I just don't get the whole "Expose to the Right" philosophy

[Ancient_Mariner]

The Nikon D40 was one of the first Nikons to have read noise non-proportional to ISO, dropping at higher ISOs. You or your converter are probably doing something non-equal.

Most camera either will have less read noise (relative to absolute signal) at the higher ISO, or the same. The only camera I've seen where higher ISOs are absolutely noisier is the Panasonic FZ50 I have, and only by a small amount.

Here's the 800 ISO image (100% crop)



And here's the 1600 ISO one. The only difference is that the 1600 ISO has a reduced brightness to appear similar to the 800 ISO one. I think it's clear that the 1600 ISO one is (a bit) noisier

 

[John Sheehy]

Here's the 800 ISO image (100% crop)

And here's the 1600 ISO one. The only difference is that the 1600 ISO has a reduced brightness to appear similar to the 800 ISO one. I think it's clear that the 1600 ISO one is (a bit) noisier

You changed perpective between shots. You could have altered the lighting with your body in a different position.

There is no control over sharpening and NR at various tonal levels in most converters; the converter uses it's own recipe and/or one suggested by the manufacturer. The NR and sharpening settings may have to be different for different ISOs, for equal treatment. If I sharpen the ISO 800 by just 25% with a radius of 0.4 pixels (with PS' USM) the 8000 is just as noisy as the 1600. Whose to say this amount of difference in sharpening didn't occur in conversion? With OOF subjects, it is kind of hard to get an anchor on how much sharpening is done.

I didn't promise that converter would cooperate with camera potential! I concern myself first and foremost with RAW data; your RAW data will stay the same as converters come and go.

--
John

 

[Ancient_Mariner]

You changed perpective between shots. You could have altered the lighting with your body in a different position.

A fair assumption - I knew you would say that - but we're talking about a few centimeters and the light source was on the ceiling

There is no control over sharpening and NR at various tonal levels in most converters; the converter uses it's own recipe and/or one suggested by the manufacturer. The NR and sharpening settings may have to be different for different ISOs, for equal treatment. If I sharpen the ISO 800 by just 25% with a radius of 0.4 pixels (with PS' USM) the 8000 is just as noisy as the 1600. Whose to say this amount of difference in sharpening didn't occur in conversion? With OOF subjects, it is kind of hard to get an anchor on how much sharpening is done.

I didn't promise that converter would cooperate with camera potential! I concern myself first and foremost with RAW data; your RAW data will stay the same as converters come and go.

No other change but brightness was made. No shaprening, no NR. ACRaw (at least the CS2 version that I have) is completely blind to camera settings anway and it applies the same settings to all files opened except WB. The two images are JPEGs saved from RAW with only difference the ISO value and the reduced brightness (ISO 800 had the ACRaw default of 50, the ISO 1600 had 22).

 

[GaborSch]

For me, it goes against all photographic instinct that higher ISO could ultimately result in less noise

You are misinterpreting your experience. Normally you select higher ISO AND lower exposure, and the latter causes higher noise. The higher ISO does reduce the noise, but it does not make up fully for the lower exposure.

The reasonable way to look at it is measureing the noise on patches with the same amount of light collected. Of course, if you increase the ISO by one stop, then the pixel intensity of the same spot with the same illumination will be one stop higher.

I actually experimented, I took two shots at same speed/aperture, one at ISO 800, appearing slightly underexposed, the other at ISO 1600 appearing slightly overexposed

The ISO 1600 sample was a bit noisier than the ISO 800, as I expected.

Not only was the expectation was wrong, but the judgement of noise as well. Your ISO 1600 image is less noisy than the ISO 800.

Pls load both in PS. Make a selection on a relatively uniform patch on both images (if you overlay them as layers, it's even easier). For example at the bottom left corner, on that nice whatever color, dark orangy. The Histogram panel shows Mean 74.31, Std Dev 64.70 on the ISO 800 image. This makes 87.06% noise. The same on the ISO 1600 shot is Mean 68.35, Std ev 72.74, which yields 81.41% noise - LESS than the ISO 800 shot.

Now make selections on different areas, the darker the better, and compare the noise. I wonder what result you get on the small, very dark patch, about 1/3 from the top and from the left.

--
Gabor

http://www.panopeeper.com/panorama/pano.htm

 

[John Sheehy]

No other change but brightness was made. No shaprening, no NR.

You can't seriously believe that there is no noise reduction in those images!

ACRaw (at least the CS2 version that I have) is completely blind to camera settings anway and it applies the same settings to all files opened except WB. The two images are JPEGs saved from RAW with only difference the ISO value and the reduced brightness (ISO 800 had the ACRaw default of 50, the ISO 1600 had 22).

