Dual ISO - massive dynamic range improvement (dual_iso.mo)

Started by a1ex, July 16, 2013, 06:33:50 PM

Previous topic - Next topic

0 Members and 4 Guests are viewing this topic.

Audionut

Can you describe a use for manually setting the recovery ISO value other then testing?

In real world shooting, SNR settings are perfectly fine for determining the required amount of recovery ISO.  The SNR settings are a fine balance of midtone/shadow SNR requirements vs dual ISO setting.  ie:  Rather then using recovery ISO 1600 just because, by setting the SNR limits based on actual noise level figures, you could shot the scene with a recovery ISO of 400.

Why is this useful?

SNR limits take into account actual noise figures which directly effect the result.
By allowing the automated code to use these values, your scene at recovery ISO 400 vs recovery ISO 1600 has 2 stops better highlight detail, while still retaining the desired amount of SNR in the midtones/shadows.

Let go of any misconstrued conceptions you have about the SNR setting.

Think of SNR limits as DR control (remember highlight control is directly controlled by highlight ignore and is no longer affected by SNR limits (with dual ISO enabled)).  Want more DR (less noise in the midtones/shadows), increase SNR value (which in turn will increase the recovery ISO as needed).

Danne

@audionut. Question. How do I work with the test builds? I used to download Lourencos test builds but there seems to be more going on at this forum? For instance If I would like to try any testing with the dual iso module. Do I download the Audionut test build or do I simply add the dual iso module to my build from Lourenco build which is from august 21?
Thanks guys
//D

Audionut

Looks like the dual_ISO module has made it's way into the new nightly builds  :)

http://builds.magiclantern.fm/#/

Danne


a1ex

Quote from: Audionut on September 12, 2013, 02:54:10 AM
There appears to be a problem with the noise calculation for ISO 6400

Fixed; the noise in top-left black corner was not white, and the blending algorithm was interpreting it as detail (which uses data from both ISO - a lot noisier).

There might be a tiny difference in the output because of slightly different noise values (they are used as thresholds in a few places). Unlikely to be noticeable.

Now I also print the noise level from the high-ISO image (ideally, the noise in very dark areas should reach that value). If it's slightly bigger, it's not really a problem (the noise may sometimes be mistaken for detail); if it's a lot higher (say 1 stop difference or more), it may indicate a bug or a problem in the sensor output (e.g. lots of hot pixels).

Next step: I want to print the SNR values too, in order to check whether ETTR works properly.

Audionut

Can the spot meter sample different areas of the image simultaneously?

a1ex

What's the use case? You can move it around.

There was a demo script with many spotmeters, if you want to try to fix the scripting engine.

Audionut

Using it to print out SNR values for the midtone/shadows on LCD.  Overlayed on the midtone/shadow areas of the image?

RenatoPhoto

I repeated the test to check the snr 8/3 setting.  The maximum ISO ranged achieved was 100-400.

In DUAL ISO  you can see that there is some improvement in noise available above 100-400.  I expected to see less noise at hgher ISO like 100-1600 but did not get it.. I cannot explain this but I had seem lower noise at higher ISOs with previous builds.  But anyway there is room for a bit of noise reduction and the SNR algorithm does not allow.

It is like saying that cannon should only give you ISO 100-1600 instead of 100-25400 because technically there is no use.  But still some people may want to use it, form the usability point of view I am suggesting to expand the range a bit more.

I can imagine that after performing the math, there is not much difference in SNR between the 100-400 and the 100-800, but I see it, so the math should conform to what I see and not to what the numbers tell.  After all we always try to make our images look like what we perceive i.e. higher dynamic range for example.

http://www.pululahuahostal.com  |  EF 300 f/4, EF 100-400 L, EF 180 L, EF-S 10-22, Samyang 14mm, Sigma 28mm EX DG, Sigma 8mm 1:3.5 EX DG, EF 50mm 1:1.8 II, EF 1.4X II, Kenko C-AF 2X

