search

deepsphere / deepsphere-cosmo-tf1

A spherical convolutional neural network for cosmology (TFv1).
https://arxiv.org/abs/1810.12186
MIT License
131 stars 30 forks source link

How to deal with mesh data #34

Closed Yaolab-fantastic closed 3 years ago

Yaolab-fantastic commented 4 years ago

Hi, thanks for developing such a great tool. I’m trying to apply healpy to my spherical mesh data for the purpose of interpolation, my mesh data contains face and vertex information and feature information at each vertex. I’m wondering how can I interpolate my mesh data so I can use DeepSphere? Thanks!

nperraud commented 4 years ago

As you know, the current implementation does not take faces informations. A bad solution would be to use a function to send the faces information onto the adjacent vertices. A good solution requires more work/research and would be to use a different type of GNN that uses faces information as well.

mdeff commented 4 years ago

Thanks for your kind words @daidaista. In theory, because the sphere is a 2D manifold, vertex-valued and face-valued functions are equivalent (while edge-valued functions are different). What are your vertex and face features? Couldn't you sample them at the same positions? As you both suggested, interpolation is otherwise a solution (which is usually not bad in practice). Another solution would be to use a different graph for your vertex and face data (the graph supporting the face data would have one vertex per face of the original graph).