Closed kwea123 closed 4 years ago
You're right, this is leftover functionality from an experiment that didn't go anywhere (trying to use an additional 5th channel to predict weighting in the coarse model that was distinct from the sigma opacity channel).
Normally the output has only 4 channels: rgb and sigma, but here it is set to 5 when
N_importance>0
, which is always the case in default experiments.https://github.com/bmild/nerf/blob/c39def3944c982aa7aabeacdd2e856c07da39b9e/run_nerf.py#L388
So what does this 5th output do? I checked the code of
raw2outputs
and it seems the 5th element is never accessed, so it means the network has a redundant output?