Page 1 of 1

Building Noise Profile - Percentage Graph confusion..

Posted: Fri Jun 07, 2013 4:20 pm
by Zach
Hi,

I have the general idea of how to build and fine-tune a noise profile for a given video, but I am confused on one issue, and I don't think I saw it addressed in the documentation..

On the graph, there is a percentage range that runs vertical with the graph on the left side.. 0% 100% 200% etc..

I don't understand the relationship between a graph sample and its position on the percentage scale. i.e I can tweak a video and the overall Quality % rating will continue to climb, but I may have values sitting slightly above / below the 0% baseline, with peaks up into 100% or slightly higher.

I could tweak a sample point, and end up with a new value putting it higher or lower in the graph, but the overall percentage rating may not change at all. What am I supposed to be targeting in relation to this percentage scale?

Do I want to have all samples as close to the 0% line as possible? Or the 100% line, or ???

If anyone could explain this, it would be a great help.

Posted: Fri Jun 07, 2013 9:37 pm
by Zach
Here is an image to illustrate what I am referring to.


Image

Posted: Sat Jun 08, 2013 9:27 am
by NVTeam
Re: Do I want to have all samples as close to the 0% line as possible? Or the 100% line, or ???

You don't want that.

The quality indicator takes into account the completeness of the analysis: how many parts of the graph have been actually measured (as compared with the default unmeasured/unknown state or with the interpolated state - when they are estimated on the basis of neighbour nodes).

The quality indicator also takes into account the accuracy of individual measurements. If you use some unsuitable area for fine-tuning then the resulting measurement will be bad and that may reduce the overall quality of the profile.

So when building a noise profile, your goals should be:
(1) to use the right image samples for analysis;
(2) to analyze many samples to cover as many nodes of the graphs with actual measurements (when you select an image area, the graphs show the nodes to be affected by the measurement) as possible;
(3) to not see a high quality figure as a goal in itself but just as an indicator of your progress with (1) and (2).

Measurements can produce a complicated graph with highs and lows, above and below the +0% line. That is normal and not a problem but a reflection of the actual properties of the noise you work with. Usually, those graphs are smooth and do not go to very extreme values like higher than +200%. If they do go there in your measurements, try to re-build the profile using the most noisy sample (select it and click Auto Profile).

Hope this helps,
Vlad

Posted: Sat Jun 08, 2013 5:31 pm
by Zach
Ok, thanks. That makes sense 8)

Posted: Fri Oct 18, 2013 9:54 pm
by Zach
I have a follow up question.

I haven't done a lot of filtering / encoding work since my initial appearance here but I am getting back in to it, with a current project (a new Anime BD set I picked up).

I have noticed with this series in particular, I have a lot of trouble filling in the first R/G/B and last R/G/B sample boxes on the graph with some episodes. Because the series seems to use a lot of complicated lighting, but rarely is there anything to sample that reaches both extreme ends of the spectrum.

Of course it kind of bothers me that I either have to spend 20+ minutes sampling a 24 minute episode, or I either say "good enough" and stop at around 98%.

Speaking in general terms, is there any real detriment to not completing a profile to 100%, if it is the case that you are just lacking in actual footage to sample within the missing ranges?

Posted: Fri Oct 18, 2013 10:10 pm
by NVTeam
There is no need to make it 100% complete by measuring samples. You can simply check individual graphs and manually adjust the unmeasured nodes to make them fit into the general curve formed by the measured ones. Usually that curve is quite smooth and easy to extend at both ends. Manually-set points work the same way measured ones do. So it may be faster to complete profiling this way than finding actual samples for fine-tuning. Yet another even faster way is to use Auto Complete function. It tries to do all that automatically for you, just one click.

Vlad

Posted: Sat Oct 19, 2013 6:10 pm
by Zach
Yes, I've fiddled with auto complete before.


I've also tried to use Auto-Fine Tune, but can't make sense of the thing.
The profile will jump up and down all over the graph and in some cases previous samples become undone and revert to yellow/missing.

