HashNeRF-pytorch
HashNeRF-pytorch copied to clipboard
Added meshing of density field using marching cubes
Added meshing during training and meshing-only once trained, creating a grid of densities and meshing using marching cubes
Thanks for the useful contribution, @amanshenoy! I will review the added code soon. In the meantime, do you have any mesh outputs to showcase?
Hey, yeah - had largely tested on synthetic lego dataset. The following mesh screenshot is the synthetic lego dataset trained to 8k iterations with default training configs. Meshing was done at resolution 512 and level set 40.
The mesh does have a lot of artifacts; a lot of which probably is due to the nature of NeRF (no clear definition of where the surface is) and also since it wasn't trained to completion. Higher resolutions should generate better meshes and texturing the mesh would definitely improve appearance!

This looks great! You can see much of the geometry here. Ofcourse, the surface isn't high quality but that's an inherent weakness of NeRF. Thanks for the effort here and adding this feature! Will review and merge as soon as I get some time.
This is what I was looking for. Any plan to merge it? Or is there still work needed for this?
This is what I was looking for. Any plan to merge it? Or is there still work needed for this?
Was randomly checking this PR and saw that you accidentally included a datadir change in configs/lego.txt
:
+ datadir = /home/ubuntu/aman/datasets/lego
Might want to leave that out.
@amanshenoy thanks a lot! i have created a mesh file. but, how do i apply color to the mesh?