Is the reason that sampling the light source area produces a more consistent gray because when sampling solid angles, you could have 0's from the dark area in your average and estimate that would cause random pixels in the solid angle picture to be randomly more dark than its nearest neighbors?
gprechter
@jackcsullivan I think in this specific case the reason that sampling the light source area produces a more consistent gray is because we are specifically sampling from rays that are going toward the light source. The reason there were random pixels that were more dark than others in the previous method was because some of samples could have the random chance that they have no rays that are going towards the light source, so yes I think your idea is correct. With solid angles you'd have 0s from sampling areas that aren't in the light source. I think it's interesting too that the closer to the light source, the less variance there seems to be in the image, probably because the probability that a ray goes into the light source is increased.
hershg
If my understanding is correct, with SAS we have a point on the ground surface, and take 100 random samples on the hemisphere (direction of hemisphere from surface normal); our high noise comes from the significant chance that our small number of samples won't effectively capture the true light distribution at play with the surface in question. Makes lots of sense.
With LAS, we sample exclusively from the actual area of the light source, and use shadow rays to determine if light hits or not/how much light hits. this gives a much cleaner image without so much more computational intensity.
hershg
My question is, what's the relationship between this and importance sampling? Why even bother creating some probabilistic model for where light sources are w.r.t. a given point, when we could instead just literally know the location of the area of the light source and work from there?
Is it that from our model/understanding of the dimensions and location of light source area, that itself is implicitly already our probabilistic model for "locations in the PDF to focus on"? Like in the project we have a uniform, rectangular light source so we have a very simple easy model, and this model itself implicitly encodes this importance sampling PDF we use for LAS?
hershg
Second question, how do we extend this to indirect lighting? After we cast a shadow ray from light source to our surface point, we can calculate the direction of reflection, but where do we go from here? If a ray bounces from light source -> our point -> some other surface point -> etc... how do we accumulate the bounces on other points into our lighting value for our point?
Is the reason that sampling the light source area produces a more consistent gray because when sampling solid angles, you could have 0's from the dark area in your average and estimate that would cause random pixels in the solid angle picture to be randomly more dark than its nearest neighbors?
@jackcsullivan I think in this specific case the reason that sampling the light source area produces a more consistent gray is because we are specifically sampling from rays that are going toward the light source. The reason there were random pixels that were more dark than others in the previous method was because some of samples could have the random chance that they have no rays that are going towards the light source, so yes I think your idea is correct. With solid angles you'd have 0s from sampling areas that aren't in the light source. I think it's interesting too that the closer to the light source, the less variance there seems to be in the image, probably because the probability that a ray goes into the light source is increased.
If my understanding is correct, with SAS we have a point on the ground surface, and take 100 random samples on the hemisphere (direction of hemisphere from surface normal); our high noise comes from the significant chance that our small number of samples won't effectively capture the true light distribution at play with the surface in question. Makes lots of sense.
With LAS, we sample exclusively from the actual area of the light source, and use shadow rays to determine if light hits or not/how much light hits. this gives a much cleaner image without so much more computational intensity.
My question is, what's the relationship between this and importance sampling? Why even bother creating some probabilistic model for where light sources are w.r.t. a given point, when we could instead just literally know the location of the area of the light source and work from there?
Is it that from our model/understanding of the dimensions and location of light source area, that itself is implicitly already our probabilistic model for "locations in the PDF to focus on"? Like in the project we have a uniform, rectangular light source so we have a very simple easy model, and this model itself implicitly encodes this importance sampling PDF we use for LAS?
Second question, how do we extend this to indirect lighting? After we cast a shadow ray from light source to our surface point, we can calculate the direction of reflection, but where do we go from here? If a ray bounces from light source -> our point -> some other surface point -> etc... how do we accumulate the bounces on other points into our lighting value for our point?