Question about training of different timestamp

hustvl / 4DGaussians

[CVPR 2024] 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering

https://guanjunwu.github.io/4dgs/

Apache License 2.0

2.26k stars 189 forks source link

Question about training of different timestamp #41

Closed yuxuanfanOrion closed 11 months ago

yuxuanfanOrion commented 12 months ago

Thank you for your great contribution to dynamic gs!

You know that 3d-gs needs a process of convergence

I didn't figure out that in your paper, do every S' of different timestamp needs a similar process of convergence? Or it just do some computing based on the previous S ?

Forgive me if the question is silly. Thanks again!

yuxuanfanOrion commented 12 months ago

I carefully read your paper again and I notice that you only maintain one set fo gs which means the process of convergence is only for the initial gs right?

guanjunwu commented 12 months ago

Hi, Yes, only one 3D Gaussians and a deformation field is maintained. we will release our latest code and paper in this week, in which more details will be discussed.