Have you implemented Combined Sampling(Sec 4.4 in the paper)?

[flow-specter]

Awesome work! I wanna ask that if you have implemented Combined Sampling(Sec 4.4 in the paper), which combines occupancy grid and proposal networks?

[fyang235]

+1

[AIBluefisher]

I'm also curious about how the combined sampling is implemented. It is not clear whether the proposal net uses the transmittance obtained from the occupancy grid sampler for importance sampling, or just simply utilizes the occupancy grid to skip empty space. Could you please make it clear here? @liruilong940607

[wanger-666]

+1

[liruilong940607]

Hi I'm not supporting that in this repo because I was hacking some CUDA code to make that happen, but I don't want to mess up the API in this repo.

The proper combination is just like you said, that should be the proposal net uses the transmittance obtained from the occupancy grid sampler for importance sampling. However what we implemented is a simplified version of that, which we simply use the occ grid to return a per-ray near far, and use the per-ray near far to do importance sampling for proposal net.

This is essentially a simplified version of occ grid estimator that assumes linear CDF between the per-ray near far plane, and CDF = 0 or 1 for the regions outside of the near far range.