Sony A7RII ISO Invariance

[bjanes]

As mentioned in the article (but no harm is saying it again), the intent wasn't to espouse this approach (and yes, the low ISO / pushed image is worse), but simply to draw the concept to the attention of an audience that may not be as well informed and technical as some on this forum.

Michael

ISO 640 pushed 5 stops is 20480 as shown below. You might repeat the experiment by comparing the 5 stop push of the shot taken at ISO 640 to an exposure at 20480 or the image taken at ISO 12800 to the ISO 640 shot pushed 4 stops.

Regards,

Bill

[AlterEgo]

ISO 640 pushed 5 stops is 20480 as shown below. You might repeat the experiment by comparing the 5 stop push of the shot taken at ISO 640 to an exposure at 20480 or the image taken at ISO 12800 to the ISO 640 shot pushed 4 stops.

Regards,

Bill

also the theory is that Sony starts may be some extra processing for raw data @ ISO25600 (and around - may be 12800 affected somewhat) under some conditions... consider histograms of darkframe shot @ ISO25600 from A7R2, but with different shot parameters (and lenses) = see the difference in green channels :

1)

2)

[Hans Kruse]

In my book, ISO 640 pushed 5 stops equals ISO 20480 (640*2^5)
Am I missing something?

Well for my part I did not even calculate 5 stops up from 640. You are right and therefore it should be no surprise that the pushed ISO 640 looks worse than the ISO 12800. The comparison would have been better done at ISO 25600. Anyway this is old news for the more technical savvy people and still new for most people

[landscapephoto]

By all means, preserve the highlights you're interested in preserving, but you're better off using the analog amplifiers to whatever extent you can than you are amplifying in the digital domain, because, quantization noise.

Quantization is a property of light, because photons. That is the reason, I do not understand what we are talking about here.

Let me give an example. I actually own a truely ISO invariant camera. It is truly ISO invariant, because it does not have an adjustable analog amplifier. The camera is an H4D-50, but other MF cameras work in exactly the same way.

The camera works in the following manner:
-it has a 16 bits ADC, meaning 2^16= 65536 levels
-interestingly, the full well (the maximum number of photons-electrons that fill the pixel) is also around 65000
-therefore the values that are recorded by the camera are something like 0 photon-electron, 1 photon-electron, 2 photons-electrons, 3 photons-electrons, ..., 65534 photons-electrons, end.
-the quantization of the ADC corresponds to the quantization of light.

Just for the record: that camera has poor low-light capabilities, for the following reasons:
-it has poor quantum efficiency, meaning your will need something like 2-3 photons to get one photon-electron
-it has high internal noise, meaning something like 8-16 photons-electrons may be added to each pixel, randomly.

But this is irrelevant to the subject of ISO invariance and quantization noise. And, if we consider that we will get 2^16 photons-electrons at ISO 100, at high ISOs, say ISO 6400, we only get 2^10. That is the problem of high ISOs: there is less light, so there are less photons. So when we are using high ISOs, we will only get values between, say, 1 photon-electron and maybe 1024 photons-electrons. The data is already very coarsely quantized in the photons domain.

So what exactly are we talking about here when we discuss quantization? Is what happens in the digital domain relevant, when the data is already quantized coarsely because of the properties of light?

[amolitor]

Quantization noise means something very specific in digital signal processing.

Suppose you have light levels present in the scene that would, in the digital domain, be rendered as: 0.0, 0.25, 0.3, 0.7, 0.75, 0.9, 1.1 or something.

Since in the digital domain you only have integers, you're going to get something like 0, 0, 0, 1, 1, 1, 1 out. That difference between the analog levels and the digital ones is quantization noise.

