instant-nsr-pl icon indicating copy to clipboard operation
instant-nsr-pl copied to clipboard

Published 20 hours ago • verified publisher icon bennyguo

Can I run the original version of NeuS in this code?

Open WanquanF opened this issue 2 years ago • 4 comments

Hi, thanks for your great code! I am wondering if I can run the original version of NeuS (which is much slower) in this repo? Are there any configs that can achieve this?

WanquanF avatar Nov 15 '22 07:11 WanquanF

Hi! This yet cannot be done with only changes in the config file, mainly because (1) importance sampling is not implemented (in favor of dense sampling + grid pruning); (2) background model is not implemented; (3) some training strategies are not aligned with the original NeuS (for example the warm-up training).

But it's a good point. I'll provide a separate implementation for the original NeuS in this framework when I have time.

bennyguo avatar Nov 16 '22 17:11 bennyguo

╭(○｀∀´○)╯

WanquanF avatar Nov 17 '22 06:11 WanquanF

Hi! This yet cannot be done with only changes in the config file, mainly because (1) importance sampling is not implemented (in favor of dense sampling + grid pruning); (2) background model is not implemented; (3) some training strategies are not aligned with the original NeuS (for example the warm-up training).

But it's a good point. I'll provide a separate implementation for the original NeuS in this framework when I have time.

Thanks for your reply and great code. I tried my custom data set on original NeuS (followd by colmap preprocess pipline decribed here) So if this means we can manauly modify the database and point3D.bin accroding to the intereted .ply file (like, cleaned in MeshLab), and port it to your code, like with NeuS-Colmap config? (as original NeuS runs very slow and does not consider distributed training)

floatshadow avatar Dec 15 '22 04:12 floatshadow

Hi! This yet cannot be done with only changes in the config file, mainly because (1) importance sampling is not implemented (in favor of dense sampling + grid pruning); (2) background model is not implemented; (3) some training strategies are not aligned with the original NeuS (for example the warm-up training).

But it's a good point. I'll provide a separate implementation for the original NeuS in this framework when I have time.

Hi Bennyguo, since the background model is supported now, is it possible to modify the config to replace hash grid with vanilla MLP ?

GCChen97 avatar May 10 '23 03:05 GCChen97