PDA

View Full Version : Metropolis vs Bidirectional Path Tracing...


Nyx2095
04-10-2006, 06:41 PM
Just out of curiosity, do you guys know much about these two algorithms. The inventors of Metropolis light transport claim it is very efficient, but there are also claims that it is impractical. The documentation about it is also very scarce.

Bidirectional path tracing also seems interesting, but it seems like it would be hard to optimize this with any kind of stratified/importance sampling, or even photon mapping.

And just out of curiosity, what algorithm does Maxwell Render use?

playmesumch00ns
04-10-2006, 08:03 PM
Metrpolis light transport is much more efficient than plain bi-directional path tracing. But it is a pain in the arse to implement.

it looks like Maxwell uses some form bi-directional mlt, which another guy is doing here: http://homepages.paradise.net.nz/nickamy/

Nyx2095
04-10-2006, 10:07 PM
Why is MLT so difficult to implement?

I heard it had some weaknesses with getting stuck along "known" paths.

playmesumch00ns
04-11-2006, 12:58 PM
It's just a pain because it's quite complex.Personally, my knowledge of probability mathematics is fairly limited at the moment, so if you have a better grasp of this stuff it might be easier for you.

The traditional MLT algorithm as outlined by Veach and Guibas had problems getting stuck on "peaks" in the integrand since any sample moving away from this peak would be lower, and thus would have a smaller chance of being accepted.

This was addressed in a paper by kelemen et al. that proposes modifying the transfer function to adapt to the size of the current integrand. You can download it here (http://www.cg.tuwien.ac.at/research/publications/2001/Szirmay-2001-METR/). This is the method used in indigo, and i'll bet you it's used in Maxwell too.

nurcc
04-11-2006, 07:02 PM
I know a guy (David Cline) who has implemented MLT, and there are a few downsides. The first is that the original papers are horribly dense, so not much follow-up work has been done. The second is that path tracing normally has high-frequency noise, while MLT normally has low-frequency noise, and is therefore not as suitable for animation. I guess it's because the low-frequency noise is more perceptible in your peripheral vision.

He wrote a paper that explains how to implement a simpler algorithm with similar properties. If you're interested, you can get it at http://rivit.cs.byu.edu/a3dg/publications.php - look for "Energy Redistribution Path Tracing".

playmesumch00ns
04-12-2006, 01:04 PM
Thanks for posting that, I'd completely forgotten about ER!

It's refreshing to read a paper where the authors actually take time to explain things properly, instead of just assuming you understand what they're talking about already :)

nurcc
04-13-2006, 04:50 AM
Yeah, I enjoy his writing too.. he's very much into writing for the implementer.
He's also got an overview of fluid simulators that I think he's trying to get accepted as a SIGGRAPH course, or in the JGT.

Carina
04-15-2006, 07:51 PM
Thanks for posting that, I'd completely forgotten about ER!

It's refreshing to read a paper where the authors actually take time to explain things properly, instead of just assuming you understand what they're talking about already :)

Such papers exist?!? :eek:

This is the problem with these sort of research areas though, the research can be however brilliant, but if you can't express it in a language the people that are going to implement it can understand, it is a struggle to move past academia :hmm:

I remember trying to understand the basics of radiosity and getting stuck on "solid angles". Simplest concept ever, but it took me hours on end trying to figure out what it was because the only literature I could find that covered it was written by mathematicians, for mathematicians (i.e. not for me!:) )

MattTheMan
04-15-2006, 08:22 PM
and, for me, and I'm 13, I need a paper like that for GI (cuz im adding that to my raytracer). Fortunatly, somebody (i forget who) posted an AWESOME monte carlo GI paper.

Carina
04-15-2006, 08:37 PM
and, for me, and I'm 13, I need a paper like that for GI (cuz im adding that to my raytracer).

At least you have the excuse of being only 13 years old. I'm 25 with a degree in Computer Science, and I would like papers like that too damnit;)

MattTheMan
04-15-2006, 09:44 PM
haha lol, well, secretly, I bet everyone does :p

Nyx2095
04-15-2006, 10:15 PM
I often feel things are much better described in terms of implementation than abstract mathematical notation. Because code has some meaning in terms of what happens... So if the meaning of the code is precisely defined, then it's the most meaningful thing you can have.

The math in graphics papers, on the other hand, is often incomplete, abstractly defined, and misleading. It's almost like they don't want you to understand what they mean exactly. If they could at least publish source code for their concept implementation, it would be a great help to everyone.

mindymike
04-15-2006, 11:14 PM
If they could at least publish source code for their concept implementation, it would be a great help to everyone.

They never wouldn't, i think =)

Array
04-16-2006, 08:44 AM
and, for me, and I'm 13, I need a paper like that for GI (cuz im adding that to my raytracer). Fortunatly, somebody (i forget who) posted an AWESOME monte carlo GI paper.

Ask and you shall receive:

http://www.amazon.com/gp/product/1568811470/sr=8-1/qid=1145173433/ref=pd_bbs_1/104-2631140-6763926?%5Fencoding=UTF8

MattTheMan
04-16-2006, 01:33 PM
Yeah, thanks, but I think that photon mapping is above me :p and that I need just some basic monte carlo GI and maybe some path tracing...

I was thinking about geting Physically Based Rendering, but its like $80... ouch

billrobertson42
04-16-2006, 09:39 PM
Might I suggest the library?

Banty
04-16-2006, 10:41 PM
I was thinking about geting Physically Based Rendering, but its like $80... ouch

It's a very good book, the bible of raytracing and monte carlo raytracing, it's a little expensive but it's well written. I think that it's a MUST HAVE. Also there are some free chapters on website. :D
http://www.pbrt.org/downloads.php

If you want to spend less I think that "Realistic Raytracing" is good, the price is half but there are only 235 pages againt 1042 pages. Take a look:
http://www.amazon.com/gp/product/1568811985/102-0690975-5443304?v=glance&n=283155

MattTheMan
04-17-2006, 05:14 PM
Yeah, after all, it is the easter holidays, I have a $30 Barns And Noble's gift card... I think my parents can afford $50 for easter... if they afford licenses for Maya and 3ds Max :p

Nyx2095
04-18-2006, 11:44 PM
Get PBRT. Realistic Image Synthesis Using Photon Mapping is not too clear on the implementation part and doesn't really explain GI in general that well (only a tiny amount of what you need to know to implement photon mapping).

CGTalk Moderation
04-18-2006, 11:44 PM
This thread has been automatically closed as it remained inactive for 12 months. If you wish to continue the discussion, please create a new thread in the appropriate forum.