search

lzhengning / SubdivNet

Subdivision-based Mesh Convolutional Networks.
MIT License
251 stars 34 forks source link

question about the input data "meshes" #26

Closed amiltonwong closed 2 years ago

amiltonwong commented 2 years ago

Hi, @lzhengning ,

According to data loading from here, meshes['faces']:(bs, num_f, 3), meshes['feats']:(bs, 13, num_f), meshes['Fs']:(bs,) are used for training iteration. Could you elaborate what those 13 channels mean in meshes['feats']?

Thanks

lzhengning commented 2 years ago

Hi @amiltonwong ,

The 13-channel input features are:

  1. 7-dimensional shape descriptor: the face area, the three interior angles of the triangle, and the inner products of the face normal with the three vertex normals (characterizing curvature).
  2. 6-dimensional pose descriptor: the position of the face center and the face normal.

The implementation is the following function: https://github.com/lzhengning/SubdivNet/blob/3c56f9fe8edf08b16fe015ad09684f8ced78a51d/subdivnet/dataset.py#L43