Mask implementation for splatfacto

[ChemrovKirill]

Can someone explain how splatfacto uses alpha masks of train images? I only see loss masking in the code.
However, I believe it is not enough. There must be an algorithm to remove gaussians that are outside the masked region. Without it, gaussians initialized outside the masked region will remain there. Also, gaussians may shift outside the masked region during optimization, and since the loss function is not applied to this region, they will remain there.