Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Adaptability of the CUDA implementation #71

Open
MulinYu opened this issue Jul 17, 2024 · 10 comments
Open

Adaptability of the CUDA implementation #71

MulinYu opened this issue Jul 17, 2024 · 10 comments

Comments

@MulinYu
Copy link

MulinYu commented Jul 17, 2024

Hello Zehao,

Congratulations on your impressive work!

I attempted to integrate the GOF CUDA implementation into the original 3DGS and Scaffold-GS. However, I noticed a decrease in rendering quality, approximately 1 dB drop in PSNR. The issue seems to come from the densification. Adding points in distant regions appears to be challenging, resulting in a blurred background. Please see the attached image for reference.

Do you have any insights or suggestions regarding this issue?

bicycle

Best,
Mulin

@niujinshuchong
Copy link
Member

niujinshuchong commented Jul 23, 2024

Hi, I am not familiar with Scaffold-GS codebase. Could you be more specific about what you changed? What is the strategy for adding new points?

@MulinYu
Copy link
Author

MulinYu commented Jul 24, 2024

Hi, I am not familiar with Scaffold-GS codebase. Could you be more specific about what you changed? What is the strategy for adding new points?

Hello, thanks for the reply. I observed the same decrease when integrating the GOF CUDA implementation into the original 3DGS. The only change I made was switching the rendering function.

@niujinshuchong
Copy link
Member

OK, I think switching to original 3DGS should work out of box. Do you also use the 3D smoothing filter?

@MulinYu
Copy link
Author

MulinYu commented Jul 24, 2024

OK, I think switching to original 3DGS should work out of box. Do you also use the 3D smoothing filter?

I didn't add the 3D filter but keep the 2D filter in the CUDA part.

@niujinshuchong
Copy link
Member

OK, maybe you should set the kernel_size to 0.

@MulinYu
Copy link
Author

MulinYu commented Jul 25, 2024

OK, maybe you should set the kernel_size to 0.

I already set it to 0.

@niujinshuchong
Copy link
Member

Then I guess the only difference is the 3D smoothing filter. Maybe you can also try to use it.

@MulinYu
Copy link
Author

MulinYu commented Jul 25, 2024

Then I guess the only difference is the 3D smoothing filter. Maybe you can also try to use it.

Got it, I'll have a try. Thanks again for the reply.

@usamex
Copy link

usamex commented Jul 25, 2024

Connected to the CUDA version, is there a reason why this project (and mip-splatting) is using CUDA version 11.3 and not 11.6? Will it work if I install CUDA 11.6 packages instead of 11.3? (IIRC, original 3DGS uses 11.6)

@niujinshuchong
Copy link
Member

Hi, I think both cuda version work.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants