Gaussian volume masking and frustum culling.

[EryiXie]

Hi guys,

I have a question regarding the Gaussian volume masking. As shown in the lines below, the "mask" and/or the "_pc.mask", which is the Gaussian Volume Mask (I hope I at least get this correct):

Compact-3DGS/gaussian_renderer/__init__.py

Lines 67 to 73 in e7372a3

	mask = ((torch.sigmoid(pc._mask) > 0.01).float()- torch.sigmoid(pc._mask)).detach() + torch.sigmoid(pc._mask)
	if rvq_iter:
	scales = pc.vq_scale(pc.get_scaling.unsqueeze(0))[0]
	rotations = pc.vq_rot(pc.get_rotation.unsqueeze(0))[0]
	scales = scales.squeeze()*mask
	rotations = rotations.squeeze()
	opacity = pc.get_opacity*mask

I am wondering if the "_mask" of some 3G Gaussians outside of the current frustum will also be updated since the frustum culling is done inside the Rasterization function.
So, my question is: will this potentially result in removing Gaussians unexpectedly? Or we can simply rely on the random viewpoint of training iterations (around 3,000) between every mask pruning to avoid that?