RaduAlexandru / permuto_sdf

Code for our CVPR'23 paper - "PermutoSDF: Fast Multi-View Reconstruction with Implicit Surfaces using Permutohedral Lattices"
https://radualexandru.github.io/permuto_sdf/
MIT License
433 stars 26 forks source link

A Pytorch extension for your occupancy grid and volume rendering modules #6

Open jingsenzhu opened 1 year ago

jingsenzhu commented 1 year ago

Thank you again for your great work!

In your implementation of Permutohedral Encoding, you implement it as a Pytorch CUDA module so it's scalable and easy to migrate to other NeRF and SDF projects. However, I found that you implemented the Permuto SDF modules as pure CUDA classes requiring a lots of dependency libraries ,which I struggled in installing on my system without using a docker :( . Therefore, for code migration needs, I slightly modified your codes which removed most of the dependencies (of course, Pytorch and CUDA are still needed :) , making it into a installable Pytorch module.

If you agree, I'll release my modified codes on github. Thank you very much :D

RaduAlexandru commented 1 year ago

Hi there!

Thanks again for taking the time and for the feedback!

Currently I am also unhappy with how entangled the OccupancyGrid and VolumetricRendering is with the rest of the dependencies. The whole codebase went throught a pretty big refactoring for the public release and I spent most of the time in getting the PermutohedralEncoding into it's own self-contained package. I'm not sure when I'll have time to clean up the rest of it so if you already performed the conversions and made it more modular, feel free to release it :)

Also if you need any help with the refactoring or with the code migration let me know.

jingsenzhu commented 1 year ago

Hi, thanks for you feedback. I've release my version of refactored code on github HERE. It would be better if you could link this and notify me if you have major revisions of your code, so that I can update my release accordingly.

s-esposito commented 1 year ago

Hi @jingsenzhu, thanks for your work! Does your refactored code still need PyTorch to be installed with CXX_ABI=1?

jingsenzhu commented 1 year ago

Hi @jingsenzhu, thanks for your work! Does your refactored code still need PyTorch to be installed with CXX_ABI=1?

Hi @s-esposito , can you describe your problem in detail? My refactored code should be able to be installed with the default Pytorch from PIP or Conda.

s-esposito commented 1 year ago

Hi @jingsenzhu, thanks for the answer. I was just asking because the official PermutoSDF requires PyTorch to be installed with CXX_ABI=1 (from README.md)

jingsenzhu commented 1 year ago

Hi @jingsenzhu, thanks for the answer. I was just asking because the official PermutoSDF requires PyTorch to be installed with CXX_ABI=1 (from README.md)

I don't think my code has this requirement, for I've successfully run my code under a Pytorch directly installed from PIP without any other constraints.

Tell me if you meet other problems when you install my code.