Effect of normalisation?

[kaikai23]

Hi, really a great work for people to get a fist hand on gaussian splatting!

I have one question regarding the normalisation in function generate_2D_gaussian_splatting:
The kernel was first computed by torch.exp(z)/ 2*pi*sqrt(det(covariance)), and then normalised such that the max pixel value for it is 1. Since torch.exp(z) already satisfies that the max pixel value is 1, why do we bother to do this redundant computation?

As a simple test, I print the value for kernel_normalized and torch.exp(z), and they look the same: