Closed chensiyuan030105 closed 4 months ago
Hi, we first get the original scale of the Gaussian by self.get_scaling
and after applying square it becomes the variances (left term in L105). Then we apply the 3D filter (sum in L105 in your screenshot, eq.9 in the paper). Finally applying sqrt to the variance results in scales after 3D filtering.
Thank you for your reply! I have some other questions.
From the ablation study, I think it is 2D Mip filter solves dilation, not 3D Smooth filter, which is different from the conclusion from the games 310
And how about dilation(zooming out)?
I think 3D Smooth filter solve high frequency artifacts by limiting the smallest scale of gaussian (zooming in) and 2D Mip filter solve dilation and aliasing (zooming out).
From my own point of view,
I agree with your first and second statement. However, for dilation, I think the reason for the artifacts caused by dilation is too small gaussian(too high frequency) for zooming out.
The coefficient of gaussian filter in EWA splatting is a good correction but not contribute for solving dilation.(I think)
In Eq.9, it is the square of covariance plus the square of 3D Filter, but in the code, it is the square of exp(scaling) plus the square of 3D Filter squared. Why?