Consider amplification. So keep things simple, let's amplify 10x. In reality, increasing exposure is done with a curve, not a constant, and 10x is unrealistic, but the ideas are the same:

Underexpose and push it in post (after the ADC): 0, 0, 0, 10, 10, 10, 10
Increase ISO instead: 0, 2, 3, 7, 7, 9, 11

By pushing in post, you're amplifying quantization noise, which shows up as posterization, and loss of detail in low contrast areas.
 

[ErikKaffehr]

Hi,

On the other hand, the signal is inprecise. We are talking about just a few photons, that is particles of light. So assume 16 photons on average. Natural variation would be something 4 photons (standard of deviation would be 4 photons, by Poisson statistics). Now, let's assume 14 bit representation, that is 16384 values. Let's assume something like 64000 electron full well capacity and one electron per photon. So 16 photons -> 16 electrons. By and large 64000/16000 -> 4, so our 16 photons  would correspond to 12 - 20 photons, natural variation taken into account. The range would result in digital numbers 3-5, and it would be decently resolved.

Es shot noise that is noise in light arriving at each pixel varies it will mask (dither) quantisation noise.

Another point may be that the A7RII is not iso invariant. It makes used of something often called "The Aptina Patent". In modern sensors a small capacitor is connected to each pixel, in order of increasing full well capacity. Having a large FWC is good for noise, and also give low ISO capability. But, a large FWC yields a low output voltage. So Aptina developed and patented a technology to connect the capacitor to the photo diode trough an extra transistor. At high ISO the transistor is closed and FWC is reduced resulting in cleaner readout.

Jim Kasson has analysed this, and found that this is happening at ISO 640 on the A7rII. So he recommends to shoot either 100 or 800 ISO. The reason he uses 800 ISO and not 640 ISO is that 100 to 800 ISO is three full EV steps. Esaier to set 800 ISO on the camera than 640 ISO.

Best regards
Erik



Best regards
Erik

Quantization noise means something very specific in digital signal processing.

Suppose you have light levels present in the scene that would, in the digital domain, be rendered as: 0.0, 0.25, 0.3, 0.7, 0.75, 0.9, 1.1 or something.

Since in the digital domain you only have integers, you're going to get something like 0, 0, 0, 1, 1, 1, 1 out. That difference between the analog levels and the digital ones is quantization noise.

Consider amplification. So keep things simple, let's amplify 10x. In reality, increasing exposure is done with a curve, not a constant, and 10x is unrealistic, but the ideas are the same:

Underexpose and push it in post (after the ADC): 0, 0, 0, 10, 10, 10, 10
Increase ISO instead: 0, 2, 3, 7, 7, 9, 11

By pushing in post, you're amplifying quantization noise, which shows up as posterization, and loss of detail in low contrast areas.

[ErikKaffehr]

Hi,

As a rule, quantisation does not matter as it is masked (dithered) by shot noise coming from the physics of light.

This is an interesting article on some of the issues: http://www.strollswithmydog.com/how-many-bits-to-fully-encode-my-image/

Best regards
Erik

Quantization is a property of light, because photons. That is the reason, I am not I understand what we are talking about here.

Let me give an example. I actually own a truely ISO invariant camera. It is truly ISO invariant, because it does not have an adjustable analog amplifier. The camera is an H4D-50, but other MF cameras work in exactly the same way.

The camera works in the following manner:
-it has a 16 bits ADC, meaning 2^16= 65536 levels
-interestingly, the full well (the maximum number of photons-electrons that fill the pixel) is also around 65000
-therefore the values that are recorded by the camera are something like 0 photon-electron, 1 photon-electron, 2 photons-electrons, 3 photons-electrons, ..., 65534 photons-electrons, end.
-the quantization of the ADC corresponds to the quantization of light.

Just for the record: that camera has poor low-light capabilities, for the following reasons:
-it has poor quantum efficiency, meaning your will need something like 2-3 photons to get one photon-electron
-it has high internal noise, meaning something like 8-16 photons-electrons may be added to each pixel, randomly.

