search

jonbarron / robust_loss_pytorch

A pytorch port of google-research/google-research/robust_loss/
Apache License 2.0
656 stars 88 forks source link

Using this loss function for depth priors in NeRF? #32

Closed morsingher closed 2 years ago

morsingher commented 2 years ago

Hi Jon, first of all, thanks for all the amazing work you keep doing.

I am working on integrating pixelwise depth priors into NeRF, in a similar way with respect to your recent paper: https://arxiv.org/pdf/2112.03288.pdf. I was wondering why you didn't try to use this robust loss function instead of GNLL. My intuition is that this should be more robust to wrong depth estimates in textureless areas or non-Lambertian surfaces. Is there something I'm missing or do you think it might be worth to try it out?

jonbarron commented 2 years ago

Hey thanks for the kind words. That seems like a totally reasonable idea!

morsingher commented 2 years ago

Cool, I'll try and let you know!