search

bennyguo / instant-nsr-pl

Neural Surface reconstruction based on Instant-NGP. Efficient and customizable boilerplate for your research projects. Train NeuS in 10min!
MIT License
856 stars 84 forks source link

Feature MLP use FullyFusedMLP or VanillaMLP? #71

Open YZsZY opened 1 year ago

YZsZY commented 1 year ago

Hello author! Thanks for your great work! Here I want to ask a question. Why NeRF corresponding Feature MLP with FullyFusedMLP and Neus with VanillaMLP, is there anything to this design?

bennyguo commented 1 year ago

Hi! Theoretically, FullyFusedMLP should be faster than VanillaMLP, so we use it for NeRF (although we do not observe significant speedup lol). But for NeuS, computing the analytic normal requires computing second-order derivatives, which is not supported by FullyFusedMLP, so we use the vanilla torch MLP. Note that if you use finite difference to compute the normal like here, you should be able to use FullyFusedMLP, but finite differencing seems to be slower and less effective.

YZsZY commented 1 year ago

Get that! Thanks a lot~