ISO Invariance

[ImageMatters]

Have any of you looked into the claims of some camera manufacturers that their cameras are 'ISO invariant'? It appears to be equivalent to saying that an under-exposed image need not be boosted in camera but can be recovered in ACR later on. We've all taken advantage of this by boosting shadows in ACR so this is nothing new. What they appear to be saying is that some cameras have such a low noise floor that you can get away with multiplying the sampled brightness values after ADC conversion without unduly bringing out noise in the shadows.

How true is this?
Do we still need to dial in ISO?

As a former electronics engineer, I can't quite go along with these claims 100%. All sensors will be limited by shot noise due to the variation in the numbers of electrons collected by each pixel which is a function of quantum efficiency and incident flux density. Then there is thermal noise, dark current noise and read noise before conversion from analogue to digital. In any system, the contribution of noise is worst at the start of the signal path where the signal is weak and so it makes sense to apply amplification early in the signal chain. If you wait until later (after A/D conversion) you are effectively amplifying the contribution of read noise and quantisation noise. Boosting the sampled pixel values at this stage brightens the image but does not increase the signal-to-noise ratio (SNR).

I use a 5D Mk III and it has good low-light performance because the pixels are large and the sensor electronics are reasonably quiet. The signal from each pixel is amplified locally so the ISO setting is only amplifying the effects of shot noise and dark current (with a minimal contribution from thermal noise in the amplifier itself) and therefore the signal plus some noise is greater when the effects of read noise and quantisation noise are taken into account. For this reason, I think it's best to adjust the ISO (as far as this controls local amplifier gain and not just multiply sampled values in software) rather than rely on a low noise floor and boosting the signal plus noise in post.

Theory aside, I'd be interested to hear about anyone who has been pushing under-exposed images from so called ISO invariant cameras. Do you think the quality is good enough or will we be setting ISO ourselves for a while longer?

[Dave Humphries]

Interesting question Dave,

Welcome to the CiC forums from me (another former electronics engineer).

I understand what prompts the question, but wonder also whether modern electronics, which achieved large SNR in analogue audio electronics, when applied to the (arguably restricted dynamic range of) video, means the effects are less relevant?

I dunno, just a thought - and not one I can back up with dB figures without possibly making myself look silly, or spending time researching. e.g. Given that audio A-D are 16 or 24 bit and still cameras only about 14 bit - and when you come to display it, the most you're likely to need is say, 6-7 bits (aka 6 stops DR).

Cheers, Dave

[xpatUSA]

Have any of you looked into the claims of some camera manufacturers that their cameras are 'ISO invariant'? It appears to be equivalent to saying that an under-exposed image need not be boosted in camera but can be recovered in ACR later on. We've all taken advantage of this by boosting shadows in ACR so this is nothing new. What they appear to be saying is that some cameras have such a low noise floor that you can get away with multiplying the sampled brightness values after ADC conversion without unduly bringing out noise in the shadows.

How true is this?
Do we still need to dial in ISO?

<theory>

Theory aside, I'd be interested to hear about anyone who has been pushing under-exposed images from so called ISO invariant cameras. Do you think the quality is good enough or will we be setting ISO ourselves for a while longer?

Welcome!

All my Sigma cameras are truly "ISO-invariant". Select more ISO and the captured raw data will have proportionally lower levels. That works as follows:

Instead of there being a PGA or VGA conventionally between the sensor and the ADC, the ISO selection is passed to the converter as meta-data in the raw file and said converter applies the appropriate multiplication factor to the raw data during conversion.

Because the Foveon's three sensor outputs are video-amplified on-chip, their outputs are up around 500mV so there is no real SNR advantage to be gained by using the conventional system. Plus, when you look at the raw histogram, it is pretty obvious if the sensor was under-exposed (never a good thing with the Foveon, I guarantee) - unlike conventional raw data.

