Smoothness of registration result

[AMeynen]

Hi all,
Thanks for the interesting algorithm and work. My first test results are really promising!
I notice that the registration results are accurate but are rough and less smooth that both the source and target meshes.
Are there any solutions to counteract this? I am working with anatomical structures and the noise registration results is not ideal.

[Yang-L1]

Since surface mesh in available in your case.
You can modify the chamfer distance loss to incorporate the normal directions of the target surface.
I.e. replacing point-to-point distance with point-to-plane distance.
This will significantly improve the smoothness.

[rabbityl]

check here

[AMeynen]

Dear @Yang-L1 and @rabbityl
The chamfer distance loss already had a normal component but was not fully implemented it seems. After adjusting this commented line:

# cham_normals = cham_norm_x + cham_norm_y if return_normals else None

I added a relative weight factor as a parameter for the loss function.
The results are more satisfactory!

@rabbityl
I tried using the output of the point-to-distance function, this also yields good results!

[binfengjiao]

Excuse me, how do you install MVRegC? I didn't find this installer

[Yang-L1]

@binfengjiao you do not need to install MVRegC. just delete "import MVRegC"