Skip to content

Latest commit

 

History

History
81 lines (65 loc) · 3.5 KB

data.md

File metadata and controls

81 lines (65 loc) · 3.5 KB
Raw

ZJU-Mocap dataset

  1. Follow the instructions in NeuralBody to request for the dataset.

  2. Modify the yaml files in tools/prepare_zju_mocap, take 387.yaml as an example, update zju_mocap_path to the root path of the downloaded ZJU-Mocap data.

dataset:
  zju_mocap_path: /path/to/ZJUMocap
  subject: '387'
  sex: 'neutral'
  1. Run the scripts of prepare_data.py and render_smpl to preprocess the dataset and render SMPL meshes for training. For example:
export PYTHONPATH=/path/to/elicit:$PYTHONPATH
cd tools/prepare_zju_mocap
python prepare_data.py --cfg 387.yaml
python render_smpl.py --cfg 387.yaml

Human3.6M dataset

  1. Since the license of the Human3.6M dataset does not allow us to release its data, we cannot publicly release the Human3.6M dataset. Please request the dataset here and follow this instruction to prepare the data in ZJU-Mocap format. You can also refer to this page for the Human3.6M dataset preprocessed by animabtable_nerf.

  2. Modify the yaml files in tools/prepare_h36m, take s1.yaml as an example, update h36m_path to the root path of the Human3.6M data.

dataset:
  h36m_path: /path/to/H36M/
  subject: 'S1/Posing'
  sex: 'neutral'
  1. Run the scripts of prepare_data.py and render_smpl to preprocess the dataset and render SMPL meshes for training. For example:
export PYTHONPATH=/path/to/elicit:$PYTHONPATH
cd tools/prepare_h36m
python prepare_data.py --cfg s1.yaml
python render_smpl.py --cfg s1.yaml

Customized data

For customized data with single image input, we provided examples from DeepFashion datasets in dataset/fashion. Please follow step 1-2 to add new data, and follow step 3 for preprocessing.

  1. Estimate SMPL parameters of the subject. Here we use the pose estimation model of CLIFF and PyMAF-X. Please convert the estimated parameters into the same format of CLIFF (refer to data/fashion/fs2818_m394/gt_params.npz as an example).

  2. Prepare the input image and the subject mask (rescale to 1024*1024), and put the data with the folowing structure.

dataset
    └── fashion
        └── ${subject_name}
            ├── gt_image.png # input image
            ├── gt_mask.png # input mask
            └── gt_params.npz # SMPL params
  1. Edit the yaml file in tools/prepare_fashion, here we use the motion sequence of subject 394 in ZJU-Mocap dataset.
dataset: # motion data!
  zju_mocap_path: /path/to/ZJUMoCap/ # root of the motion dataset
  subject: 'CoreView_394' # MOTION_ID
  sex: 'neutral'

If you want to use a customized motion sequence, please prepare it in the same struction of ZJUMocap:

${CUSTOM_ROOT}
    └── ${MOTION_ID}
        ├── annots.npy # annotation of the cameras, it's recommanded to use a similar camera setting of ZJUMocap with 20~30 cameras around the subject
        └── new_params # SMPL annotation of the motion sequence
            ├── 0.npy
            ├── 1.npy
            └── ...

Then run the scripts to prepare SMPL meshes renderings.

python render_smpl.py --cfg ${subject_name}.yaml