
├── cameras.pkl
├── smpl_params.npz

Preprocessing

  1. Reconstruct a template of the cloth using the label images.
  • save it as template_body.py
  1. Reconstruct a template of the body using the smpl parameters.
  • save it as template_cloth.py

  1. Given the pose and body shape from smpl_params.npz, generate template_body_offset.py and template_cloth_offset.py, which stores the body and cloth meshes in canonical shape-zero space. The side product template_body_offset.pkl and template_cloth_offset.pkl stores the shape offset information

  2. Generate original beta-shaped and shape-zero weight volume:

    1. For shape-zero: gen_data/gen_weight_volume.py -c configs/4d_dress/template.yaml -z
    2. For original-shape: gen_data/gen_weight_volume.py -c configs/4d_dress/template.yaml

  3. Generate position map for both body and cloth using:

    1. gen_data/gen_pos_maps.py -c configs/4d_dress/avatar.yaml -rc -ro -o smpl_pos_map_offset_body -t template_body_offset -lw cano_weight_volume_shape_zero
    2. gen_data/gen_pos_maps.py -c configs/4d_dress/avatar.yaml -rc -ro -o smpl_pos_map_offset_cloth -t template_cloth_offset -lw cano_weight_volume_shape_zero

Testing Dataset

Data Organization

The structure of the testing dataset we used is similar as the training data:

Preprocessing

  1. Reuse the following files from training dataset for the same subject

    • cameras.pkl
    • cano_weight_volume_shape_zero.npz
    • cano_weight_volume.npz
    • template_body_offset.py
    • template_cloth_offset.py
    • template_body_offset.pkl
    • template_cloth_offset.pkl

  2. Prepare the test pose sequence and save as smpl_params.npz

  3. Generate position map for both body and cloth using: 0. gen_data/gen_pos_maps.py -c configs/4d_dress/avatar.yaml -rc -ro -o smpl_pos_map_offset_body -t template_body_offset -lw cano_weight_volume_shape_zero

    1. gen_data/gen_pos_maps.py -c configs/4d_dress/avatar.yaml -rc -ro -o smpl_pos_map_offset_cloth -t template_cloth_offset -lw cano_weight_volume_shape_zero