Effect of -aa parameter on Rpict Rendering Accuracy

Someone with better understanding of the math of it might hopefully chime in, but my understanding has been that the phrase "and approximately represents the error that is associated with it" - which I think is from the manual page - may cause some undue concern. Similar to recommendations for setting the other parameters, if you want test for stable results, instead of jumping from 0.2 all the way to the extreme 0.01, you might try 0.15, 0.125, 0.1, 0.075 until you don't notice any difference in your results from one to the next. I think you mentioned testing various values for -aa, and what change in results did you see? -aa will sometimes have a stronger impact on rendering time than it does on your results depending on the geometric scale of pixels/points you're trying to discern accurately from your image.

A second note, if you're running -aa 0 you probably want -ps 1 and you want to render an image several times larger than your intended result and filter it down to reduce noise. I'm not sure of the numerical accuracy of this technique but I think this is the usual recommendation. If you were using rtrace for point calculations I suppose the noise introduced in -aa 0 is more visual and should be averaged out in the contributions as measured at the point (if I'm thinking about it clearly), but maybe not if using rtrace without the -I parameter?

Hopefully I'm not oversimplifying, and I make my comments hoping someone will correct my misunderstandings and further expand on it.

-Chris

···

From: Kolomenski, Andrei (JSC-SF311)[WYLE INTEG. SCI. & ENG.] [mailto:[email protected]]
Sent: Thursday, August 06, 2015 10:35 AM
To: [email protected]
Subject: [Radiance-general] Effect of -aa parameter on Rpict Rendering Accuracy

Hello Radiance Community,

The -aa input argument to rpict controls the ambient accuracy of the rendering and approximately represents the error that is associated with it. I ran some tests varying the -aa parameter while holding all other input parameters constant. To my surprise, the execution of a rendering with -aa 0 took 1.6 hours and the a rendering with -aa 0.01 has taken 15 hours and has only finished 28%.

What is considered to be the most accurate rendering? One with no ambient interpolation (-aa 0) or a low error interpolation (-aa 0.01) ?

Overall, what rpict input parameters that will produce the most accurate rendering? Currently, I'm using the following parameter settings, for my "truth" renderings that I'm assuming are as accurate as reasonably possible.

General Parameters: -lr 9 -lw 0.0005 -ps 4 -pt 0.05
Specular Parameters: -ss 2 -st 0 -as 256
Direct Parameters: -dr 3 -dp 1024 -ds 0 -dt 0 -dc 1
Ambient Parameters: -ab 11 -aa 0 -ar 256 -ad 2048

Please let me know if you think I'm missing some important parameters that affect the rendering accuracy.

Thank you for your input,
Andrei Kolomenski

____________________________________________________________
Electronic mail messages entering and leaving Arup business
systems are scanned for acceptability of content and viruses

Hi Andrei,

Chris has good advice regarding convergence testing, and the appropriate parameter settings are really dependent on your scene and lighting conditions. That said, there is a table that discusses various rendering parameters and is a good place to start. Unfortunately, it is badly in need of an update:

  http://radsite.lbl.gov/radiance/refer/Notes/rpict_options.html

This same out-of-date text file sits in the ray/doc/notes directory of the standard distribution.

The main thing to be aware of is that changing the -aa setting alone is no guarantee of overall accuracy, since this parameter *only* controls the indirect irradiance interpolation accuracy. If the indirect irradiance values themselves are wrong due to a too-low setting of the -ad parameter (for example), then improving the interpolation won't help you. This goes for the other parameters as well. You can think of each as a link in a chain. It would be great to have an overall accuracy setting, but that would require a much more basic (and necessarily slower) calculation method, such as naive Monte Carlo, to work.

Cheers,
-Greg

···

From: Christopher Rush <[email protected]>
Subject: Re: [Radiance-general] Effect of -aa parameter on Rpict Rendering Accuracy
Date: August 6, 2015 8:12:46 AM PDT

Someone with better understanding of the math of it might hopefully chime in, but my understanding has been that the phrase “and approximately represents the error that is associated with it” – which I think is from the manual page – may cause some undue concern. Similar to recommendations for setting the other parameters, if you want test for stable results, instead of jumping from 0.2 all the way to the extreme 0.01, you might try 0.15, 0.125, 0.1, 0.075 until you don’t notice any difference in your results from one to the next. I think you mentioned testing various values for –aa, and what change in results did you see? –aa will sometimes have a stronger impact on rendering time than it does on your results depending on the geometric scale of pixels/points you’re trying to discern accurately from your image.

A second note, if you’re running –aa 0 you probably want –ps 1 and you want to render an image several times larger than your intended result and filter it down to reduce noise. I’m not sure of the numerical accuracy of this technique but I think this is the usual recommendation. If you were using rtrace for point calculations I suppose the noise introduced in –aa 0 is more visual and should be averaged out in the contributions as measured at the point (if I’m thinking about it clearly), but maybe not if using rtrace without the –I parameter?

Hopefully I’m not oversimplifying, and I make my comments hoping someone will correct my misunderstandings and further expand on it.

-Chris

From: Kolomenski, Andrei (JSC-SF311)[WYLE INTEG. SCI. & ENG.] [mailto:[email protected]]
Sent: Thursday, August 06, 2015 10:35 AM
To: [email protected]
Subject: [Radiance-general] Effect of -aa parameter on Rpict Rendering Accuracy

Hello Radiance Community,

The –aa input argument to rpict controls the ambient accuracy of the rendering and approximately represents the error that is associated with it. I ran some tests varying the –aa parameter while holding all other input parameters constant. To my surprise, the execution of a rendering with –aa 0 took 1.6 hours and the a rendering with –aa 0.01 has taken 15 hours and has only finished 28%.

What is considered to be the most accurate rendering? One with no ambient interpolation (-aa 0) or a low error interpolation (-aa 0.01) ?

Overall, what rpict input parameters that will produce the most accurate rendering? Currently, I’m using the following parameter settings, for my “truth” renderings that I’m assuming are as accurate as reasonably possible.

General Parameters: -lr 9 -lw 0.0005 -ps 4 -pt 0.05
Specular Parameters: -ss 2 -st 0 -as 256
Direct Parameters: -dr 3 -dp 1024 -ds 0 -dt 0 -dc 1
Ambient Parameters: -ab 11 -aa 0 -ar 256 -ad 2048

Please let me know if you think I’m missing some important parameters that affect the rendering accuracy.

Thank you for your input,
Andrei Kolomenski

____________________________________________________________
Electronic mail messages entering and leaving Arup business
systems are scanned for acceptability of content and viruses

_______________________________________________
Radiance-general mailing list
[email protected]
http://www.radiance-online.org/mailman/listinfo/radiance-general