UGSCNN: Spherical CNNs on Unstructured Grids
In this project, we present an alternative convolution kernel for deploying CNNs on unstructured grids, using parameterized differential operators. More specifically we evaluate this method for the spherical domain that is discritized using the icosahedral spherical mesh. Our unique convolution kernel parameterization scheme achieves high parameter efficiency compared to competing methods. We evaluate our model for classification as well as semantic segmentation tasks. Please see
experiments/ for detailed examples.
Generate or download mesh files
To acquire the mesh files used in this project, run the provided script
To locally generate the mesh files, the Libigl library is required. Libigl is mainly used for computing the Laplacian and Derivative matrices that are stored in the pickle files. Alternatively the script will download precomputed pickles if the library is not available.
To run experiments, please find details instructions in under individual experiments in
experiments. For most experiments, simply running the script
run.sh is sufficient to start the training process:
chmod +x run.sh ./run.sh
The script will automatically download data files if needed.
Please contact Max Jiang if you have further questions!