search

xxlong0 / NeuralUDF

Volume rendering based surface reconstruction using Unsigned Distance Fields
MIT License
279 stars 11 forks source link

Results without Patch loss #2

Open Dawars opened 1 year ago

Dawars commented 1 year ago

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 1 year ago

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.