Closed gkouros closed 1 year ago
In the implementation used for generating the results of the paper, which was based on the mip-NeRF codebase, setting the bottleneck noise parameter to 0.1 worked slightly better. This was especially useful for the ball
scene, where it prevented the network from getting stuck at a local minimum in the appearance model. In the MultiNeRF codebase, the behavior is slightly different, and we found that zero worked a little better in terms of quantitative performance (but it didn't really make much of a difference).
Hope this helps!
Thank you for the thorough explanation!
Hi,
I have a question regarding the value of bottleneck_noise for Ref-NeRF. In the paper, Appendix C, you mention that it is set to 0.1 which stabilizes the results. However, in the code it is 0 by default and not adapted to 0.1 in the refnerf config file. Was that intentional?
Also, what's the effect of setting it to 0 and not preventing the network from using the bottleneck early in training? Does it prevent the optimization of the roughness, tint etc?
Thank you in advance!