Sigma did try a few models with the conventional AFE system (cost reduction? heat/current reduction?) but users soon complained about "lack of headroom" and "blown highlights" and are still doing so with the AFE-equipped Quattro models.

Apropos of "pushing under-exposed" captures: if I do that by applying negative EC in-camera or actually selecting a higher ISO (heaven forbid), I get my just reward - more noise/blotching and/or less DR.

So, for Sigmas, the quality is still not good enough and my ISO stays at 100 permanently. For modern Bayer systems, I'd guess that the quality is perfectly acceptable.

[xpatUSA]

- and when you come to display it, the most you're likely to need is say, 6-7 bits (aka 6 stops DR).

Sorry, Dave, here I am again. Equating DR to the number of bits is technically incorrect, but of course depending on what the definition of "DR" is (there are several).

When thinking only in bit depth, say 8-bit, then the stumbling point is that a range of 0-255 has infinite SNR/DR even tho' zero is a valid number in that range. Specifically, 255/0 = infinity - neither acceptable in logarithmic calculations nor in the Real World Going further about that, suppose our fancy new 14-bit camera has a noise level of 30/16383? That would be an SNR of about 9 EV. Now suppose that appears in 8-bit form as 0 or 1? SNR is now either infinity or 8 EV, a whole EV somehow lost . . .

When thinking of the display on the other hand, zero does give you a finite brightness (as does 255). So the display can have say a contrast of 1000, e.g. 255 = 250 cd/m2 and 0 = 0.25 cd/m2, and that would be 60dB DR or 10 stops - both of which are far more than e.g. 255/1.

[Dave Humphries]

Sorry, Dave, here I am again. Equating DR to the number of bits is technically incorrect, but of course depending on what the definition of "DR" is (there are several).

I agree, it's a fudge, but I don't think we need to go in to numbers and noise floors, etc. it was the general principle and comparison with audio I was trying to make.

If you assume linear processing and coding in (and before) the A-D converter, each bit is a doubling of signal amplitude and a doubling of signal amplitude is a doubling of light, which is 'a stop' of light gain (or DR = Dynamic Range).

Very, very imprecise - as is that statement

[xpatUSA]

I agree, it's a fudge, but I don't think we need to go in to numbers and noise floors, etc. it was the general principle and comparison with audio I was trying to make.

If you assume linear processing and coding in (and before) the A-D converter, each bit is a doubling of signal amplitude and a doubling of signal amplitude is a doubling of light, which is 'a stop' of light gain (or DR = Dynamic Range).

Very, very imprecise - as is that statement

Almost makes me want to go out and buy a 16-bit camera, Dave . .

[dje]

Have any of you looked into the claims of some camera manufacturers that their cameras are 'ISO invariant'? It appears to be equivalent to saying that an under-exposed image need not be boosted in camera but can be recovered in ACR later on. We've all taken advantage of this by boosting shadows in ACR so this is nothing new. What they appear to be saying is that some cameras have such a low noise floor that you can get away with multiplying the sampled brightness values after ADC conversion without unduly bringing out noise in the shadows.

How true is this?
Do we still need to dial in ISO?

As a former electronics engineer, I can't quite go along with these claims 100%. All sensors will be limited by shot noise due to the variation in the numbers of electrons collected by each pixel which is a function of quantum efficiency and incident flux density. Then there is thermal noise, dark current noise and read noise before conversion from analogue to digital. In any system, the contribution of noise is worst at the start of the signal path where the signal is weak and so it makes sense to apply amplification early in the signal chain. If you wait until later (after A/D conversion) you are effectively amplifying the contribution of read noise and quantisation noise. Boosting the sampled pixel values at this stage brightens the image but does not increase the signal-to-noise ratio (SNR).

