Location to the Implementation of Feature Consistency Loss

[vhvkhoa]

Hello,

I would like to ask about the implementation of feature consistency loss in your source code. I was looking through the nnutils/mesh_net.py but could only find match loss and cycle (reprojection) loss, and imatch loss (maybe inverse match loss), which seems not reported in the paper.

Would you please help me locate the implementation of feature consistency loss ?

Thank you in advance.

[gengshan-y]

Hi,

Feature consistency loss is not used in the code release since we found it can be removed without loss of reconstruction accuracy. Feature matching loss + re-projection loss was good enough.

But if you are interested in the implementation of it, check this https://github.com/gengshan-y/viser-release/blob/502ac589b7be0a5a0307625cca167b27e19c6914/nnutils/mesh_net.py#L99-L117

https://github.com/gengshan-y/viser-release/blob/502ac589b7be0a5a0307625cca167b27e19c6914/nnutils/mesh_net.py#L784-L800

[vhvkhoa]

Hello,

Thank you very much for your clarification. I would like to follow up by another question related to reprojection loss. I notice that at Line 849 in the code part below:

https://github.com/gengshan-y/viser-release/blob/eac3a8e0599e56781e6bf69a0428e62d7907eccc/nnutils/mesh_net.py#L846-L869

csm_points is detached from the whole network. Therefore, I am wondering which parts of your model is optimized by re-projection loss ?

[gengshan-y]

The articulation parameters and camera parameters (e.g., G, W in eq 1.) are updated by the reprojection loss. feature embedding is only updated by the 3d matching loss (eq 9).