search

Fictionarry / DNGaussian

[CVPR'24] DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Normalization
https://fictionarry.github.io/DNGaussian/
Other
246 stars 15 forks source link

Optimal Parameters for 360 scenes #19

Open sairisheek opened 4 months ago

sairisheek commented 4 months ago

Thank you for the great work! I was wondering what the optimal parameters would be for 360 degree captures of scene level data with relatively high frequency detail. I also intend on training for the full 30k iterations. Thanks in advance!

Fictionarry commented 4 months ago

Hi, in fact there's not such an optimal hyperparameter setting. Experientially, if the views are not extremely sparse, you can try to reduce the intervention of depth regularization by decreasing the position_lr and densify_grad_threshold. Recommend using the script for MVS initialization as the base and taking adjustments on it.

I do not suggest extending training time if unnecessary, since it is easy to cause overfitting.

sairisheek commented 4 months ago

I believe this is the LLFF MVS script? Traditional 3DGS takes about 20k-25k iterations to accurately converge in the non-sparse view setting for my dataset. Would you still recommend running for a bit less, say 15-20k?

Fictionarry commented 4 months ago

Yes.

You can see original 3DGS can get very good quality even if only at 7k steps, and the remaining in 30k is mainly for refine. If the available views are sparse and can only provide limited information, this refine may be helpless and cause overfitting. Nevertheless, if the views are not sparse, our work may not bring help, as depth would introduce additional errors. It may be better just use original 3DGS

sairisheek commented 4 months ago

Okay, it is in fact a sparse view setting, I am also planning on fitting to a resolution of 1600x1200 or 800x600. Is there anything else I should look out for? Thanks for the help btw!

Fictionarry commented 4 months ago

According to my limited attempts at higher resolutions, I guess the strategies would still work. Higher resolutions may take more time to build the details, thus more training iterations may be needed. You can have a try and see if any problems would happen :)