You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I've got some question about the gaussian map regressing algorithm.
(1) Is it possible to allocate more than one pixel on each pixel?
(ex. high frequency region)
(2) For the high resolution (2048x2048) novel view synthesis, how is it rendered based on the original resolution (1024x1024)?
I've checked that gaussians maps are still in size of 1024x1024
The text was updated successfully, but these errors were encountered:
(1) It is a good idea to allocate more than one pixel on each pixel in high-frequency regions or merge some Gaussians in low-frequency regions. The only concern is whether the depth of such non-pixel-aligned Gaussians can be precisely adjusted according to the image frequency or the distribution of Gaussian points. I think it is worth further research. If you need any help or further discussion, please feel free to contact me.
(2) Yes, you are right. GPS-Gaussian uses two 1K-resolution Gaussian maps from both views to render 2K-resolution novel views. The learned scale map enables a higher-resolution rendering based on the binocular input with redundant information in overlapping regions.
Hello, firstly, thanks for your awesome work.
I've got some question about the gaussian map regressing algorithm.
(1) Is it possible to allocate more than one pixel on each pixel?
(ex. high frequency region)
(2) For the high resolution (2048x2048) novel view synthesis, how is it rendered based on the original resolution (1024x1024)?
I've checked that gaussians maps are still in size of 1024x1024
The text was updated successfully, but these errors were encountered: