Multidimensional Adaptive Sampling and Reconstruction ?

[damoguyan]

Anybody is trying or planning to implement the algorithm in the paper "Multidimensional Adaptive Sampling and Reconstruction for Ray Tracing"?
I'm expecting the feature in luxrender, and want to involve the interesting implementation work.

[ktegan]

MDAS is a great method. Another possibility is Adaptive Wavelet Rendering. I would expect that each method may be preferable in different situations, but the AWR algorithm has a few advantages over MDAS:

- ease of implementation, you don't have to change how your integrator samples higher dimensions such as lens position, light sample position, and time using AWR. AWR only uses x, y and radiance values for reconstruction and adaptive sampling
- MDAS reconstruction is slower, at least in the reference implementations provided by the MDAS authors (this makes sense because it's working with high-dimensional kd-Trees)
- MDAS has a high memory consumption (in my paper, "A Frequency Analysis and Sheared Reconstruction for Motion Blur" we found MDAS required 1GB of memory to render a 512 × 512 image with 8 samples per pixel when using the author's implementation). In contrast AWR reports 88 MB memory for a 1k image.
- the "curse of dimensionality" starts to hurt MDAS. From the AWS paper: "The adaptive sampling portion of the algorithm provides diminishing returns as the dimensionality increases (as predicted by Bahvalov’s theorems) and the computational cost of the signal reconstruction stage is exponential in the dimensionality".

http://graphics.cs.berkeley.edu/papers/ ... index.html

Note: I am biased because I worked with Ryan and I think his paper is awesome. Also MDAS is a really cool technique that is very general, and captures a huge number of effects, and works on "the hard parts" of the image in a very elegant and general manner. I don't want to take anything away from their work, and I certainly haven't stress tested both systems to see where one might have a relative advantage over the other.

Multi-threaded versions of either of these will require some modifications.

Kevin

[Lord Crc]

*droool*

[jeanphi]

Hi,

Thanks Kevin for this most informative contribution. I was looking after this AWR paper.
When reading the MDAS paper back when it was published, I also feared that its memory consumption would be impractical for a general purpose renderer.

Jeanphi

[Dade]

Thanks Kevin for this most informative contribution. I was looking after this AWR paper.
When reading the MDAS paper back when it was published, I also feared that its memory consumption would be impractical for a general purpose renderer.

Didn't we talked about the MDAS paper when it was first publish ? ... searching on forums ... yup, viewtopic.php?f=8&t=452&p=4978&hilit=Multidimensional+Adaptive+Sampling+and+Reconstruction+for+Ray+Tracing#p4978

Quoting myself: "Isn't the memory required a bit high ? 400MB for a 1024x1024x1 sample image sounds like a lot. Anyway memory is cheaper than CPU cycles and the results are awesome."

[jeanphi]

Hi,

I had a quick glance at the AWS paper and it looks much more interesting than MDAS: part of it could even be used as noise reduction for all other parts of Lux. The implementation details are a bit scarce though so it will require a more thorough reading to implement.

Jeanphi

[jeanphi]

Hi,

The only drawbacks I see to this method are:
- samples are purely random, maybe it could use some stratification
- it will be quite inefficient for bidirectional (for the light paths directly connected to the eye most notably)
But it should be pretty efficient for animations.

Jeanphi

[ktegan]

The only drawbacks I see to this method are:
- samples are purely random, maybe it could use some stratification
- it will be quite inefficient for bidirectional (for the light paths directly connected to the eye most notably)
But it should be pretty efficient for animations.

Light paths to the eye will land in places that you can't control, so it hurts the adaptive sampling part of the algorithm, but the smooth reconstruction will still be fine, yes? Stratification might help (with / without comparisons would be pretty interesting), but the main reason for stratification is to reduce noise, and if your reconstruction makes nice smooth images it may not be as crucial.

Kevin

[ktegan]

I talked with Ryan. He said that he's happy to help people with questions and he might be able to give out some parts of the source code that are mostly closely related to the paper.

I don't want to write post his email address in the forum, but I think a determined person could find it (COUGHresumeCOUGH).

Kevin

[damoguyan]

I talked with Ryan. He said that he's happy to help people with questions and he might be able to give out some parts of the source code that are mostly closely related to the paper.

I don't want to write post his email address in the forum, but I think a determined person could find it (COUGHresumeCOUGH).

Kevin

Hi, kevin
Thank you for useful information. I'm trying to implement some interesting research work into luxrender because i like it and hope it become stronger. As far, so regretful that nobody have the same idea as i. Therefore i plan to begin with some work after Spring Festival when i buy a high-performance desktop. By the way, i have read about your paper about frequency analysis of motion blur, and it is a great work. You are so nice to share your work.