Audionut


RenatoPhoto

Yes, my mistake but the snr 8/4 also had the same 100-400.  I will try again and see if I can get the higher ISO.  As I said the module limits the ISO range so it not always possible to get higher ISO. 
http://www.pululahuahostal.com  |  EF 300 f/4, EF 100-400 L, EF 180 L, EF-S 10-22, Samyang 14mm, Sigma 28mm EX DG, Sigma 8mm 1:3.5 EX DG, EF 50mm 1:1.8 II, EF 1.4X II, Kenko C-AF 2X

RenatoPhoto

Yes, I was correct, I took some pictures and the same thing at 8/4 get iso of 100-400.  The image shows as 8/3 is in fact 8/4
http://www.pululahuahostal.com  |  EF 300 f/4, EF 100-400 L, EF 180 L, EF-S 10-22, Samyang 14mm, Sigma 28mm EX DG, Sigma 8mm 1:3.5 EX DG, EF 50mm 1:1.8 II, EF 1.4X II, Kenko C-AF 2X

RenatoPhoto

Here is the scene, bright sun with deep shadows:

This one is DUAL ISO 100-3200

http://www.pululahuahostal.com  |  EF 300 f/4, EF 100-400 L, EF 180 L, EF-S 10-22, Samyang 14mm, Sigma 28mm EX DG, Sigma 8mm 1:3.5 EX DG, EF 50mm 1:1.8 II, EF 1.4X II, Kenko C-AF 2X

Audionut

Without the SNR 8/4 CR2 to analyze, it's going to be near impossible for a1ex to fix any issues.

My first best guess would be that the area you are showing in your cropped observations above, is below the 5% threshold.

Quote from: a1ex on August 25, 2013, 09:11:48 PM
For shadows, I've chosen the 5% percentile. So, in this context, the shadows having a SNR of 3 EV means 5% of the image pixels have a SNR lower than 3 EV, and the other pixels will be brighter than that.

Keep in mind that 5% is bigger than you may think, since it refers to image area, not linear size. The linear percentage is roughly 22% (1/4.5).

Secondly, I assume those crops are of the extremely dark shadow area in that wood shed?  Would you really tone map that to a brightness level where detail is visible?
Would you tone map that area to a brightness level where the detail difference between 100/400 and 100/800 is visible?

If the answer is yes to the second tone map question, have you observed the highlight detail trade offs?  It's not fair to only observe the shadow areas when highlight detail is also affected by the EV difference between the 2 ISOs.

RenatoPhoto

Quote from: Audionut on September 13, 2013, 11:09:56 PM
Without the SNR 8/4 CR2 to analyze, it's going to be near impossible for a1ex to fix any issues.
Should be able to get to some bandwidth sometime next week.  It seems that this type of condition could be found in many situations where the sun is bright.

Quote from: Audionut on September 13, 2013, 11:09:56 PM
My first best guess would be that the area you are showing in your cropped observations above, is below the 5% threshold.
I kind of feel that I am splitting hairs so the resolution to the issue would be better addressed by allowing the user to simply choose between this type of calculation or not.  I was quite happy with the previous version  and a simple ON-OFF will be great.

Quote from: Audionut on September 13, 2013, 11:09:56 PM
Secondly, I assume those crops are of the extremely dark shadow area in that wood shed?  Would you really tone map that to a brightness level where detail is visible?
Would you tone map that area to a brightness level where the detail difference between 100/400 and 100/800 is visible?

I have already brighten the shadows to expose the noise difference.  Although there is less noise at higher ISO combinations it is hardly noticeable.  In retrospect the level of noise given the super dark are is supper acceptable at 100-800 or 100-400.

Quote from: Audionut on September 13, 2013, 11:09:56 PM
Secondly, I assume those crops are of the extremely dark shadow area in that wood shed?  Would you really tone map that to a brightness level where detail is visible?
Would you tone map that area to a brightness level where the detail difference between 100/400 and 100/800 is visible?

