Lecture 13: Global Illumination and Path Tracing (83)

RichardChen9

Is cpdf a bit of a misnomer, since it should return a suitable probability between 0 and 1 instead of some pdf function that could be unbounded?

andrewdcampbell

Could we perhaps use known properties of the material in determining the continuation probability? Intuitively it seems that using a high continuation probability for a reflective surface would result in less noise while using a low probability for matte materials would be more efficient.

theandrewchan

@RichardChen9 yeah pdf here is a misuse of terminology since the space (continue, terminate) is discrete.

jchen12197

How come we divide by the cpdf every time we add to L? I included this division in my project 3-1 but I'm not sure why.

samparadis

I believe it relates to making the estimator unbiased. Notice how we only add the radiance to L with probability cpdf. Thus, we need to scale it to account for this fact.

Is cpdf a bit of a misnomer, since it should return a suitable probability between 0 and 1 instead of some pdf function that could be unbounded?

Could we perhaps use known properties of the material in determining the continuation probability? Intuitively it seems that using a high continuation probability for a reflective surface would result in less noise while using a low probability for matte materials would be more efficient.

@RichardChen9 yeah pdf here is a misuse of terminology since the space (continue, terminate) is discrete.

How come we divide by the cpdf every time we add to L? I included this division in my project 3-1 but I'm not sure why.

I believe it relates to making the estimator unbiased. Notice how we only add the radiance to L with probability cpdf. Thus, we need to scale it to account for this fact.