Adaptive Sampling - PSNR

[cheind]

Hey,

first: congratulations on such a useful contribution - The results you achieve are amazing! Lately, I've been working on a pure PyTorch version of your approach

https://github.com/cheind/pure-torch-ngp

and already get quite nice results. However, I still get no more than 27 dB on the synthetic scenes and I've come to the conclusion that it might be related how we treat samples along the rays differently. I'd like to confirm the following:

Is it true that your adaptive sampling approach skips empty voxels and continues to sample from the next non-empty voxel using constant step sizes until a maximum number of steps have been considered? In my approach I pre-generate all samples along the array and then skip those, which fall into empty voxels. Doing it this way allows for a better batch utilization than marching rays iteratively. The drawback is however, that samples in empty voxels are ignored and not resampled in active voxels.