🔒 Locked

Estimating imprecise probabilities

Started by GameCat Jun 23, 2004 at 4:36 PM 3 replies 1.5k views

Original Post

Jun 23, 2004 at 4:36 PM

In this old thread (I'd post there and revive it, but it doesn't work, probably because of the forum upgrade) I got help on how to estimate imprecise probabilities. When I help someone with something I always think it's kind of annoying not to find out what the final solution turned out to be so heres my update :) This was done for my Master's thesis "Optimal occlusion testing" and I was trying to optimise occlusion culling by only performing the occlusion test for objects likely to be hidden. Actually it was a little more complicated since there was a utility function involved but that was the basic idea. I ended up using a Markov model due to it's computational efficiency and decent theoretical grounding in this particular case. To my (and my supervisors ;) great joy this technique did in fact offer a speed up compared to other heuristics i tested. I learned lots about imprecise probabilities along the way, coherent lower probabilities and interval probabilities among other things, but in the end they actually weren't suitable for my particualr problem. Anyway, thanks for the help, especially Timkin that got me exploring using a Markov model. If you're interested (though it's really more graphics than AI related) the thesis is available online here Optimal occlusion testing.

Timkin

Jun 23, 2004 at 7:54 PM

Fantastic GameCat!

It's great to see someone successfully solving their problem. I took a quick read through your thesis and it looks like you did some valuable work. I was particularly intrigued by the result that the pseudo-naive SkipN strategy performed so well on the more complex scences. I'm not surprised that you found a performance increase from your Markov methods... the computational cost of performing the model update and query is low compared to the expected cost from stalled culling queries.

Do you have any plans to follow up this work or are you looking in a different direction now? Considering (or have you started) a PhD? ;)

Cheers,

Timkin

GameCat

Jun 24, 2004 at 4:05 AM

Thanks!

Actually, the SkipN strategy turns out to be a special case of the Markov strategy (when occlusion testing cost for all objects is constant over time rendering and occlusion testing loss is constant over all objects), so the fact that they all perform similarly isn't that strange. And it makes a good case for the elegance of the Markov models as well :)

Currently I'm looking for work and not pursuing a PhD (although I will be applying for at least one "research engineer" type job). While there are a number of very smart AI guys at my institution and my supervisor has published extensively within decision theory, there really are no graphics guys and very little graphics research done in Sweden at all. Thomas Akenine-Möller has a couple of doctoral students but he really is one of the few graphics researchers in Sweden that I know about (there are some at the Royal Institute of Technology in Stockholm as well, but they're very usability oriented AFAIK). There are some biotech related visualisation jobs i might get though, so we'll see...

/Joakim

Timkin

Jun 24, 2004 at 7:42 PM

Have you considered changing countries to do a PhD?

GameCat

Jun 27, 2004 at 6:10 AM

Well, yes, but I'd rather stay in Sweden until my girlfriend graduates (new years) and then have her tag along. But if the research is interesting enough, sure. I doubt any gold plated offers will magically drop into my inbox though, so while I'm looking I'm applying for a number of jobs both here in Stockholm and abroad.

Topic Locked

This topic has been locked by a moderator. New replies are not allowed.