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

Question about 'forgetting' #67

Closed
lif314 opened this issue Feb 4, 2024 · 3 comments
Closed

Question about 'forgetting' #67

lif314 opened this issue Feb 4, 2024 · 3 comments

Comments

@lif314
Copy link

lif314 commented Feb 4, 2024

I noticed that during the training process, some frames are initially optimized very well, with PSNR values exceeding 30. However, when I perform evaluation after the entire map optimization process, the rendering results for the initial frames are not satisfactory. Is there any good way to alleviate this issue? (replica office2 result)

metrics

@lif314 lif314 changed the title Question about 'forgetting' or 're-optimization' Question about 'forgetting' Feb 4, 2024
@lif314 lif314 closed this as completed Mar 4, 2024
@ljjTYJR
Copy link

ljjTYJR commented Mar 14, 2024

I think it is a trade-off between the efficiency and the accuray.
One option might be enlarging the reply buffer;
The other option would be adding some regularization terms to regularize the optimization of Gaussian parameters.

@Nik-V9
Copy link
Contributor

Nik-V9 commented Mar 15, 2024

I believe that enlarging the mapping window and some updates to our keyframing similar to this paper (https://arxiv.org/abs/2402.03246) should help.

@AutoSenseTech
Copy link

I believe that enlarging the mapping window and some updates to our keyframing similar to this paper (https://arxiv.org/abs/2402.03246) should help.

What do you mean by "enlarging the mapping window"? Do you mean changing the parameter mapping_window_size in configs/splatam.py?

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

4 participants