facebookresearch / NSVF

Open source code for the paper of Neural Sparse Voxel Fields.
MIT License
805 stars 92 forks source link

Forward facing real scene on NSVF #17

Open pureexe opened 4 years ago

pureexe commented 4 years ago

Hi Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt,

As talked in Email, we converted nerf NDC coordinates to support with your code and ran your code on a scene from NeRF, “trex” with minor changes on launch script. Is this the result that you expect?

Here is trex on NSVF. https://drive.google.com/file/d/1EoUnrAEbN6vnnXy7WpWCLqhjex8AByrL/view?fbclid=IwAR27zUKFwvS3gATUMUfQH8Ij_n4EaYJpep94ejFRr2aLbIB9lBFL_FssUaE

Here is our trex dataset. https://drive.google.com/drive/folders/1bKfaNuafR46qfdcxSiAlFyi3Ox7yIkxO?usp=sharing

Here is bash script to train. https://gist.github.com/pureexe/52486aa8f62dff0a9a8770b5861e7deb

Here is bash script to render. https://gist.github.com/pureexe/e9cda0b0adf671a84a1711f6e3e19c4a

Kind regards, Pakkapon

CC: @domejiraphon @sWizad @supasorn

MultiPath commented 4 years ago

Thanks for sharing your data and script. I will take a look and run on our side. But as we mentioned in the email, our method currently only works in Euclidean space as we need to build voxels, which is quite inefficient for such large even unbounded scenes. The original nerf which works in NDC space usually works easier in this case.

kwea123 commented 4 years ago

Also in this case normal map/depth map should be converted in some way, because they don't correspond to real distance anymore.