Correspondence query meshes and reconstructed meshes

[aniskacem]

No description provided.

[aniskacem]

Dear Authors,

First, I would like to thank you for this great work and for making your codes publicly available.

I have an issue with visualizing the reconstructed meshes from the input meshes. I succeeded in running the codes in the different modes, but when I wanted to visualize the reconstructed meshes I was not able to find a correspondence between the predicted meshes and the input meshes. I saw that in your code, you introduced the variable "unperm_predictions" instead of predictions for the visualization to highlight that the predicted meshes are not in the same order as for the input meshes. Can you recommend a way to find the predicted mesh of each input mesh?

Thanks in advance!

Best regards,

[anuragranj]

Oops. This unperm stuff is leftover code from previous version. Please use vec2mesh to get a mesh, and then visualize the mesh using mesh.show().

https://github.com/anuragranj/coma/blob/master/facemesh.py#L65-L67

Something like this would work.

predicted_mesh = FaceData.vec2mesh(predicted_vec)
predicted_mesh.show()

[anuragranj]

This is with the provided pretrained model? or did you train from scratch? I haven't seen anything like this before? You should at least get something close to a face, because you are learning only the deviations from a mean face.
Here, the mean face is added if you call vec2mesh:
https://github.com/anuragranj/coma/blob/master/facemesh.py#L66