It seems increasing the number of rays using MC is very helpful to reduce the variance. But sampling more rays is also most costly.
We can easily estimate the magnitude of variance but at some sharp area(edge), we need more samples than the smooth area(ground). Maybe we can more wisely allocate different numbers of sampling rays to accelerate and improve the quality of sampling.
Rishiparikh
What sort of criteria would go into allocating number of rays? I agree that edges would need more samples such as the hazy corners of the black rectangle on the floor. But that might be hard to detect where to increase sampling rate and where to decrease.
One separate thought I did have was a gaussian blur or smoothening might give us an average / smoothened image, as a cheap? way to smoothen the image and reduce variance.
shreyaskompalli
I commented on the earlier slide of this example with an idea about how to get the original Monte Carlo example closer to the "True Answer". However, after seeing how this technique is actually performed, it makes sense that the best way to achieve a more realistic effect is to take many samples and average them out, thus reducing the variance of the overall color values.
greeknerd1
So the idea here is that Monte Carlo simulation will work very well as long as we use a large number of samples, since we will naturally converge to the expected value
melodysifry
Putting aside computational difficulty, is there any point at which increasing the number of samples reduces the variance too much? How do we decide how many samples is enough, or too much?
NKJEW
I've always wondered why the Blender renderer looks noisy in the particularly way it does, but now I'm guessing it's probably using a Monte Carlo estimator (with gradually more and more samples over time so that the scene converges). This lets people get a fairly responsive overview of their scene as they move around, but one that still eventually converges to a physically correct final result.
It seems increasing the number of rays using MC is very helpful to reduce the variance. But sampling more rays is also most costly.
We can easily estimate the magnitude of variance but at some sharp area(edge), we need more samples than the smooth area(ground). Maybe we can more wisely allocate different numbers of sampling rays to accelerate and improve the quality of sampling.
What sort of criteria would go into allocating number of rays? I agree that edges would need more samples such as the hazy corners of the black rectangle on the floor. But that might be hard to detect where to increase sampling rate and where to decrease.
One separate thought I did have was a gaussian blur or smoothening might give us an average / smoothened image, as a cheap? way to smoothen the image and reduce variance.
I commented on the earlier slide of this example with an idea about how to get the original Monte Carlo example closer to the "True Answer". However, after seeing how this technique is actually performed, it makes sense that the best way to achieve a more realistic effect is to take many samples and average them out, thus reducing the variance of the overall color values.
So the idea here is that Monte Carlo simulation will work very well as long as we use a large number of samples, since we will naturally converge to the expected value
Putting aside computational difficulty, is there any point at which increasing the number of samples reduces the variance too much? How do we decide how many samples is enough, or too much?
I've always wondered why the Blender renderer looks noisy in the particularly way it does, but now I'm guessing it's probably using a Monte Carlo estimator (with gradually more and more samples over time so that the scene converges). This lets people get a fairly responsive overview of their scene as they move around, but one that still eventually converges to a physically correct final result.