• Please use real names.

    Greetings to all who have registered to OPF and those guests taking a look around. Please use real names. Registrations with fictitious names will not be processed. REAL NAMES ONLY will be processed

    Firstname Lastname

    Register

    We are a courteous and supportive community. No need to hide behind an alia. If you have a genuine need for privacy/secrecy then let me know!
  • Welcome to the new site. Here's a thread about the update where you can post your feedback, ask questions or spot those nasty bugs!

Measuring camera's saturation point in ADU; two different readings

Alex Weiss

New member
Hi,

first of all, this is a great, great forum, and I've already learnt a lot from reading here - heck, Doug Kerr's posts alone led me to register.

I bought a Canon 550D a while ago and would now like to measure its various characteristics, such as read noise and dynamic range (total and above/below middle gray). I've already measured read noise at various ISOs, and the results are consistent enough for me to conclude that my measurements were correct.

To calculate a theoretical dynamic range (theoretical because it is the engineering definition of the term), I simply want to find out the ratio between the highest possible signal level (i.e. at saturation) in ADU and the noise level.
But here's the problem: I get two different sets of readings for the highest possible ADU; in pictures with "normal" overexpose where only parts are blown out I get a number in the 13000s (don't remember the exact number, but it is always the SAME number, +/- 1). However, in pictures with severe overexposure (read: a white picture with pretty much no details at all), I always get a number in the 15000 (again, always the same one). I cannot explain this discrepancy, and it does not make sense at all to me. It also seems that these numbers, especially the first, are rather low, as these are the raw numbers that still include the offset Canon applies to all RAW data (2048 in the case of the 550D).
Has something similar ever happened to anybody?

Thanks,
Alex
 
Hi,

first of all, this is a great, great forum, and I've already learnt a lot from reading here - heck, Doug Kerr's posts alone led me to register.

I bought a Canon 550D a while ago and would now like to measure its various characteristics, such as read noise and dynamic range (total and above/below middle gray). I've already measured read noise at various ISOs, and the results are consistent enough for me to conclude that my measurements were correct.

Hi Alex,

Which application did you use to read the Raw Bayer CFA data before demosaicing?

To calculate a theoretical dynamic range (theoretical because it is the engineering definition of the term), I simply want to find out the ratio between the highest possible signal level (i.e. at saturation) in ADU and the noise level.
But here's the problem: I get two different sets of readings for the highest possible ADU; in pictures with "normal" overexpose where only parts are blown out I get a number in the 13000s (don't remember the exact number, but it is always the SAME number, +/- 1). However, in pictures with severe overexposure (read: a white picture with pretty much no details at all), I always get a number in the 15000 (again, always the same one). I cannot explain this discrepancy, and it does not make sense at all to me. It also seems that these numbers, especially the first, are rather low, as these are the raw numbers that still include the offset Canon applies to all RAW data (2048 in the case of the 550D).

Hard to say where the difference comes from, but make sure it's not in the data collection process. Barring some specific Custom Function setting, the only thing I can imagine is that on the partially blown image there is still the tail of a noise distribution around saturation, and perhaps a not perfectly uniform brightness distribution (partly clipped, partly not clipped).

Also, from a statistical point of view, you get more reliable results by judging the data from the difference of pairs of exposures and black frames. That will e.g. reduce the effect of amplification differences between individual sensels to the individual sensel noise contribution.

I've described my testing methodology in an earlier thread about a different camera.

The posts by Emil Martinec are also worth reading for those interested in the technical details of dynamic range, noise, and ISO gain effects.

Cheers,
Bart
 

Alex Weiss

New member
Hi Alex,

Which application did you use to read the Raw Bayer CFA data before demosaicing?

Hi Bart,

I used iris. Just to make sure, I also took a couple of readings with Rawnalyze, and they matched.

Also, from a statistical point of view, you get more reliable results by judging the data from the difference of pairs of exposures and black frames. That will e.g. reduce the effect of amplification differences between individual sensels to the individual sensel noise contribution.

I did that with the black frames, but not with the overexposed pictures - wouldn't a subtraction just leave me with noise?

I've described my testing methodology in an earlier thread about a different camera.

The posts by Emil Martinec are also worth reading for those interested in the technical details of dynamic range, noise, and ISO gain effects.

Thanks for the links, I've read them. In your analysis, how did you determine the saturation point? I don't think you mention it in your post (unless I didn't see it).

