Alain <ele### [at] netscapenet> wrote:
> Bruno Cabasson nous apporta ses lumieres en ce 2007/09/21 09:18:
> > I take the risk of appearing stupid, but is it possible to make the
> > assumption that if a ray hits an object, the rays next to it (next pixel
> > for example, or the next ray shot, or so ...) have a high probability to
> > hit the same object, thus avoiding useless tree-traversal and intersection
> > test? Would it be costful: if the test fails, then we go on with the normal
> > traversal with little overhead, and if success, we save a non neglectable
> > time?
> >
> > Is it possible to collect ray-object intersection statistics while shooting
> > rays and intersecting, in order to make prediction?
> >
> >     Bruno
> >
> >
> It may be worth investigating. The potential benefit would be largely scene
> dependent. May help in scenes containing hundreds to thousands of objects.
>

This is in such scenes that ray-object intersection is costly. I think
POV-team members are likely to give an opinion. How could such a predictor
work for, say, a tree? Could it be adaptive? Can we imagine rules like:

   While in a csg object, if the previously predicted object is missed by
the ray, then try to traverse only the parent branch, and if still missed,
traverse the whole scene (or try a level further).

   Keep in memory a map of object/ray hits and exploit the info herein,
using local and global statistics.

   Etc ...

My knowledge in that is very poor, and I am absolutely not sure it's a good
idea... But who knows? Perhaps I am lucky with that (for once ;o))!


Bruno

Post a reply to this message