That's what ACR tells you, but that is not what ACR is doing. The sliders mean different things for different cameras, even the same camera at different ISOs and tonal levels. ACR is not a valid program for converting with no noise reduction and no sharpening.

--
John

 

[GaborSch]

Why wasn't this test conducted OVER "correct" exposure
(which is what expose to the right is)
and then pulled back to correct exposure level ?

In other words, instead of cranking up underexposed shots
which would be short on information ,
why not take overexposed ones which will have all/too much/blown
information and compare ?

AM's starting position was:

If the scene is dynamic (i.e. moving subjects), the first option is not applicable. The third option does not make sense, since exposing to the right is used to improve the signal/noise ratio

In other words, the issue is how to handle situations, where the ideal solution (from the noise perspective), namely increasing the exposure can not be achieved.

I have demonstrated that increasing the ISO can be used to improve the signal/noise ratio.

--
Gabor

http://www.panopeeper.com/panorama/pano.htm

 

[richardplondon]

See if you can get him to admit (or even believe) that...

ah, but I have no interest in circus tricks ;-)

RP

 

[Ehrik]

Aren't the DR curves at DxOmark useful for this purpose?

Yes, as long as the highlight headroom isn't changing for some ISO(s).

Yes, good point.

*) I've looked at blackframes from the G1/GH1/GF1 in IRIS and can confirm that the read noise doesn't go up proportional with ISO. Estimating the exact magnitude is hard due to black clipping.

You can do a histogram analysis of the blackframe to get a rough estimate of where black actually is.

Yes, I've tried various ideas and come up with rough estimates.

Let z = number of zeros. Let y = number of non-zeros. Let p(potential real zeros) = z-y. Does p seem like the peak in a natural, gaussian progression, going

+3 -> +2 -> +1 -> p? If so, black is probably clipped in the right place. If not, slide the histogram and try again.

Two problems: At ISO 100, for the tested GH1, z = 0.994n and y = 0.006n (!) (n=number of values, "zero" is 16 BTW as you probably know from the FZ50). So there is very little left to make estimations on. And then the distribution deviates a little from a pure Gaussian. It has a little more peak and tail and little less around the inflection point.

Actually, there's a third problem too, there's a slight gradient over the frame, so one has the trade-off of either having too few samples or a too high reading of the read noise.

At ISO 1600, z = 0.7n so it's easier there.

The G1 and GF1 clip closer to the middle.

I have found that if the blackframe is clipped exactly where it should be, and there is enough bit depth so that there is little quantization (a nice half-bell forms), then the real standard deviation is about 1.63x the measured one.

Is there a way to get a std.dev. reading straight from IRIS?

The GH1 has a lot less read noise than the other two. It uses another 12Mp sensor.

What was the std dev as measured, and what is the bit-depth?

12 bits, with 16 the lowest value and 4095 the highest. Noise and blackpoint in ADUs:

ISO 100:
G1 & GF1: 3.9+-0.05 (Black = 16)
GH1: . .. . 3.5+-0.5 (very unsure estimate) (Black = 8+-2)

With p = 0.994n for the GH1 at ISO 100, I wonder if the practical DR gain (assuming it's 3.5 -v- 3.9) over the G1 is lost to clipping? Your opinion?

ISO 1600:
G1 & GF1: 35.5+-5.5 (Black = 19.9+-0.5)
GH1: . .. . 12.15+-0.35 (Black = 10.6+-0.35)

The tested GH1 had quite a bit of line noise, though.

Note I did this on all channels together because splitting the channels as described in the bundled documentation gave "unknown command" error. When I later found how to split in the Astrosurf site tutorial, I redid one test on a single channel to check but the result was essentially the same.

I've run an FFT PS plugin on the red channel of a dcraw -D TIFF but found no evidence of noise reduction on files with the setting NR+2. But with the setting NR-2 (sic) some ISOs have a vertical local smearing which reduces the std.dev. So one has to make sure one is testing files taken with NR+2 (or NR+1).

Björn Utpott ( viztyger ) has kindly made five GH1 blackframes available here:
http://www.esnips.com/web/GH1-Black-Frames
And poster oluv , G1 ISO 100 here:
http://dl.getdropbox.com/u/893528/iso100.RW2
and G1 ISO 1600 here:
http://dl.getdropbox.com/u/893528/iso1600.RW2

 

[Derge]

Is there a way to get a std.dev. reading straight from IRIS?

Typing "> stat" at the command prompt will provide the standard deviation, along with some other information.



