Closed ZhenhuiL1n closed 9 months ago
Hi, sorry for my late reply. This is because we use softplus, which is also adopted in TensoRF. Again, we just want to do a fair comparison with TensoRF so we keep it the same.
alpha = 1 - exp(-softplus(density + shift) interval) = 1 - exp(-log(1 + exp(density + shift)) interval) = 1 - exp(log(1 + exp(density + shift)) ^ (-interval)) = 1 - (1 + exp(density + shift)) ^ (-interval)
Best
Hi,
the raw2alpha implementation is different from the original nerf as shown below in the Cuda implementation.
Could you please explain why it is implemented like this, thanks a lot!