Monte Carlo code where it divide the number of samples

[NhatMinh2001]

Hi Mitsuba developing team and everyone else, I want to ask this silly question. As i known, a monte carlo estimator aggregates samples and the divide for the number of samples.

But I've not yet found the code where it do that. could you point me the part where the algorithm divide for spp?.
Also, i want to set different parts of image with different, adaptable spp (not known before, like adaptive sampling), what are the right ways to do that?
Big thank to all the developers and contributors who making this project happen.

[njroussel]

Hi @NhatMinh2001

This is split over a few different files. An integrator will typically return a "weight" for each sample, and its value is usually 1. These weights are accumulated into their own channel in the film, like RGB values. Finally the weight is divided out when developing the film. So effectively, this weight channel is equivalent to spp.

For your questions about varying spp, it is handled implicitly. The function Sensor.sample_rays() generates the rays with a constant spp, but it could do so according to a completely arbitrary distribution. So if you want to do some adaptive sampling like technique, modifiying the sample_rays() method should be enough. The change in distribution will automatically be reflected in the accumulated weight channel, and hence the develop() method will divide out the correct value -- nothing needs to be changed there.