--
'So what do you take pictures of?' 'Mostly nouns.'

 

[samusan]

I have found that if the blackframe is clipped exactly where it should be, and there is enough bit depth so that there is little quantization (a nice half-bell forms), then the real standard deviation is about 1.63x the measured one.

With one R command you find it's a bit higher than that, more like 1.658 + -0.001:
x=replicate(100, 1/sd(abs(rnorm(1e5))))
c(mean(x), 1.96*sd(x) sqrt(length(x)))
yields
1.658, 9.2e-4

But a better way to work around the clipping is to use quantiles.
For instance, using the distance between 60% quantile (q1) and 90% (q2) :
d=qnorm(0.9) - qnorm(0.6)
1/d
yields 0.9752, so you can estimate the sd of a normal sample by 0.9752*(q2-q1).

This sd estimator's error is just twice the error of the usual sd estimator but is resistant to clipping and bias.

--
Samusan

 

[samusan]

Two problems: At ISO 100, for the tested GH1, z = 0.994n and y = 0.006n (!) (n=number of values, "zero" is 16 BTW as you probably know from the FZ50). So there is very little left to make estimations on.

You can use
4.243815 * (q2 - q1)

as an estimator, using the distance between the 0.994 quantile (q1) and 0.997 (q2). It estimates the sd of the (normal) unclipped distribution with a precision of a bit under 1% (for 12 Mpixels)

Maybe hard to compute with IRIS though, unless you can compile plugins to extend its features.

Or you can also estimate the unclipped sd with 29.6 times the sd of the sample.
It's only 3 times less precise (3.6% for 12 MPixels)

And then the distribution deviates a little from a pure Gaussian. It has a little more peak and tail and little less around the inflection point.

Oh, is there any hypothesis why it is so ?

Actually, there's a third problem too, there's a slight gradient over the frame, so one has the trade-off of either having too few samples or a too high reading of the read noise.

you can compute estimates for each line and then average them. Once again I'm guessing IRIS can't be made to do that ?

--
Samusan

 

[Ehrik]

Two problems: At ISO 100, for the tested GH1, z = 0.994n and y = 0.006n (!) (n=number of values, "zero" is 16 BTW as you probably know from the FZ50). So there is very little left to make estimations on.

You can use
4.243815 * (q2 - q1)

as an estimator, using the distance between the 0.994 quantile (q1) and 0.997 (q2). It estimates the sd of the (normal) unclipped distribution with a precision of a bit under 1% (for 12 Mpixels)

Thanks for your suggestion, you clearly understand this better than me.

Maybe hard to compute with IRIS though, unless you can compile plugins to extend its features.

I'm pretty new to IRIS but I didn't find such functionality in the bundled documentation at least.

Or you can also estimate the unclipped sd with 29.6 times the sd of the sample.

Yes, that could be had from IRIS (thanks, Derge).

It's only 3 times less precise (3.6% for 12 MPixels)

... if the distribution is Gaussian. But I'm thinking it could be way off with even a modest deviation in the distribution?

And then the distribution deviates a little from a pure Gaussian. It has a little more peak and tail and little less around the inflection point.

Oh, is there any hypothesis why it is so ?

Not from me. But it's been observed that some Canon model (50D?) had more outliers (at a similar std.dev.) than a Nikon model (D300?). Some think Nikon had filtered those out from the raw files, but ejmartin did some elaborate test that came out negative.

Actually, there's a third problem too, there's a slight gradient over the frame, so one has the trade-off of either having too few samples or a too high reading of the read noise.

you can compute estimates for each line and then average them. Once again I'm guessing IRIS can't be made to do that ?

Not that I'm aware of. Your suggestion would as a bonus remove the influence of locally mis-matched black-level offsets for individual lines ("banding") in one dimension.

Failing plug-in functionality in IRIS, it would be possible to alter dcraw to get the data needed, but I'm not willing to invest the time to do it.

 

[MarkNicholas]

My understanding is that the main reason / benefit of shooting to the right (RAW only !!) is that you get more detail. An improvement in noise (or not as the case may be) is only a side product of shooting to the right.

This increase in detail is of no use if you let your camera produce Jpegs. It is only of use if you convert your own RAW files using propritary software.

There was a very good thread on this forum a while ago but I cannot locate it at the moment.

 

[Razr]

My understanding is that the main reason / benefit of shooting to the right (RAW only !!) is that you get more detail. An improvement in noise (or not as the case may be) is only a side product of shooting to the right.

This increase in detail is of no use if you let your camera produce Jpegs. It is only of use if you convert your own RAW files using propritary software.

There was a very good thread on this forum a while ago but I cannot locate it at the moment.