If the answer is yes to the second tone map question, have you observed the highlight detail trade offs?  It's not fair to only observe the shadow areas when highlight detail is also affected by the EV difference between the 2 ISOs.
In my opinion there was very little issue with the highlights, the algorithm was working well and keeping overexposure to the 0.2% h.i, so there is no tradeoff.  This kind of result is achievable thanks to DUAL ISO.

-----------

Finally and most importantly

I shoot a lot of wildlife and the speed of congruence is very crucial; now when the modules are linked the calculations have become slower so more shots are required.  This also is another reason to leave AETTR do its thing and the add DUAL ISO just on top of AETTR as it was before with +-XEV option for Recovery ISO.  I had become very happy with setting my AETTR at 4/2 and DUAL ISO at +3EV and shooting without too many delays.  Now I cant do this, I just takes one or two or three more shots to get optimum results.

And finally dont take me wrong the images are great and the technology is great, but please put an ON-OFF switch for the more precise calculations.  Please!

http://www.pululahuahostal.com  |  EF 300 f/4, EF 100-400 L, EF 180 L, EF-S 10-22, Samyang 14mm, Sigma 28mm EX DG, Sigma 8mm 1:3.5 EX DG, EF 50mm 1:1.8 II, EF 1.4X II, Kenko C-AF 2X

Audionut

Quote from: RenatoPhoto on September 14, 2013, 12:20:11 AM
I have already brighten the shadows to expose the noise difference.  Although there is less noise at higher ISO combinations it is hardly noticeable.  In retrospect the level of noise given the super dark are is supper acceptable at 100-800 or 100-400.

So ideally, the fundamental code base is sound.  And you're just splitting hairs.


Quote from: RenatoPhoto on September 14, 2013, 12:20:11 AM
I shoot a lot of wildlife and the speed of congruence is very crucial; now when the modules are linked the calculations have become slower so more shots are required.  This also is another reason to leave AETTR do its thing and the add DUAL ISO just on top of AETTR as it was before with +-XEV option for Recovery ISO.  I had become very happy with setting my AETTR at 4/2 and DUAL ISO at +3EV and shooting without too many delays.  Now I cant do this, I just takes one or two or three more shots to get optimum results.


I have noticed that it's ability to converge quickly is directly related to the initial exposure settings.  Start off overexposed and expect problems (especially with the following).
Scene conditions and expectations play a large part.  In a scene with a very large DR where you have no, or very little highlight clipping, and you have the SNR settings @ 8/4, you're asking quite a bit.

Relax your expectations of what should be achievable, perhaps use SNR 7/3 instead, and convergence is met much faster.

Having said all that, I have mentioned,

Quote from: Audionut on September 11, 2013, 06:19:53 AM
Or maybe a setting in dual ISO, "link to AETTR".  Else dual ISO is always enabled at the settings specified.

I guess then a1ex would have to ignore any SNR settings applied (just consider highlight ignore and leave the user to make his own mistakes).  That sounds ugly and isn't something I would expect a1ex to be open to.
Or perhaps fully disable SNR settings when dual ISO is not linked to AETTR.

For the absolute fastest results, you should look at AutoExpo/Dual ISO combination.  There are no LV calculations.  Point and shoot at burst speed :)
You lose some highlight precision sure, but lets face it, 0.5EV below FWC and 1.0EV below FWC (with the recover ISO taking up the slack), is probably splitting hairs also.

RenatoPhoto

Quote from: Audionut on September 14, 2013, 12:55:46 AM
So ideally, the fundamental code base is sound.  And you're just splitting hairs.

These are my observations, I cannot analyze code.

Quote from: Audionut on September 14, 2013, 12:55:46 AM
I guess then a1ex would have to ignore any SNR settings applied (just consider highlight ignore and leave the user to make his own mistakes).  That sounds ugly and isn't something I would expect a1ex to be open to.
Or perhaps fully disable SNR settings when dual ISO is not linked to AETTR.

Just give the user a choice between slow converging exact process or a quick process that produces excellent results. 