But this is irrelevant to the subject of ISO invariance and quantization noise. And, if we consider that we will get 2^16 photons-electrons at ISO 100, at high ISOs, say ISO 6400, we only get 2^10. That is the problem of high ISOs: there is less light, so there are less photons. So when we are using high ISOs, we will only get values between, say, 1 photon-electron and maybe 1024 photons-electrons. The data is already very coarsely quantized in the photons domain.

So what exactly are we talking about here when we discuss quantization? Is what happens in the digital domain relevant, when the data is already quantized coarsely because of the properties of light?

[AlterEgo]

Esaier to set 800 ISO on the camera than 640 ISO.

yes, something along the lines of using a dial that selects only even ISOs (to be able to do dial it faster vs 1/3 for example) and auto ISO only using even ISO numbers...

[amolitor]

Yeah, you can make quantization noise go away with dithering, and there's quite a bit of natural dither built into the system.

I don't know how it works on the Sony, but if you want to test, find yourself a very low contrast test target of some sort with a fair bit of fine detail. At some point you will see the fine detail "almost vanish" in the sense of the pattern will be vaguely visible in the noise. You'll get a sort of "modulated noise" sort of thing.

If it looks the same either way, great. In the samples I've seen (from other sensors in other cameras) the underexpose+push picture just had mush where the increase-ISO shot had the "modulated noise" appearance, at that limit.

[landscapephoto]

Quantization noise means something very specific in digital signal processing.

I know that, thank you.

But, if memory does not fail, you are a mathematician. You should know the difference between a discrete and a continuous signal. Digital quantization only makes sense when we have a continuous signal as source.

Here we have at source a discrete signal (photons) and out digital levels are actually counting individual photons.

[amolitor]

Yes, yes, I am aware that it's all discrete at the quantum level etc.

I've seen quantization noise in pictures taken with the "underexpose and push" method. It was quite close to the noise floor presented by everything else, to be sure. As you'd expect from a well-designed system. If your quantization noise is NOT on the hairy edge of visible, then you're simply wasting bits, and I am assuming that  the chaps that design these things are not morons.

[amolitor]

I just dug out my Oppenheim & Schafer, and I think I remember again how this goes.

The thing about quantization noise is that it doesn't actually go away, it simply adds on to whatever your dithered-in noise is. This is actually very useful if you are oversampling because it lets you trade sample rate for bit depth.

In effect, that smooth change across the sky which gets posterized can be completely reconstructed with a blur (de-noising).

In the simple model of posterization, you just get a patch of one tone separated from a patch of another tone, and a sharp line between them.

With dither noise, you get that, plus a speckled pattern of noise (which you may consider is "random spots of the darker tone in the lighter region" and "ransom spots of the lighter tone in the darker region") that gets, statistically, darker as the sky got darker. The posterization is still there, it's just hard to see, because the storm of random dots is quite dense at the "sharp line".

Then you apply a de-noiser, and the sky's lovely gradient returns, in all its glorious detail.

This doesn't work for fine detail which was less than one bit in amplitude.

The photon counting argument doesn't work, because perhaps the detail was all tucked in to 12.0 through 13.0 (well, wait, I guess maybe it does, if we assume a completely linear system that's actually counting photons? are there any sensors that actually are?

Let the ADC round it all down to 12 and the detail is simply gone, there's no getting it back.

If you let the analog amplifier do its thing, then you get that 12.0-13.0 range expanded out before the ADC gets its filthy little hands on it, and you can recover some of the detail. Yes, there will also be a lot of noise, since you're probably dealing with various kinds of noise at roughly the same amplitudes as the signals we're dealing with, but your signal will be, in some cases, present and visible. De-noising will, of course, almost certainly destroy it.

All this assumes that there IS an analog amplification step, which isn't at all clear is universally true. There's an incredible amount of speculation and blather out there, though.

