Going back to the slide where we discussed how the noise came around when either the sampling rate is too high or too low for a single receiving point, we can now see that with the same randomization idea in sampling, if we simply change the sample space from the receiver (the hemisphere where we sample based on the spherical solid angles) to the light emitters instead, the blur noise will become much better since we are not letting unnecessary samples hit the blocker as they do in sampling from the receiving point hemisphere.
eliot1019
Maybe a way to get results similar to this in quality but more efficiently would be to not fire rays from every pixel but do it from every other pixel. Then the pixels that were not chosen can be given lighting values that are averaged by the chosen pixels around it.
GitMerlin
I feel like this idea of only sampling light source area works pretty well when materials are either transparent or totally blocking, but when it comes to reflection/refraction/diffusion, the complexity can be exponentially higher.
avinashnandakumar
@GitMerlin, that is a very interesting point, I'm trying to picture how this importance sampling would work with occlusion that is not completely transparent or completely blocking. Could we introduce some parameter that characterizes the occlusion and take a proportion out of it from our incident ray?
Going back to the slide where we discussed how the noise came around when either the sampling rate is too high or too low for a single receiving point, we can now see that with the same randomization idea in sampling, if we simply change the sample space from the receiver (the hemisphere where we sample based on the spherical solid angles) to the light emitters instead, the blur noise will become much better since we are not letting unnecessary samples hit the blocker as they do in sampling from the receiving point hemisphere.
Maybe a way to get results similar to this in quality but more efficiently would be to not fire rays from every pixel but do it from every other pixel. Then the pixels that were not chosen can be given lighting values that are averaged by the chosen pixels around it.
I feel like this idea of only sampling light source area works pretty well when materials are either transparent or totally blocking, but when it comes to reflection/refraction/diffusion, the complexity can be exponentially higher.
@GitMerlin, that is a very interesting point, I'm trying to picture how this importance sampling would work with occlusion that is not completely transparent or completely blocking. Could we introduce some parameter that characterizes the occlusion and take a proportion out of it from our incident ray?