Thanks,
Alex
 

Doug Kerr

Well-known member
Hi, Alex,

I haven't actually done a lot of this type of investigation (and none recently, so that I could remember the results). I applaud your looking into this.

My first thought is that the entire actual process of administering the sensor system is no doubt extremely complex. I can well imagine that things go on we do not think of in our basic simplistic model of sensel detector operation and the attendant analog-to-digital conversion.

One question I have is do you by any chance have highlight tone priority turned on (I assume that camera has it). I don't know what that actually does (there is a lot of folklore about it) and where in the chain this occurs, but it just seems as though if that were in effect it could intrude into the process.

Clearly, this result needs further probing.

Perhaps it would be useful to look into the matter of "normal" and "extreme" overexposure. Perhaps you could see how the maximum ADU value varies as the actual peak photometric exposure varies. Then perhaps you could see how the distribution of luminance across the frame affects this relationship - that is, a frame of uniform luminance vs. a frame that was mostly of a low luminance but had a modest area of that luminance.

Another issue would be whether the camera exposure (actual Ev) was manually chosen or resulted from the operation of the metering system. (I could imagine some type of "adaptability" in the sensor administration based on cues from the metering system.)

Again, these are just random thoughts (many of which might already be on the table).

Again, my thanks for undertaking this investigation. I may decide to parallel it with my EOS 20D.

Best regards,

Doug
 

Doug Kerr

Well-known member
Hi, Alex,

Now that I have done breakfast, let me drop some thoughts on the table. Exactly what, if anything, these have to do with the mystery at hand I don't know. You are probably well aware of these things, but perhaps they deserve illumination.

Of course saturation can have many meanings, depending on the variable of interest, where it is observed, and whether we are really interested in that variable or in another variable of which it is a function.

At the beginning of the chain, we can think of saturation as the maximum photometric exposure (H) on a sensel for which incremental changes in H can be perceived downstream, in terms of the sensel physics proper.

In this regard, it is important to note that the "charge well" metaphor often used in this connection an be rather misleading; in sensors of the "COMOS" type, the charge in the sensel does not increase for each photon that is "recognized". Rather, an initial charge is depleted (ideally by 1 e) each time a photon is "recognized".

Thus, as exposure proceeds, and H increases, the charge (Q) declines, as does the voltage that reveals the charge (V=C/Q).

We visualize detecting this with a differential amplifier, measuring the difference between the "end-of-exposure" voltage on the sensel and a reference voltage which, simplistically, is regulated to be equal to the initial voltage to which all sensels are (hopefully) charged. The negative of that difference is the "analog output" of the sensel.

You have perhaps read my article "The CMOS APS Digital Camera Sensor", available on The Pumpkin, here:

http://dougkerr.net/Pumpkin/#CMOS-APS

It discusses this concept at length (but realize that it is not based on any definitive information from Canon).​

The next issue is the digital clipping of the ADC output. Clearly, we will not have precisely the same differential voltage for each sensel at the end of each "overexposed" cycle. There may be residual voltage differences, and of course there is the matter of noise at the input to the differential amplifier. But we expect to get a consistent ADU value each time (and in fact I think you experience such over all sensels for any given "overexposure" situation).

This precisely consistent maximum (for any given overexposure situation) must be a result of digital (logical) clipping of the actual basic ADC output, not a manifestation of the limiting end-of-exposure voltage on the individual sensels or the observed difference between that and the reference voltage.

Now, why might the camera use a different clipping level in different "situations", and what property of the "situation" governs that? This is of course a major thrust of your quest. But we must assume that it is done in pursuit of some behavior Canon thinks is advantageous to the user.

I am reminded of the engine temperature gauge system in my 1998 VW Passat, in which, for actual detected temperatures in the range 185°F - 195°F, always indicated 190°F. The objective was believed to be to prevent the user from obsessing over continuous small changes in engine temperature as the car operated. (Paranoia is its own reward!)​

Well, just some observations that might lubricate the thought process.

Best regards,

Doug
 

Ken Tanaka

pro member
Much of the (barely) practical data you seek is already available at DxOMark.

BTW, I also recently bought a T2i (550D) as a small DSLR. Frankly, engineering mumbo-jumbo aside, I've been floored by how good this camera really is. (And this is coming from a fellow who has a range of tools in his arsenal that reaches to the pinnacle, a Phase One P65+.) This camera just feels good to use and is just plain fun! I hope you can just enjoy yours to explore photography and get past the mostly irrelevant techie stuff. (Unless, of course, that's how you enjoy photography!)
 

