Previous | Next --- Slide 47 of 52
Back to Lecture Thumbnails
CMUScottie

I think the speed up is not significant enough?

L100magikarp

It seems to me that we did not save that much time and produced a much lower quality image. I wonder how the images would compare if we allowed each rendering with Russian Roulette to have an equal amount of compute time, say 6.4 seconds. Maybe then the body of the car would look better with a small sacrifice to the shadow that we care less about.

rgrao

I don't think the amount of compute time can be equalized, since the contributing ray is just terminated and so the algorithm/computation ends quicker. It isn't specifically about how many times you run it or anything that can be changed to increase the computations to equalize the time. Correct me if I'm wrong?

keenan

You all have identified the key trade off: Russian roulette reduces compute time, at the cost of increasing variance. The upside is that the algorithm remains unbiased, so you can at least be sure you'll still get an estimate of the right image. Russian roulette can also help avoid pathological behavior, e.g., imagine bouncing a ray back and forth off of two parallel mirrors millions of times, after the ray already hit some black carpeting. This is just a waste of time, and you're better off killing it early.