We know that DUAL ISO should be only used in scenes where the dynamic range is greater then what the camera can capture.  So that means we are going to capture the highlights as well as the shadows.   We have also determined that h.i. controls the highlights and so to increase dyamic range we protect the highlights with h.i. and capture the shadows with DUAL ISO.  That is what I have been doing before.  The SNR numbers give me a scale of shadow recovery which is very complicated 32 different options while DUALISO gives me 6 options.  For this reason, when I was using AETTR+DUAL ISO I was leaving SNR OFF and then enabling the DUAL ISO at a reasonable number (ie +3EV) so I could capture the details in the shadows at acceptable noise level.

In the scenes where the dynamic range is not great, or in scenes where I am not concerns with the highlights, the I USE SNR 4/2 and 6/3 as I see fit based on experience.  This has always given me good results.  I even tested this in Timelapse with very good results see here:

Dual ISO vs Auto ETTR
http://www.magiclantern.fm/forum/index.php?topic=7835.0

This new system took away everything that I like and replaced it with a better, more accurate, splitting hairs system, which I may want to use from time to time when time allows and I can take a few shots at a static scene.

Quote from: Audionut on September 14, 2013, 12:55:46 AM
For the absolute fastest results, you should look at AutoExpo/Dual ISO combination.  There are no LV calculations.  Point and shoot at burst speed :)
You lose some highlight precision sure, but lets face it, 0.5EV below FWC and 1.0EV below FWC (with the recover ISO taking up the slack), is probably splitting hairs also.

Interesting, I will have to look into this option.  I tried it a couple of times but so many freaking curves that I could not make scense of what this thing is doing.  It needs a user manual for dummies like me.
http://www.pululahuahostal.com  |  EF 300 f/4, EF 100-400 L, EF 180 L, EF-S 10-22, Samyang 14mm, Sigma 28mm EX DG, Sigma 8mm 1:3.5 EX DG, EF 50mm 1:1.8 II, EF 1.4X II, Kenko C-AF 2X

a1ex

Quote from: RenatoPhoto on September 13, 2013, 01:27:38 PM
I can imagine that after performing the math, there is not much difference in SNR between the 100-400 and the 100-800, but I see it

I can barely see any difference between 100/400 and 100/6400. Maybe you can use a higher contrast for that area?

Maybe I should check the standard deviation of the noise in some out-of-focus area.

Danne

@Audionut
"For the absolute fastest results, you should look at AutoExpo/Dual ISO combination.  There are no LV calculations.  Point and shoot at burst speed :)
You lose some highlight precision sure, but lets face it, 0.5EV below FWC and 1.0EV below FWC (with the recover ISO taking up the slack), is probably splitting hairs also."


Been fiddling with the Auto exposure settings and it seems to work very well with dual iso. My settings right now pointing out the windows seems to give consistent good results;

AV range, I used f5.6-f5.6 but free of choice of course
TV minimum 1/44, free of choice of course
EC range -1.5 EV- +3.0 EV try this one out a little
EC +3.0 EV probably the most important setting to change. Try +2.0 EV or til you get it as you want
ISO range 100-3200

The big plus here is you don,t need to open up liveview to get a reading as with Aettr. The low iso, though, has to be changed manually from within dual iso.
Gonna try some more...

*After developing I can only say wow. This autofeature is exactly what I,ve been looking for. If an automated feature were implemented to disable dual iso when in not so contrasty areas and maybe also have an auto setting for the lower iso setting I could finally shut of my brain and start taking pictures ;)
*So, to simplyfi. Set auto exposure to overexpose by two or three steps, then hit dual iso to get the blown highlights back. Am I right?

//D

a1ex

Quote from: Danne on September 14, 2013, 10:51:17 AM
Set auto exposure to overexpose by two or three steps, then hit dual iso to get the blown highlights back. Am I right?

Only for your particular test scene  ;)

That's because auto expo meters for midtones and ignores highlights. Try on a blank wall, then try pointing it out the window. I just did on 5D3: on the blank wall it underexposed by 2-3 stops, on the window it overexposed by 3-4 stops. I've tried evaluative and center weighted, could not see any difference between them.

