Clothed Avatar Reconstruction

Clothed avatar reconstruction

Table of Contents

• Install
• Dataset
• Training
• Testing

Install

Dataset

For single view reconstruction, please follow the THuman2.0 Data Processing Instruction from ICON. For avatar reconstruction, please follow the

Training

Single image reconstruction

Models are trained on THuman2.0 dataset using normal image as input and output human body in projected space. Users can take RGB image as input by setting option '-ii rgb'.

# PIFu  
python -m apps.train --gpu 0 --data thuman -ii normal -cfg configs/pifu.yaml  

# ICON (*: re-implementation)
python -m apps.train --gpu 0 --data thuman -ii normal -cfg configs/icon.yaml  

# Ours 
python -m apps.train --gpu 0 --data thuman -ii normal -cfg configs/pifu-sdf.yaml   

Avatar Reconstruction

# ARCH* (*: re-implementation)
python -m apps.train --gpu 0 --data mvp -ii normal -cfg configs/arch.yaml  

# ARCH++* (*: re-implementation)

# ARWild 
python -m apps.train --gpu 0 --data mvp -ii normal -cfg configs/arwild.yaml  

# Ours 
python -m apps.train --gpu 0 --data mvp -ii normal -cfg configs/ours.yaml  

Testing

1. Download the pretrained model and put it in ./out/ckpt/ours-normal-1view/.
2. Download extra data (PyMAF, ICON normal model, SMPL model) and put them to ./data.
3. Testing example images in directory ./examples. Results will be saved to ./examples/results.

python -m apps.infer --gpu 0 -cfg configs/ours.yaml  

License