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

reproduce problem #38

Open
VLadImirluren opened this issue Aug 29, 2023 · 11 comments
Open

reproduce problem #38

VLadImirluren opened this issue Aug 29, 2023 · 11 comments

Comments

@VLadImirluren
Copy link

image

I use your house geometry and apperance on 4 * 4090, but still meet problem.
How can I solve..
Thanks

@VLadImirluren
Copy link
Author

there are too much noise on rgb and geometry..

@VLadImirluren
Copy link
Author

and the rgb not right...such as the door on the left hand side. I try for many times, but problem still exist

@VLadImirluren
Copy link
Author

and the rgb not right...such as the door on the left hand side. I try for many times, but problem still exist

the rgb doesn't match the geometry....

@ARuiChen
Copy link
Collaborator

ARuiChen commented Aug 30, 2023

1、set "if_use_bump = False" in config file if your appearance has too much noise. This means closing the prediction of Kn in the paper. Actually, this may also be due to the background is fixed black. You can add "black background" in the prompt to alleviate this problem.
2、Strategy 0 is more prone to texture and geometric misalignment, and using strategy 1 or 2 can alleviate this problem. At the same time, it is necessary to adjust the range of FOV and set the minimum value higher. The problem of misalignment is mainly caused by the small FOV of certain perspectives. But please note that if the FOV range is set improperly, the details may not be rich enough.
3、 Train longer.

@VLadImirluren
Copy link
Author

1、set "if_use_bump == False" if your appearance has too much noise. Actually, this is because the background is fixed black. You can add "black background" in the prompt to alleviate this problem 2、Strategy 0 is more prone to texture and geometric misalignment, and using strategy 1 or 2 can alleviate this problem. At the same time, it is necessary to adjust the range of FOV and set the minimum value higher. The problem of misalignment is mainly caused by the small FOV of certain perspectives. But please note that if the FOV range is set improperly, the details may not be rich enough. 3、 Train longer.

thanks for you advice.
I am curious of the reason why I can't reproduce correctly by using your config.
What's the main point you think?
thanks

@ARuiChen
Copy link
Collaborator

ARuiChen commented Aug 30, 2023

Because the number of viewpoints used in one iteration is different (the number of GPUs is different). The sampling algorithm I proposed in the supplementary materials contributes to global consistency in appearance and geometry modeling, but it depends on a large batch size.

@VLadImirluren
Copy link
Author

Because the number of viewpoints used in one iteration is different (the number of GPUs is different). The sampling algorithm I proposed in the attached materials contributes to global consistency in appearance and geometry, but it depends on a large batch size.

Do you think I can solve this problem by using double batchsize on 4gpu?

@ARuiChen
Copy link
Collaborator

I think this is possible, but I haven't tried it before, so I need you to give it a try and let me know if it works.

@VLadImirluren
Copy link
Author

I think this is possible, but I haven't tried it before, so I need you to give it a try and let me know if it works.

okay, I will

@VLadImirluren
Copy link
Author

I think this is possible, but I haven't tried it before, so I need you to give it a try and let me know if it works.

yes, it works. thank you

@ARuiChen
Copy link
Collaborator

Good to hear that. Feel free to ask again if you have any questions.

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

2 participants