iOrthoPredictor

The source code of "TSynNet" for our paper "iOrthoPredictor: Model-guided Deep Prediction of Teeth Alignment" (SIGGRAPH ASIA 2020)

We propose a novel framework for visual prediction of orthodontic treatment. The entire framework is as follows:

Our TSynNet automatically disentangles teeth geometry and appearance, enabling visual prediction of orthodontics under the guidance of the synthesized geometry maps:

upper: sythesized geometry maps. lower: results.

Prerequisites

• Linux
• Python 3.6
• NVIDIA GPU + CUDA 10.0 + cuDNN 7.5
• tensorflow-gpu 1.13.1

Getting Started

• Conda installation:

  # 1. Create a conda virtual environment.
  conda create -n tsyn python=3.6 -y
  conda activate tsyn
  
  # 2. Install dependency
  pip install -r requirement.txt

• Please download the example dataset by running:

  python scripts/download_dataset.py

Testing

• Please download the pre-trained model by running:

   python scripts/download_model.py

• Test the model by running:

   python test.py \
   --test_data_dir=examples/cases_for_testing \  
   --use_gan \
   --use_style_cont \
   --use_skip

• You can check the results in examples/cases_for_testing

Training

• Before training with your own dataset, please make it compatible with the data loader in data/data_loader.py.
• Please download the pre-trained vgg weights by running:

  python scripts/download_vgg.py

• Train the model by running:

  python train.py \
  --train_data_dir=your_train_data_dir \
  --val_data_dir=your_val_data_dir \
  --use_gan \
  --use_style_cont \
  --use_skip

Citation

If you find this useful for your research, please cite the following paper.

@article{yang2020iorthopredictor,
  title={iOrthoPredictor: model-guided deep prediction of teeth alignment},
  author={Yang, Lingchen and Shi, Zefeng and Wu, Yiqian and Li, Xiang and Zhou, Kun and Fu, Hongbo and Zheng, Youyi},
  journal={ACM Transactions on Graphics (TOG)},
  volume={39},
  number={6},
  pages={1--15},
  year={2020},
}

Acknowledgement

We build our project based on StyleGAN2.