I use a 5D Mk III and it has good low-light performance because the pixels are large and the sensor electronics are reasonably quiet. The signal from each pixel is amplified locally so the ISO setting is only amplifying the effects of shot noise and dark current (with a minimal contribution from thermal noise in the amplifier itself) and therefore the signal plus some noise is greater when the effects of read noise and quantisation noise are taken into account. For this reason, I think it's best to adjust the ISO (as far as this controls local amplifier gain and not just multiply sampled values in software) rather than rely on a low noise floor and boosting the signal plus noise in post.

Theory aside, I'd be interested to hear about anyone who has been pushing under-exposed images from so called ISO invariant cameras. Do you think the quality is good enough or will we be setting ISO ourselves for a while longer?

Hi Dave

A good place to get (measured) information on the noise characteristics of sensors is DXOMark. You will find figures published for Dynamic Range and SNR. The DR figures are most useful for assessing how noise in the shadows varies with ISO as they effectively indicate what is happening with read noise. The higher the DR figure, the lower the read noise. The SNR figures vary with signal level and for all but the darkest areas of the image, shot noise tends to be dominant as it increases with the square root of the signal level.

Here is a comparison DR curve from DXOMark showing the Nikon D750 vs the Canon 5D MkIII.



Notice how the D750 varies pretty much linearly with ISO. This is what I call ISO invariance because it does't really matter with this whether you apply amplification using the in-camera ISO amplifier or externally in software. On the other hand, the 5D MkIII is not a linear curve with a fairly flat area at low ISO values. With this camera, at least up until about ISO 1600, you are better off applying amplification with the in-built ISO amplifier.

Re-iterating that the main noise of interest in this consideration (ie raising signal in the dark shadow areas) is read noise, the explanation for this difference in behaviour between the two sensors lies in where the main contribution of read noise occurs. With the Canon, there is a significant contribution of read noise after the ISO amplifier (in the input circuitry to the A/D converter). This noise is not amplified by the ISO amplifier but would be amplified if amplification were applied in software. With the Nikon, the contribution to read noise from the ADC is negligble. Nikon/Sony have lead Canon in this area for some time but with recent camera releases (eg 5D MkIV), Canon are catching up. Most people believe this is due to Canon moving the ADC onto the main sensor chip (like Sony/Nikon).

If you want further reading on the theory, I'd recommend this article by Emil Martinec.

The usefulness of ISO Invariance is probably open for debate (although it's always good to have the lowest possible read noise). The beauty of using the ISO adjustment is that exposure is handled automatically in camera and in particular a suitable shutter speed can be obtained without having to think about it too much when shooting. eg you can set a min ss when in auto ISO mode.

Dave

[Adzman808]

I've played a bit with ISO invariance on my Fujis.

A lot of it of course comes down to semantics... No digital camera has anything other than a fixed ISO, but judging by the preceding posts, y'all know waaaay more than me about this, so you'll know that!

FWIW I didn't find much benefit in pushing files that I'd circumnavigated pre-ADC brightening with. (But that's not true invariance anyway)

But once the signal was pure digital gain, then there was mileage in selectively pushing what I wanted in post, rather than the global brightening of gain increase.

The main barrier I found was what PP software I was playing with. For example LR CC offers about 5 stops of push and retains colour gamut. SilkyPix offers 3 and doesn't

There's a lot of BS on ISO invariance out there... Some people think it means that they can shoot 1/500 ISO200 in darkness and sort it all out in post. It doesn't.

True ISO invariance should mean its about the same to let the camera add the gain as it is to do it yourself. They'll be shots where you'd rather a lower iso, protect the highlights and push the shadows in post, and that can work well with a ISOless sensor.

You'll IMO not really gaining DR by (say) pushing iso800 3 stops in post, you're just using the greater amount of DR available at 800 to push it to 6400 eqv.

Which will also increase noise.

In fact not that much different than letting the camera select 6400 in the first place!

But (like I said) if you take the 800 shot and lift the shadows 3 stops, but not the highlights, then you'll theoretically have a cleaner image than a global gain push.

Use with care and with an attempt at understanding was what I told myself. YMMV