Alex Weiss

New member
Hi, Alex,

I haven't actually done a lot of this type of investigation (and none recently, so that I could remember the results). I applaud your looking into this.

One question I have is do you by any chance have highlight tone priority turned on (I assume that camera has it). I don't know what that actually does (there is a lot of folklore about it) and where in the chain this occurs, but it just seems as though if that were in effect it could intrude into the process.

Hi Doug,

thanks for your detailed replies. To answer your question: I made sure that HTL wasn't turned on, so unfortunately, that cannot be the reason.

Perhaps it would be useful to look into the matter of "normal" and "extreme" overexposure. Perhaps you could see how the maximum ADU value varies as the actual peak photometric exposure varies. Then perhaps you could see how the distribution of luminance across the frame affects this relationship - that is, a frame of uniform luminance vs. a frame that was mostly of a low luminance but had a modest area of that luminance.

That is a very good idea. I'll do that once I find some more time. I like that idea a lot!

Another issue would be whether the camera exposure (actual Ev) was manually chosen or resulted from the operation of the metering system. (I could imagine some type of "adaptability" in the sensor administration based on cues from the metering system.)

Yes, that's what I thought at first, too. But all the exposures were chosen manually, so theoretically the metering system should have no influence on that. But then again, who knows...

Again, my thanks for undertaking this investigation. I may decide to parallel it with my EOS 20D.

You're most welcome, I enjoy doing these things. And yes, do measure your EOS 20D, I'd be interested in your results.


In this regard, it is important to note that the "charge well" metaphor often used in this connection an be rather misleading; in sensors of the "COMOS" type, the charge in the sensel does not increase for each photon that is "recognized". Rather, an initial charge is depleted (ideally by 1 e) each time a photon is "recognized".

That is very interesting. I, too, thought that charge increases rather than decreases. Regarding your last sentence: Why is it ideal if the relationship between e- and photon is 1:1? And isn't this relationship usually defined by the gain of the camera (which seems to be a misnomer)?

You have perhaps read my article "The CMOS APS Digital Camera Sensor", available on The Pumpkin, here:

http://dougkerr.net/Pumpkin/#CMOS-APS

It discusses this concept at length (but realize that it is not based on any definitive information from Canon).
I haven't read it yet, but I definitely will. There's a wealth of information on your web site, thanks so much for making it available to everyone.

I am reminded of the engine temperature gauge system in my 1998 VW Passat, in which, for actual detected temperatures in the range 185°F - 195°F, always indicated 190°F. The objective was believed to be to prevent the user from obsessing over continuous small changes in engine temperature as the car operated. (Paranoia is its own reward!)
Haha, that's great!!

One observation that I find interesting is that the difference between the two ADUs is conspicuously close to 2048, the offset Canon applies to the RAW data. But that might also be completely coincidental.

Much of the (barely) practical data you seek is already available at DxOMark.

Ken Tanaka said:
BTW, I also recently bought a T2i (550D) as a small DSLR. Frankly, engineering mumbo-jumbo aside, I've been floored by how good this camera really is. (And this is coming from a fellow who has a range of tools in his arsenal that reaches to the pinnacle, a Phase One P65+.) This camera just feels good to use and is just plain fun! I hope you can just enjoy yours to explore photography and get past the mostly irrelevant techie stuff. (Unless, of course, that's how you enjoy photography!)

Hi Ken,

don't worry, I do enjoy my 550D and use it every single day. And yes, the camera is great, and I'm not trying to find faults.
I understand your point, but I disagree with it. It is not irrelevant, and it does matter; first of all, I like to know what exactly my gear is capable of and what I bought, and I like to push my gear to its limits. Second, some of the measurements do have practical value: for example, knowing the dynamic range above middle gray allows me to get spot-on ETTR exposures without having to look at the histogram (which is inaccurate anyway), because I can simply take a spot meter reading at the brightest spot and then overexpose by that dynamic range above middle gray. And to be honest: it works even better than I thought.
Another example: the noise measurements debunk the myth that higher ISO = more noise, or rather, that it is simply the increasing ISO that's responsible for the added noise. Stuff like this does matter, as it allows you to gain a lot more control over your final product. Granted, I could have simply looked up some of the relevant data, but I believe that knowing why certain things happen matters.
Also, thanks for pointing out the DxOMark website. I've seen it before, but the problem is that various websites report vastly different data, which is yet another reason to conduct one's own measurements.
 

