DATA.md

Data Preprocessing

Example dataset can be found in README.md. Please check it for data organization and format.

Additional Requirement

• libigl

Training Data

1. Input: 3D scans & fitted SMPL poses.
2. The input data is organized as the following structure:

training_data_dir
├── scan
│   └── 000.ply
│   └── 001.ply
├── smpl
│   └── pose_000.txt
│   └── pose_001.txt
│   └── shape.txt

3. 3D scans (*.ply) contain vertex colors as texture, the SMPL pose in pose_*.txt is defined as [global_trans, global_rot, joint_1_rot, ...], where global_trans, global_rot, joint_*_rot are all 3-dimensional vectors.
4. Download PoissonRecon.exe, and place it in ./gen_data/bin.
5. Run the following script to process the training data.

python -m gen_data.preprocess_training_data --data_dir=training_data_dir

6. Note that this code uses the executable of Poisson Reconstruction, so it can only run on Windows.

Testing Data

1. Input: image sequence & fitted SMPL poses.
2. The input data is organized as the following structure.

testing_data_dir
├── imgs
│   └── color
│      └── color_0000.jpg
│   └── mask
│      └── mask_0000.png
│   └── camera.yaml
├── smpl
│   └── pose_0000.txt
│   └── shape.txt

3. The smpl folder is the same as that in the training data, and the imgs folder contains color & mask sequences and camera information (camera.yaml).
4. Make sure that the normal network checkpoint is in ./pretrained_models/normal_net. This checkpoint is extracted from the trained model of PIFuHD. Many thanks to the authors!
5. Run the following script to process the testing data.

python -m gen_data.preprocess_real_data --data_dir=testing_data_dir --normal_net=./pretrained_models/normal_net/netF.pth