Would a grouping algorithm such as k-means be faster or detrimental to this process? In other words, would being more accurate and smarter about the way we split these sets up be worth the potential asymptotic slowdown caused by implementing a more complicated grouping method? Additionally, would this trade-off be minimized as we add more elements or only worsen?
Carpetfizz
@ryanpmeyer
Clustering would require some kind of metric (perhaps Euclidean distance of vertices of each mesh) so objects that are closer together in space would be in their own box.
It's not clear if this would create smaller bounding boxes which avoid empty space (favorable for ray tracing).
julialuo
The tradeoff here and in the next slide is that with the bounding box with a lot of extra space, rays will have to test against the objects inside more often when they don't intersect any of them. But putting more objects into the box on the right will mean more intersection tests when a ray intersects the right box. I guess we'll learn next lecture how we should deal with this tradeoff.
julialuo
The tradeoff here and in the next slide is that with the bounding box with a lot of extra space, rays will have to test against the objects inside more often when they don't intersect any of them. But putting more objects into the box on the right will mean more intersection tests when a ray intersects the right box. I guess we'll learn next lecture how we should deal with this tradeoff.
chenwnicole
This example shows that splitting at the median isn't always the best - in this case, our bounding box on the left has a ton of white space but we're still testing against every single pixel, causing our algorithm to be inefficient with unnecessary operations.
Would a grouping algorithm such as k-means be faster or detrimental to this process? In other words, would being more accurate and smarter about the way we split these sets up be worth the potential asymptotic slowdown caused by implementing a more complicated grouping method? Additionally, would this trade-off be minimized as we add more elements or only worsen?
@ryanpmeyer Clustering would require some kind of metric (perhaps Euclidean distance of vertices of each mesh) so objects that are closer together in space would be in their own box.
It's not clear if this would create smaller bounding boxes which avoid empty space (favorable for ray tracing).
The tradeoff here and in the next slide is that with the bounding box with a lot of extra space, rays will have to test against the objects inside more often when they don't intersect any of them. But putting more objects into the box on the right will mean more intersection tests when a ray intersects the right box. I guess we'll learn next lecture how we should deal with this tradeoff.
The tradeoff here and in the next slide is that with the bounding box with a lot of extra space, rays will have to test against the objects inside more often when they don't intersect any of them. But putting more objects into the box on the right will mean more intersection tests when a ray intersects the right box. I guess we'll learn next lecture how we should deal with this tradeoff.
This example shows that splitting at the median isn't always the best - in this case, our bounding box on the left has a ton of white space but we're still testing against every single pixel, causing our algorithm to be inefficient with unnecessary operations.