What about those of us who do not ever use the histogram for any reason?
You have a lot of confidence in your skills .

And those of us who wait to see the "Real" image on the monitor?
You have a lot of confidence in your skills

What about those of us who shoot wiith the LCD turned off , shooting like the old-timers used to do?
Mad skills!

 

[Sammy Yousef]

Your interpretation is just plain wrong and unfortunately it's confusing others.

All you've demonstrated here is that under-exposing and post processing at low ISO gives more noise than correct exposures at a higher ISO without post processing.

That's no great surprise. There is a reason to strive for correct exposure in camera. It makes for the best picture. Even if you use RAW (which does give more leeway by recording more information but doesn't change your shutter speed or physical aperture).

Some people like to think of the "exposure triangle" with ISO, shutter speed and aperture at each corner. The trick is to get them all right in the first place. The minute you brighten one picture but not the other you're not playing on a level field.

If higher ISOs were indeed less noisy, cameras would be built to default to higher ISOs rather than low ISOs. (In fact where the lower ISOs are below the base ISO of the sensor that's exactly how they work because the sensors don't do so well with dynamic range below their base ISO. My D90 defaults to ISO 200 not Lo 1)

In practical terms what you've shown (for a limited set of circumstances), is that if you have a choice between shooting low ISO and post processing, or higher ISO to get correct exposure, correct exposure with higher ISO will give you better results. But if you can shoot at lower ISO AND get correct exposures (eg. if you can control the light as with flash), that's preferable to both previous options. Correct exposure is king.

--
Sammy

 

[Les Olson]

If you have 12-bit encoding there are 4096 levels of intensity across the bottom of the histogram. Digital sensors (unlike your retina) encode linearly: level 2048, halfway across the histogram, represents exactly half the number of photons in level 4096, level 1024 half as many again and so on. Now - key point - remember that each stop of exposure is a doubling of the light intensity. That means that if you have 4096 levels and 6 stops of dynamic range (= six doublings of light intensity), 2048 of those levels - the whole right hand half of the histogram - are devoted to the brightest stop, 1024 to the second brightest, and so on, and the darkest stop is represented by only 64 levels.

That is, if your image data is mainly on the left hand side of the histogram it is much more coarsely encoded than it would be if it were on the right hand side of the histogram, and that is a bad thing.

Adobe has a good pdf on this issue (go to http://www.adobe.com/ and search for linear gamma - the first resource that comes up is the one you want).
--
A man of swallows and acquainted with beef.

 

[richardplondon]

What about those of us who shoot wiith the LCD turned off , shooting like the old-timers used to do?

Mad skills!

Not really. There are specific techniques one can use even with the LCD turned off that do not rely on superhuman expertise (or informed guesswork, as it is more properly described).

For example: with the camera meter set to Spot, you can point the spot circle at a mid-toned part of the subject, or even at a specific greycard target. This would not be ETTR; the camera manufacturer has set a constant representation of midtone that will be achieved in the output when the meter is used in this way, and the remainder of the scene takes its chances regardless. The midtone is in the driving seat.

Or, with the camera set to Spot, one can instead start from a knowledge of how many stops above the camera manufacturer's midtone it is safe to go without clipping. Then meter off the brightest part of the scene where detail is to be retained, while applying compensating negative exposure correction by that known number of stops. This is, precisely, ETTR: the midtone level in the scene will be variably represented depending on the specifics of the scene and on the capabilities of the camera. The highlights are now in the driving seat.

Real world use of Matrix (Pattern) metering often falls closer to the latter in reality, though a compromise of both. The metering cells that encounter the brightest parts of the scene cast a vote for an ETTR-related exposure, and the ones that encounter less bright areas cast a vote for a midtone-related exposure. The camera takes a look at the pattern of these responses and weights them in a pre-programmed way, and comes up with a decision - partly, this is affected by a guess of whether highlight detail is more or less important in a given case.

So there is a form of automatic ETTR-style judgement (highlight protection) embedded even in green easy mode with the histogram turned off - no "mad skills" required.

Anyone who uses Matrix / Pattern metering cannot deny the effectiveness of ETTR as a concept - at least, should not, IMO. When they turn to more manual forms of metering, it is all up to them of course and they may not choose to conceive the task in that way.

RP

 

[John Sheehy]

That is, if your image data is mainly on the left hand side of the histogram it is much more coarsely encoded than it would be if it were on the right hand side of the histogram, and that is a bad thing.

That's a totally theoretical approach to the issue, which ignores the basic fact that there is so much noise in digital captures that bit depth or "levels" is not an issue - neighboring pixels which should be the same intensity vary by many RAW levels at all tonal levels, even with only 12 bits.