It's a corner case, to be sure, but if you want to see if the system is or is not working for you, find some midtone region with very fine detail that's very low contrast, so low contrast that it's only visible as modulated noise. Test it one way, test it the other, if you can see a difference, then you can be sure the analog ISO amplifier system is bringing some value.

[tony field]

ISO Invariance is certainly something to be wished for.   However, many cameras are iso-invariant when the ISO amplification overcomes the read/downstream noise of the sensor system.   For example, my Canon 5D-III and 1D-IV are virtually iso-invariant at/over ISO 1600, and "usefully iso-invariant" at/over ISO 800.  Testing this in a practical sense is trivial.  My testing comparison is approximate based on "observing" the point where there is less  that 1/3 stop improvement  in visual S/N.  If anyone is interested, I can show the results of my 5D-III tests.

This is useful when shooting high ISO scenes.   If I am shooting a scene that may require 1600 to 6400 ISO (say a dance performance) using my preferred exposure, I will park my camera at  ISO 1600 (or even ISO 800) and shoot.   This will give some highlight headroom and avoid blown highlights that might happen if I shot ISO 6400.   The (for me very) minor downside is that chimping images on the camera screen will look dark.

[landscapephoto]

The photon counting argument doesn't work, because perhaps the detail was all tucked in to 12.0 through 13.0 (well, wait, I guess maybe it does, if we assume a completely linear system that's actually counting photons? are there any sensors that actually are?)

We have roughly the same number of photons as we have levels on the camera I was talking about. On modern cameras, we are actually counting a few (that is: less than 10) photons on the low light levels of the picture at high isos. Basically, photon shot noise (which is quantization noise for photons) is about the same level as ADC quantization noise, so you can't really ignore one and say you are only interested in the other.

[hjulenissen]

...Why do we want to do this tests at all? In my world, just expose properly for every shoot and do a good picture. ...

I think one should always try to expose properly - i.e. set exposure time and aperture so as to have sufficient (but not too much) light falling onto the sensor.

The problem is what to do when conditions and artistic ambitions exclude "proper exposure". Perhaps you want significant DOF and not too much motion blur in a dark scene. Then your image sensor is going to be under-exposed.

You might still like to take the picture, but knowing how to deal with it can be a good thing. One solution is to increase the in-camera ISO setting. This works well with my Canon camera up to at least ISO1600. Another solution is to increase the "brightness" slider in your raw developer.

-h

[Stefan12345]

I wonder: If this sensor is really ISO-invariant, wouldn't it then be better if the ISO set in the camera would only be registered as a setting in the RAW file, just like the white balance?

It seems that having the ISO setting applied to the RAW files only results in possible loss of highlight information, with no gains in return.

[Hans Kruse]

I wonder: If this sensor is really ISO-invariant, wouldn't it then be better if the ISO set in the camera would only be registered as a setting in the RAW file, just like the white balance?

It seems that having the ISO setting applied to the RAW files only results in possible loss of highlight information, with no gains in return.

A similar method has been suggested before by Michael for ETTR that a tag was registered and brightness on the LCD was adjusted accordingly. This would also be required with your suggested approach.

[AlterEgo]

I wonder: If this sensor is really ISO-invariant, wouldn't it then be better if the ISO set in the camera would only be registered as a setting in the RAW file, just like the white balance?

this was already done for a long time ... for example some (old) MFDBs and old Sigmas

[AlterEgo]

ISO Invariance is certainly something to be wished for.   

actually no... for as long as we have the same DR @ base ISO I'd prefer less (as much less as possible) then 1 stop DR drop with the every next full stop in gain increase. - I think you 'd prefer that too upon thinking a little bit... that means that if my situation allows or forces me to expose less (no sensel saturation) then I actually wish to trade unused sensel capacity for lesser readout-related noise...

[amolitor]

As long as there is noise that can be overcome with some kind of analog amplification step, ISO invariance isn't a very good idea.

When there is no such noise, you might as well amplify in the digital domain.