Lecture 9: Ray Tracing & Acceleration Structures (84)
aidangarde
In what situations is this heuristic accurate, and when is it less reliable? Is it immediately obvious when this heuristic is good, or is there an easy and efficient way to tell if there is a better heuristic to use for this particular rendering?
KevinXu02
In scenes that a lot of objects clustering in a small portion of the area, I think this heuristic might overestimate the benefit of splitting sparsely populated regions, which could be a case that it is less reliable.
rcorona
I'm wondering if there are any machine learning approaches for estimating cost?
For example, could one collect a dataset of 3D scenes along with their ground truth costs and train a light-weight neural network to get better estimates than this heuristic at comparable cost to generate the estimate?
Alternatively, I'm wondering if there might be any ML methods which skip the cost estimating process altogether and directly infer an efficient KD-tree or BVH?
Hamme122
I want to know why we include a ratio C_trav when calculating the cost, rather than a numerical value.
Alina6618
@Hamme122 I was also confused by this point. Then I realized that the whole expression on the right of the equation * cost_to_intersect_a_triangle would be the actual expression of the cost of processing a cell. Since the cost to intersect a triangle is a constant throughout ray tracing we could just divide the actual cost expression by this constant, so we are left with the equation on the slide. (e.g. for the second term the actual expression should also be SA(L)*TriCount(L)*cost_to_intersect_a_triangle rather than simply SA(L)*TriCount(L))
In what situations is this heuristic accurate, and when is it less reliable? Is it immediately obvious when this heuristic is good, or is there an easy and efficient way to tell if there is a better heuristic to use for this particular rendering?
In scenes that a lot of objects clustering in a small portion of the area, I think this heuristic might overestimate the benefit of splitting sparsely populated regions, which could be a case that it is less reliable.
I'm wondering if there are any machine learning approaches for estimating cost? For example, could one collect a dataset of 3D scenes along with their ground truth costs and train a light-weight neural network to get better estimates than this heuristic at comparable cost to generate the estimate?
Alternatively, I'm wondering if there might be any ML methods which skip the cost estimating process altogether and directly infer an efficient KD-tree or BVH?
I want to know why we include a ratio C_trav when calculating the cost, rather than a numerical value.
@Hamme122 I was also confused by this point. Then I realized that the whole expression on the right of the equation * cost_to_intersect_a_triangle would be the actual expression of the cost of processing a cell. Since the cost to intersect a triangle is a constant throughout ray tracing we could just divide the actual cost expression by this constant, so we are left with the equation on the slide. (e.g. for the second term the actual expression should also be SA(L)*TriCount(L)*cost_to_intersect_a_triangle rather than simply SA(L)*TriCount(L))