Signal-to-noise ratio is what drives ETTR, in practice. Higher absolute exposure increases SNR. Higher relative exposure with fixed absolute exposure through a higher ISO may reduce late-stage system noises, relative to absolute exposure.

--
John

 

[John Sheehy]

Your interpretation is just plain wrong and unfortunately it's confusing others.

The truth is often confusing to people incapable of grasping it, but that is no excuse for propagating myths. The common ways of looking at things are often quite wrong. High ISOs, except in possibly a very small number of poorly designed cameras, do not add more noise to an image. They add less, or the same amount of noise, than lower ISOs. High-ISO images are generally noisier only because they dictate weaker exposures, to have the higher exposure index. Anyone who believes that higher ISOs "add" more noise to their images is incapable of coming consistently to accurate conclusions.

All you've demonstrated here is that under-exposing and post processing at low ISO gives more noise than correct exposures at a higher ISO without post processing.

Nonsense. There is no processing difference. The images are simply scaled to equal intensity; if you consider that "processing", you are very easily impressed.

That's no great surprise. There is a reason to strive for correct exposure in camera. It makes for the best picture. Even if you use RAW (which does give more leeway by recording more information but doesn't change your shutter speed or physical aperture).

Some people like to think of the "exposure triangle" with ISO, shutter speed and aperture at each corner. The trick is to get them all right in the first place. The minute you brighten one picture but not the other you're not playing on a level field.

Nonsense. Brightening an image is a meaningless, transparent operation, just like raising the volume on a stereo when switching from FM tuner to CD.

If higher ISOs were indeed less noisy, cameras would be built to default to higher ISOs rather than low ISOs.

This statement makes me wonder if you are agreeing to disagree, or if you even understand at all what you are disagreeing with. No one has suggested that higher ISOs, by default, give less noisy images. It is only when ISO is not dictating exposure, but rather, a variable trading SNR against headroom, that what Gabor is demonstrating is claimed to be true.

(In fact where the lower ISOs are below the base ISO of the sensor that's exactly how they work because the sensors don't do so well with dynamic range below their base ISO. My D90 defaults to ISO 200 not Lo 1)

That's just because the wells run out room for extra charge. I'm not one for substituting truth with simplicity, but sometimes simplicity is the truth, and this is it. Base ISO of a camera is simply determined by how much highlight headroom is decided to be enough, and the capacity of the wells. No electronic magic or gain issues ("do so well") involved here at all; the wells are filled, it's that simple (of course, the digitization process often clips slightly below sensor saturation). Most current DSLRs have maximum well capacities which allow a base ISO of 120 to 160 with a standard of 3.5 stops of highlights. Below that, there isn't enough capacity for 3.5 stops. I've seen a range of 2.5 to 4 stops of headroom in DSLRs by design.

In practical terms what you've shown (for a limited set of circumstances), is that if you have a choice between shooting low ISO and post processing,

What post-processing? You are weaving things into this which have no place or relevance. Processing was the same (basically, just demosaicing), except for the trivial, transparent act of arithmetic gain.

or higher ISO to get correct exposure, correct exposure with higher ISO will give you better results. But if you can shoot at lower ISO AND get correct exposures (eg. if you can control the light as with flash), that's preferable to both previous options. Correct exposure is king.

"Correct exposure" is an embarrassing term. It offers no concise meaning, and is open to interpretation. You have to shoot a scene, and there are lights in the shot which are much brighter than the ambient background and subject. no matter what exposure you choose, you will blow out a certain amount of detail in the bright lights. You want to have as much detail from the lights as possible. What is "correct exposure"?

You shoot a subject that is darker than the gray background. What is "proper exposure"? I know what optimal RAW exposure is; it is one where the gray background is just below clipping of the RAW (which would be an "over-exposure for a default JPEG). This could be at a low ISO for immobile subjects from a tripod; handheld or with subject motion, this may be at a high ISO with +2.6 or +3 EC. ISO 800 with +3 EC will give less read noise and the same shot noise as ISO 100 with 0 EC.

--
John

 

[John Sheehy]

My understanding is that the main reason / benefit of shooting to the right (RAW only !!) is that you get more detail. An improvement in noise (or not as the case may be) is only a side product of shooting to the right.

This increase in detail is of no use if you let your camera produce Jpegs. It is only of use if you convert your own RAW files using propritary software.

There was a very good thread on this forum a while ago but I cannot locate it at the moment.

Given the same optics, and shutter speed, and conversion, noise is the only thing obscuring detail.

--
John