search

autonomousvision / monosdf

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

About the figure 19 in your final version paper #36

Closed zhao-yiqun closed 2 years ago

zhao-yiqun commented 2 years ago

Hi,

Thanks for your code sharing, I wonder what the different columns of Figure 19 means. Are there both with depth and normal cues? The caption of Fig.19 did not mention about that

Since I try to reproduce the single-res grids with all views without any cues and found the surface is noisy like followings. Therefore I wonder whether your single-res without any cues may cause this.

Regards

image

niujinshuchong commented 2 years ago

Hi,

thanks for pointing this out! The order is the same as the figure in Table 5. You result looks reasonable to me. Do you use our code base?

zhao-yiqun commented 2 years ago

Thans for your quickly reply.

Is it resonable for single-res Voxel-based dtu without any cues?

I did not use the code from your code base, but I just changed from volsdf to NeuS and used the mask. without mask for single-res NeuS may be even worse. I found you have implement some similar experiments, therefore just consulting you to check whether it is a difference caused by volsdf and NeuS or the single-res voxel.

Regards

niujinshuchong commented 2 years ago

Hi,

I got similar results at the beginning of this project. You could try to increase the resolution of the feature grids, or use smooth trilinear interpolation https://github.com/autonomousvision/monosdf/blob/main/code/hashencoder/src/hashencoder.cu#L87 (also see appendix in instant-ngp) or simply use our provided multi-res. girds .

zhao-yiqun commented 2 years ago

Thanks for your quick reply. I will have a try.

Best regards.