Russian Roulette was a smart way to deal with this problem of balancing the amount of computation done with getting accurate results. It makes sense to discard rays with probability proportional to their potential contribution.
ndiamant
Do we only do Russian Roulette on direct illumination samples? Otherwise isn't the Li term expensive to compute?
rlpo
Do we use russian roulette instead of directly pruning low intensity rays since it allows our estimator to remain unbiased?
Bananya
@rlpo I think so. The car at slide 50 is kind of similar to directly pruning low intensity rays since 90% of them are terminated. As a result the shadow in the picture becomes very noisy.
keenan
@rlpo Exactly. If you terminate rays below a certain threshold, then you know with 100% certainty that you're excluding possible light paths from your calculation. If you use Russian roulette, then there's a nonzero probability that all these paths are still included in your calculation, and by carefully factoring in that exact probability ($p_{rr}$) the expected value of your estimate is still equal to the true value of the integral.
keenan
@ndiamant This example does indeed consider direct lighting, where $L_i$ is trivial to compute. However, you can also apply Russian roulette in other settings, e.g., in path tracing you could just use (say) the product of all scattering terms seen so far in your Russian roulette formula (to make it more likely that you terminate paths along which there is a lot of absorption).
Russian Roulette was a smart way to deal with this problem of balancing the amount of computation done with getting accurate results. It makes sense to discard rays with probability proportional to their potential contribution.
Do we only do Russian Roulette on direct illumination samples? Otherwise isn't the Li term expensive to compute?
Do we use russian roulette instead of directly pruning low intensity rays since it allows our estimator to remain unbiased?
@rlpo I think so. The car at slide 50 is kind of similar to directly pruning low intensity rays since 90% of them are terminated. As a result the shadow in the picture becomes very noisy.
@rlpo Exactly. If you terminate rays below a certain threshold, then you know with 100% certainty that you're excluding possible light paths from your calculation. If you use Russian roulette, then there's a nonzero probability that all these paths are still included in your calculation, and by carefully factoring in that exact probability ($p_{rr}$) the expected value of your estimate is still equal to the true value of the integral.
@ndiamant This example does indeed consider direct lighting, where $L_i$ is trivial to compute. However, you can also apply Russian roulette in other settings, e.g., in path tracing you could just use (say) the product of all scattering terms seen so far in your Russian roulette formula (to make it more likely that you terminate paths along which there is a lot of absorption).