search

NVlabs / nvdiffrec

Official code for the CVPR 2022 (oral) paper "Extracting Triangular 3D Models, Materials, and Lighting From Images".
Other
2.09k stars 222 forks source link

Confused about the tetrahedral grid resolution #128

Closed zhao-yiqun closed 1 year ago

zhao-yiqun commented 1 year ago

Hi

Thanks for your great works!

I found some descriptions about the resolutions that I can not fully understand. In your supplemental material, Sec 9.1, the implementation details. There are descriptions about "Unless otherwise noted, we start from a tetrahedral grid of resolution 128 (using 192k tetrahedra and 37k vertices).". I check the tetrahedral grid files provided here (tets-files)[https://github.com/NVlabs/nvdiffrec/tree/main/data/tets]. There are nearly same amount of (37k verts and 192k tets) in 64_tets.npz. But in 128_tets.npz. There are almost (270k vertices and 1.5 M tets). It seems that the resolution and the verts number are not consistent. And in your default implementation, you use the resolution of 64_tets. This confused me. Which file is correctly correspond to the 128-resolution stated in paper?

Looking forward to your reply. Thanks.

jmunkberg commented 1 year ago

Hello @zhao-yiqun,

You are right. There is an error in the supplemental. The stats should be:

The configs for the results in the paper are here: https://github.com/NVlabs/nvdiffrec/tree/main/configs Depending on the scene, we sometimes use a grid res of 64 (like the nerd_ehead.json) but most often a grid res of 128.

If you want to test with different grid resolutions, you can generate others following the instructions at: https://github.com/NVlabs/nvdiffrec/blob/main/data/tets/README.md

zhao-yiqun commented 1 year ago

Thanks for your reply!