efficiency about the Fixed Culling

[cv-lab-x]

hi, thanks for your work, did you test the efficiency about the "Fixed Culling"? training time and rendering fps compare with origin diff-gaussian-rasterization ? Looking forward to your reply, thanks, @dendenxu

[dendenxu]

Hi, the efficiency gain can be very scene-dependent and view-dependent since very little or very large amounts of points can fall out of the viewing frustum. I would suggest running the rendering on the dataset of your interest to check the efficiency difference.

As a side note, it sometimes might also be preferable to disable the culling altogether (like in the original 3DGS repo) since the current culling function is merely an approximation using the center of the Gaussian. If you're interested in a more "correct" culling, I would suggest looking at the implementation in fast-gaussian-rasterization