Path Space Regularization (aka the solution to SDS paths)

Discussion related to the Engine functionality, implementations and API.
provisory
Posts: 224
Joined: Wed Aug 01, 2018 4:26 pm

Re: Path Space Regularization (aka the solution to SDS paths)

Post by provisory » Thu Sep 19, 2019 4:48 pm

I think the paper says about the same what I wrote, just a little vaguely worded:
Our proposed method combines the simulated annealing optimization method with MCMC integration. Intuitively, the “sharp peaks” of the integrand are regularized (smoothed out) in the beginning of the integration such that the Markov chain can easily find them. In order to achieve consistency, the regularization is reduced throughout the integration.
http://cg.ivd.kit.edu/publications/p201 ... n_2013.pdf

User avatar
Dade
Developer
Developer
Posts: 3276
Joined: Mon Dec 04, 2017 8:36 pm
Location: Italy

Re: Path Space Regularization (aka the solution to SDS paths)

Post by Dade » Fri Sep 20, 2019 11:15 am

provisory wrote:
Thu Sep 19, 2019 4:48 pm
I think the paper says about the same what I wrote, just a little vaguely worded:
Our proposed method combines the simulated annealing optimization method with MCMC integration. Intuitively, the “sharp peaks” of the integrand are regularized (smoothed out) in the beginning of the integration such that the Markov chain can easily find them. In order to achieve consistency, the regularization is reduced throughout the integration.
http://cg.ivd.kit.edu/publications/p201 ... n_2013.pdf
They talk of "throughout the integration" or "mutation number" (without large or small qualifier). It is very generic and unclear.

Anyway, I have gone back to my initial idea of using a mollifies count for each pixel: this has several benefices. The "angles" will remain "large" until when a path affecting the pixel is not found, it will also shrink according the single pixel samples count, not the overall image samples count.

It seems the best compromise at the moment:

psr.jpg
Support LuxCoreRender project with salts and bounties

User avatar
Sharlybg
Donor
Donor
Posts: 1530
Joined: Mon Dec 04, 2017 10:11 pm
Location: Ivory Coast

Re: Path Space Regularization (aka the solution to SDS paths)

Post by Sharlybg » Fri Sep 20, 2019 11:32 am

Look like you render the pool always without sky diffuse light. only sun. i think HDRI or sun + Sky can make the pool more clear.
Support LuxCoreRender project with salts and bounties

Portfolio : https://www.behance.net/DRAVIA

provisory
Posts: 224
Joined: Wed Aug 01, 2018 4:26 pm

Re: Path Space Regularization (aka the solution to SDS paths)

Post by provisory » Fri Sep 20, 2019 11:56 am

Dade wrote:
Fri Sep 20, 2019 11:15 am
It is very generic and unclear.
I agree.

provisory
Posts: 224
Joined: Wed Aug 01, 2018 4:26 pm

Re: Path Space Regularization (aka the solution to SDS paths)

Post by provisory » Fri Sep 20, 2019 12:07 pm

Did you examine their sample implementation?
http://cg.ivd.kit.edu/publications/p201 ... 3_supp.pdf

I have too little knowledge for this, unfortunately.

User avatar
lacilaci
Donor
Donor
Posts: 1637
Joined: Fri May 04, 2018 5:16 am

Re: Path Space Regularization (aka the solution to SDS paths)

Post by lacilaci » Fri Sep 20, 2019 12:27 pm

starting to look good. (except the overall darkness underwater)


User avatar
epilectrolytics
Donor
Donor
Posts: 578
Joined: Thu Oct 04, 2018 6:06 am

Re: Path Space Regularization (aka the solution to SDS paths)

Post by epilectrolytics » Fri Sep 20, 2019 1:11 pm

There's also "Hierarchical Russian Roulette for Vertex Connections", though it is implemented in BiDir and may not work with Hybrid.
MBPro 15" 16GB i7-4850HQ GT750M, MacOS 10.13.6 & Win10Pro PC 16GB Ryzen 2700X, 2 x RTX 2070

Vutshi
Posts: 15
Joined: Tue Jul 17, 2018 8:24 pm

Re: Path Space Regularization (aka the solution to SDS paths)

Post by Vutshi » Fri Sep 20, 2019 1:25 pm

Dade wrote:
Fri Sep 20, 2019 11:15 am
They talk of "throughout the integration" or "mutation number" (without large or small qualifier). It is very generic and unclear.
This statement is pretty clear to me as it comes from the field of many-body physics where the simulated annealing is used extensively. The general idea is to start from a "hot" system and let the MC equilibrate such that it samples the correct distribution at the given "temperature" (mollification). Then we gradually reduce the temperature slow enough to let MC adjusts itself to the new distribution at each moment. This is why there is a particular speed of temperature drop (like gamma^n for mollification) which should not be exceeded and which follows from the Metropolis rate of convergence.

In the end it should not matter what is n - a small mutations number or large mutations number, since they are proportional to each other.

BTW, I tried the smallpt code from the Supplemental Material and it is broken :). After fixing it seems to work but produces a bit more fireflies as compared to the figure in the SM. It does not darken as far as I can see.
Last edited by Vutshi on Fri Sep 20, 2019 1:45 pm, edited 1 time in total.

User avatar
Sharlybg
Donor
Donor
Posts: 1530
Joined: Mon Dec 04, 2017 10:11 pm
Location: Ivory Coast

Re: Path Space Regularization (aka the solution to SDS paths)

Post by Sharlybg » Fri Sep 20, 2019 1:42 pm

lacilaci wrote:
Fri Sep 20, 2019 12:27 pm
starting to look good. (except the overall darkness underwater)
what about sun+sky or HDRi
Support LuxCoreRender project with salts and bounties

Portfolio : https://www.behance.net/DRAVIA

Post Reply