confusion about the gradients when taking the rho2d

[Nole21]

Thank you for sharing such great work. I have a question about the gradients when taking rho2d. Since depth is computed by depth = s.xTw.x+s.yTw.y+Tw.z. When choosing rho2d, the alpha loss has no gradients to s, but depeh loss still has gradients w.r.t s. Why you do not backpropagate the depth gradients to s in the same way when choosing rho3d and further backpropagate to TransMat(Tu, Tv, Tw)?

This is the code when choosing rho3d, you backpropagate depth loss to s, and further backpropagate it to Tu, Tv, Tw.

When choosing rho2d, you just backpropagate depth gradients w.r.t to Tu, Tv, Tw and ignore the gradients w.r.t s?

Looking foward to your reply, Thank you.

[hbb1]

Oh a good catch. I was caused by the latest commit where I modified the depth for the low pass filters.

I think it may not cause significant difference because the low pass filters are rare.

Did you observe any differences by adding the depth gradients through s back? Looking forward to your sharing!

[Nole21]

Thanks for reply. I update codes when applying low-pass filter in the below way.

The results on DTU dataset is shown as below:

The original results are slightly worse than the updated one.