New improvement on NeRF!

[HrsPythonix]

There's a recent research aiming to improve quality and robustness on nerf named robustnerf, at https://robustnerf.github.io/public/, it looks pretty straightforward to add this in nerfstudio

[tancik]

Yea, @kerrj has an implementation. We are still testing it, the results currently aren't conclusive. If it ends up working well we will merge it into the nerfacto model.

[CSU-NXY]

Do you have any update about RobustNeRF? I tried to implement it with mipnerf 360, but the model can never converge once I switch to sample 16x16 patches instead of a single pixel.

[tancik]

Here is a branch with robustNeRF implemented. https://github.com/nerfstudio-project/nerfstudio/tree/justin/robustnerf It seems to work, however it is much slower to converge. Performance also depends a bit on hyperparam choices.