Lecture 9: Intro to Ray-Tracing & Accelerating Ray-Scene Intersection (39)
kkkhanl
is the first step usually done through some kind of machine learning or hard code?
LeslieTrue
@kkkhanl, I think step 1 is not a difficult task to deal with. Unlike computer vision tasks that we process unknown 2D/3D input data, graphics tasks are usually based on manually set 3D objects, which means we usually know the coordinates of each pixel. In this case, finding a bounding box can be easy, or can be synchronously defined when artists are creating the 3D objects.
It's my own assumption. Let me know if I'm wrong.
Staffjamesfong1
@LeslieTrue That's right. Sometimes computer vision is thought of the inverse of computer graphics, and vice-versa! Vision tries to reconstruct a scene given only images of it, and graphics tries to reconstruct an image given only the scene! Oversimplification, but it is interesting to ponder.
We have the benefit of having the ground-truth 3D shape available to us, and so finding the bounding boxes is much easier than in the computer vision case.
chetan-khanna
Somewhat of a side note, but on a (very general level), using this somewhat reminds me of indexing into a database (e.g. into an RDBMS), in that we’re (in a sense) building an index on location in a database of “objects”. While this analogy is on such a general level that it would probably apply to a lot of computer science, I was curious if there were any situations in which typical database-indexing strategies were applied to Uniform Spatial Partitions or even just Acceleration Structures/BVHs in general?
I would think that maybe there would be some benefit to such a strategy in some rare/niche situations, perhaps where you have ultra-large, ultra-high resolution images (e.g. gigabytes in size), and you want to draw objects selectively based on very specific parameters (e.g. their temperature, if you are using an IR camera).
is the first step usually done through some kind of machine learning or hard code?
@kkkhanl, I think step 1 is not a difficult task to deal with. Unlike computer vision tasks that we process unknown 2D/3D input data, graphics tasks are usually based on manually set 3D objects, which means we usually know the coordinates of each pixel. In this case, finding a bounding box can be easy, or can be synchronously defined when artists are creating the 3D objects.
It's my own assumption. Let me know if I'm wrong.
@LeslieTrue That's right. Sometimes computer vision is thought of the inverse of computer graphics, and vice-versa! Vision tries to reconstruct a scene given only images of it, and graphics tries to reconstruct an image given only the scene! Oversimplification, but it is interesting to ponder.
We have the benefit of having the ground-truth 3D shape available to us, and so finding the bounding boxes is much easier than in the computer vision case.
Somewhat of a side note, but on a (very general level), using this somewhat reminds me of indexing into a database (e.g. into an RDBMS), in that we’re (in a sense) building an index on location in a database of “objects”. While this analogy is on such a general level that it would probably apply to a lot of computer science, I was curious if there were any situations in which typical database-indexing strategies were applied to Uniform Spatial Partitions or even just Acceleration Structures/BVHs in general? I would think that maybe there would be some benefit to such a strategy in some rare/niche situations, perhaps where you have ultra-large, ultra-high resolution images (e.g. gigabytes in size), and you want to draw objects selectively based on very specific parameters (e.g. their temperature, if you are using an IR camera).