QuoteIf an automated feature were implemented to disable dual iso when in not so contrasty areas...
That's exactly what auto ETTR does (and Renato is reporting it as bug).

Danne


a1ex

QuoteOr maybe a setting in dual ISO, "link to AETTR".

What about a setting like this in ETTR?

Link to Dual ISO
- OFF (old way)
- SNR priority (as it's now)
- Highlight priority (similar to SNR, but if dual ISO reaches 100/max, ignore the SNR and capture all the highlights).

I'm still working on convergence issues; with ETTR in LiveView (on request via SET button), convergence is quite fast and you don't have to waste any test shots.

The biggest problem when metering outside LV is that current heuristics (simple and fast) sometimes get tricked because of black level difference. With a black difference of around 30 (typical for ISO 100/6400), the metering error can be up to 4 stops (!).

In LiveView, ETTR is metering on only one of the two exposures (no black difference), and estimates the other exposure from DxO data. That's why it works a lot better than metering outside LV.

Another cause of non-convergence is that 1600/3200/6400 are almost identical, so the algorithm may switch from one to another just because of round-off errors. This one should be easy to fix.

RenatoPhoto

Quote from: a1ex on September 14, 2013, 11:36:48 AM
That's exactly what auto ETTR does (and Renato is reporting it as bug).

Hi Alex,
I like the feature very much, I like the concept of using DUAL ISO only when the dynamic range requires of its use.  ;)  Please dont take it out, it has its uses.
I find the implementation a bit difficult to use with ease and speed. I find the final hair splitting a bit inconsistent and difficult to obtain.

I am suggesting in order of importance:

1. Implement an ON OFF switch for the more robust and automated calculations.
2. When the switch is in OFF position, and DUAL ISO is enabled set h.i.=0.2%, SNR=OFF, have AETTR to its thing, have DUAL ISO take the Base ISO and add the user selected Recovery ISO via +-XEV
3. If user has switch in OFF position, and DUAL ISO is enabled and the user wants to use other settings for SNR or h.i., then let him but warn him that these settings work best with the automated version.  But I suppose there is no sense in setting h.i=5% when using DUAL ISO, it defeats its use since you no longer require DUAL ISO to capture the whole dynamic range.  The SNR settings could be fiddle with as long as the highlights are not clipped.  If the user boosts SNR and the scene is overexposed then warn the user.
3. When the switch is ON your calculations take over, the SNR is optimized via combination of AETTR and DUAL ISO. 

I see you posted above while I was writing..

http://www.pululahuahostal.com  |  EF 300 f/4, EF 100-400 L, EF 180 L, EF-S 10-22, Samyang 14mm, Sigma 28mm EX DG, Sigma 8mm 1:3.5 EX DG, EF 50mm 1:1.8 II, EF 1.4X II, Kenko C-AF 2X

a1ex

Quote from: RenatoPhoto on September 14, 2013, 01:47:16 PM
But I suppose there is no sense in setting h.i=5% when using DUAL ISO, it defeats its use since you no longer require DUAL ISO to capture the whole dynamic range.

Sometimes you do. I had to do this when taking photos in a cellar - there was a bright window (something like this - not my picture, but a similar scene), and it was impossible to have detail there and see something in the cellar, so I had to ignore it. Of course, dual ISO maxed out.

Also street lights: they are usually so bright that dual ISO can't cover them, and you have to ignore them (otherwise you may get a lot of noise).

RenatoPhoto

Quote from: a1ex on September 14, 2013, 01:14:23 PM
What about a setting like this in ETTR?

Link to Dual ISO
- OFF (old way)
- SNR priority (as it's now)
- Highlight priority (similar to SNR, but if dual ISO reaches 100/max, ignore the SNR and capture all the highlights).

I am sure that your fingers will do the magic!  ;D
http://www.pululahuahostal.com  |  EF 300 f/4, EF 100-400 L, EF 180 L, EF-S 10-22, Samyang 14mm, Sigma 28mm EX DG, Sigma 8mm 1:3.5 EX DG, EF 50mm 1:1.8 II, EF 1.4X II, Kenko C-AF 2X