Use MeshCNN for regression problem and adding custom edge-features

[ptotikj]

Hi,

is it possible to add some relevant custom features to the existing 5 edge-features that you already generate? I would like to use your MeshCNN for regression purposes (as analogically a "normal" CNN can be used for regression problems on images, videos and etc).

The idea is that I have a mesh object, and since in your MeshCNN the convolutions are performed on the edges, I would like to put custom relevant features on the edges, beside the already existing features that you generate (inner angles, edge-length ratio and the dihedral angle).

The reason is because I have continous values for node-features (7 features in total) and each node has 2 labels (the values that need to be predicted). Since the nodes are not relevant in the convolution process, i can calculate the corresponding attributes (features) from the two connected nodes. The question is now, is it possible to add them to the edge, along side the ones that you generate (which come strictly from the geometry of the mesh object)?

If that is somehow possible, then I could definitely use your network for my problematic.

Best, Petar

[ranahanocka]

Hi @ptotikj ,

Yes, it is not a problem to change the input features. To add different features, see my reply in this thread: https://github.com/ranahanocka/MeshCNN/issues/62#issuecomment-593124691 , where I gave an example how to add a 3-dimensional xyz coordinate.

[lanzithinking]

hi @ptotikj ,to follow up: have you modified the mesh_classifier.py for the purpose of regression?

[zhoujinhai]

I have some vertices that can divide my model into two parts. Can I use MeshCNN to regress these vertices? If that's OK, do you have any suggestions?