xxlong0/NeuralUDF

Results without Patch loss

Dawars opened this issue · 1 comments

Dawars commented

Hi,
This paper looks very interesting.

I'm wondering whether the patch loss is a necessary component of this method or it would work without this loss as well.

In the figures 11/12 you do a qualitative comparision against VolSDF and NeuS but they don't have this extra patch loss.
Wouldn't it be better to compare against SparseNeuS?

Thank you!

flamehaze1115 commented

Hello,
The patch loss is not necessary. We don't enforce the patch loss for the open surfaces, since the patch loss assumes the surfaces are continuous.
The patch loss is only included in the training of closed surfaces (DTU here), and you can see the figure.6, the patch loss can improve the reconstruction results. Without the patch loss, our method also works well.
SparseNeuS is a generalizable work targeting sparse views, which is different from our setting.