search

autonomousvision / monosdf

[NeurIPS'22] MonoSDF: Exploring Monocular Geometric Cues for Neural Implicit Surface Reconstruction
MIT License
573 stars 53 forks source link

Question about the normal computed in Eq(10) in the paper. #32

Closed hbell99 closed 2 years ago

hbell99 commented 2 years ago

Hi,

Before all, thank you for your inspiring work!

I would like to ask some questions about the \hat{N}(r) computed in equation (10). image

If I understand it correctly, the direction of the normal vector calculated at the sampling points \hat{n}_r^i on each ray might be different. Then why does the volume rendering formula used to calculate \hat{N}(r) can represent the surface intersecting the current ray？

Thank you for your time!

niujinshuchong commented 2 years ago

Hi,

It's an approximation. It's true that points on the ray will have different normals. But during training, the model will converge so the sampled points will be very close to the real surface.

hbell99 commented 2 years ago

Hi,

It's an approximation. It's true that points on the ray will have different normals. But during training, the model will converge so the sampled points will be very close to the real surface.

Many thanks to your reply!

I can see what it means now.