Which doesn't quite make sense to me given the way I understand sampling is supposed to work.

Maybe I misunderstand though?
In my view, the expected behavior would be that I initially find a frame that I am satisfied with using Auto-Profile (say 90% baseline) and from there I am expecting that I should be able to use Auto-Complete on other additional frames, to fill in more nodes as it goes, only adjusting their position on the graph in relation to all the previous samples for that specific node. So any particular node should be rising and falling but NOT reverting to an unsampled (yellow color) state, and lowering the overall profile quality....

Basically how manual fine tuning works, except letting the program identify the areas of the frame to take samples from. But that is not the behavior I am seeing.

When I try this, the results could be anything from an improvement in the quality rating, to a drastic drop in quality rating. i.e I could have a 93% Auto-Profile result. Then Auto-Fine tuning would result in a 70% result. Then Auto-Fine Tuning another frame could make it better again or even worse than before. That seems counter-productive?

I thought maybe I was being mislead by the fact it may just be averaging over the total range of samples, but I actually spent the time to jump through an entire 26 minute clip, taking several samples every few hundred frames.

I saw no discernible pattern that made sense, and any high results I occasionally got just seemed like a lucky frame sampling.

So I've just been sticking to manual fine tuning since that's the only way I know to fill my nodes in properly.

Posted: Sat Oct 19, 2013 6:33 pm
by NVTeam
Zach wrote:In my view, the expected behavior would be that I initially find a frame that I am satisfied with using Auto-Profile (say 90% baseline) and from there I am expecting that I should be able to use Auto-Complete on other additional frames
No, Auto-Complete works in another way. It should be used only once when building a noise profile, in the very end of that process. Using it twice or more times will give you the same results as using it only once.

Manual Fine-Tune used with different frames or with the same frame (but with different frame areas) on the other hand can incrementally improve the profile by filling more nodes of the graphs with measured values.

Auto Fine-Tune used with different frames completely re-builds the fine-tuning results based on each new frame. So if a new frame is less representative in terms of noise samples then the fine-tuning results may turn out less complete (fewer nodes of the graph are actually measured) and the overall quality figure may fall. Which is why it makes sense to select the most representative frame giving the largest number of measurements (highest quality figure) and then (if you feel you really need to cover all nodes) you can of course use other frames and manually fine-tune your current profile using samples covering not-yet-measured parts of the graph. Usually that is not really needed if Auto Complete works well or if you manually control/adjust the graphs as seems necessary.
Zach wrote:So I've just been sticking to manual fine tuning since that's the only way I know to fill my nodes in properly.
Which is the right way if you use multiple frames. Or you can simply use manual adjustments, which may be even faster.
Zach wrote:So any particular node should be rising and falling but NOT reverting to an unsampled (yellow color) state, and lowering the overall profile quality....
The rule implemented in Neat Video is the following: if a particular node of the graph is a result of direct measurements of the noise (by Auto Profile, Auto Fine-Tune or Fine-Tune) then it is displayed in "measured" color. If a node is set manually or by Auto Complete then it is marked as "approximate" - yellow color. Please note that Auto Complete never changes the "measured" nodes into "approximate" ones.

Vlad

Posted: Sat Oct 19, 2013 10:20 pm
by Zach
Thanks for clearing that up.

I think I understand things a lot better.

Posted: Sun Oct 20, 2013 2:53 pm
by rimage
I was going to ask a similar question.

Not clear to me from the discussion - what does the position on the vertical axis indicate?

I.e. what does +100% mean as opposed to -100%?

Posted: Sun Oct 20, 2013 3:03 pm
by NVTeam
+0% there is the noise level measured in the main analysis area (the area shown in the image when you click Auto Profile).

+100% means the noise is two times stronger than the noise in the main analysis area: that original level plus 100% extra.

-100% means there is no noise there: 100% of that original level minus 100% gives absolute 0.

Vlad