WrappingNet

Implementation of the WrappingNet architecture.
The entire framework is illustrated below.

Data Preparation

The dataset for WrappingNet should be prepared as follows:

For training

1. mkdir -p datasets/Manifold40; cd datasets/Manifold40
2. Download processed.zip from https://aspera.pub/3O5IeFo then move into datasets/Manifold40/
3. unzip processed.zip, then check the data under datasets/Manifold40/processed/

For evaluation

4. wget https://cg.cs.tsinghua.edu.cn/dataset/subdivnet/datasets/Manifold40.zip
5. unzip Manifold40.zip
6. mv Manifold40 raw then check the data under datasets/Manifold40/raw/

Dependencies

   pytorch
   pytorch-geometric
   pytorch-lightning
   pytorch-scatter
   botorch
   open3d
   numpy

To Run

To use our generalized face convolutions, follow these steps:

1. Create a python environment with the above dependencies installed
2. Go to ./nndistance/ and run python build.py install. This will build the faster chamfer distance module.
3. Run CUDA_VISIBLE_DEVICES={GPU}, bash scripts/LC.sh or CUDA_VISIBLE_DEVICES={GPU}, bash scripts/basesup3.sh to launch a training script.

Citation

Eric Lei, Muhammad Asad Lodhi, Jiahao Pang, Junghyun Ahn, Dong Tian,
"WrappingNet: Mesh Autoencoder via Deep Sphere Deformation",
To Appear in 2024 IEEE International Conference on Image Processing (ICIP).