Lecture 13: Global Illumination and Path Tracing (42)

wangcynthia

By summing all of the successive bounces of lights, we can obtain the global illumination rendering of the scene. From the previous slide, we know that this summation converges to we can find the steady-state surface light field of the scene!

fywu85

How should we decide when to stop? Apparently, we cannot really sum over a infinite series. If this were purely matrix multiplication, we can skip all the summations by directly computing the matrix inverse. Is there anything similar to matrix inversion for an arbitrary operator?

sandykzhang

In our project, we use Russian roulette to determine a stopping point as well as impose a maximum depth, so I think this can be arbitrary to an extent and is probably going to dependent on computational power.

Jordanwyli

@fywu85, since this is in reality a converging series, we could probably safely stop after a certain amount of bounces since the effects of additional bounces are diminishing. Maybe you could play around with the parameters and see at what point the effects are basically negligible.

youtuyy

We know that |K| is smaller than 1 because the energy is reduced in transportation so it is guaranteed to converge. I wonder how we can make it converge faster so that we need less computation and memory for the recursion.

By summing all of the successive bounces of lights, we can obtain the global illumination rendering of the scene. From the previous slide, we know that this summation converges to we can find the steady-state surface light field of the scene!

How should we decide when to stop? Apparently, we cannot really sum over a infinite series. If this were purely matrix multiplication, we can skip all the summations by directly computing the matrix inverse. Is there anything similar to matrix inversion for an arbitrary operator?

In our project, we use Russian roulette to determine a stopping point as well as impose a maximum depth, so I think this can be arbitrary to an extent and is probably going to dependent on computational power.

@fywu85, since this is in reality a converging series, we could probably safely stop after a certain amount of bounces since the effects of additional bounces are diminishing. Maybe you could play around with the parameters and see at what point the effects are basically negligible.

We know that |K| is smaller than 1 because the energy is reduced in transportation so it is guaranteed to converge. I wonder how we can make it converge faster so that we need less computation and memory for the recursion.