Asher Kelman

OPF Owner/Editor-in-Chief
...........knowing the dynamic range above middle gray allows me to get spot-on ETTR exposures without having to look at the histogram (which is inaccurate anyway), because I can simply take a spot meter reading at the brightest spot and then overexpose by that dynamic range above middle gray. And to be honest: it works even better than I thought.


Alex,

I'm particularly interested in being able to use a spot meter to measure the brightest parts of a face of a classical musician in performance, when the stage is lit from above, and be able to just know how to apply your findings/ideas to getting exposures that take the most advantage of the DR. So could you give examples of how this works in practice.

Based on Bart's advice, I'm overexposing at lower ISO's to get less noise and my guide is the histogram. If I had, in addition, an accurate table of DR for my camera then perhaps my work would be easier. I'm using the 5DII which may have greater DR than the T2i.

Asher
 

Ken Tanaka

pro member
Alex,

To each his own when it comes to photographic enjoyment.

Can we see some of your work where such intimate knowledge of the camera's electronics made a decisive difference in your success?
 

Asher Kelman

OPF Owner/Editor-in-Chief
Alex,

To each his own when it comes to photographic enjoyment.

Can we see some of your work where such intimate knowledge of the camera's electronics made a decisive difference in your success?

Ken,

In difficult situations, such as low light with harsh shadows, I'd be happy if Alex's approach is merely more convenient. So at least I'd enjoy seeing where it is of practical use, even though one could get the picture anyway.
 

Doug Kerr

Well-known member
Hi, Alex,

Why is it ideal if the relationship between e- and photon is 1:1?
This is a result of the fundamental physics involved - I can't explain it well! In any case, ideally, each photon causes the formation of an electron-hole pair.

And isn't this relationship usually defined by the gain of the camera (which seems to be a misnomer)?

I think the term usually refers to the scaling of the ADC - that is, what is the relationship between increments of the DN and increments of the charge in the sensel (in units of e). (Voltage on the sensel is an intermediate quantity.)

This talks about the notion:

http://spiff.rit.edu/classes/phys445/lectures/gain/gain.html

And it is of course a misnomer. We are really talking of a conversion constant from charge to a dimensionless number (gain is of course is between two dimensionally-comparable quantities, such as voltage or power).

Of course if we think not of charge but of number of electrons, then the quantities are dimensionally comparable. But we don't usually use "gain" to speak of a ratio between dimensionless numbers as a "gain" (if a block of a signal processing system multiples the binary number by 4, we don't say it has a "gain of 4".)

And of course in this case, the ratio is upside-down (input/output).

In any case, that all happens after the photon-electron "conversion" (in the sensel itself), so it doesn't affect the photon-electron relationship (the so-called quantum efficiency).

One observation that I find interesting is that the difference between the two ADUs is conspicuously close to 2048, the offset Canon applies to the RAW data. But that might also be completely coincidental.
Yes, and it is of course also 2^11, which might have something to do with it.

Best regards,

Doug
 

Alain Briot

pro member
Ken,

"Can we see some of your work where such intimate knowledge of the camera's electronics made a decisive difference in your success?"

I doubt we will see any. Focus on technique activates logical functions in the brain (left side I think if one subscribes to the left/right brain activity theory) while focus on artistic aspects activates the creative function of the brain (right side according to the same theory). I'm not so in love with the right/left side theory, but I certainly have evidence that what we think about directly influences what we do and the work we create!

Of course, if the goal is not to create art but to document the subject as accurately as possible, then a focus on technique would work. It's just not my cup of tea (not that I drink much tea, but just to use a commonplace metaphor).
 

Asher Kelman

OPF Owner/Editor-in-Chief
Of course, if the goal is not to create art but to document the subject as accurately as possible, then a focus on technique would work. It's just not my cup of tea (not that I drink much tea, but just to use a commonplace metaphor).

Alain,

Let me share with you my interest to create art where I have no control of light. I shoot wonderful music performers where lighting is rigidly outdated. It comes from above to create ghoulish eye sockets and over-exposed foreheads and fingers of the soloist. If the dynamic range of the camera was 2-3 stops better at ISO 1600-3200, I'd be fine. So I shoot at 1600 and over-expose by a stop as suggested by Bart and my mages have less noise. Still, relying on the LCD histogram is not always perfect for me. Alex's idea of spot-meterng the important brightest areas holds the possibility of more conveniently getting the exposure right. His approach would be to choose to expose about 3 stops above mid gray recommended setting for that ISO, of course depending on the dynamic range of the particular camera.

Whether or not Alex's approach has any merit for others, for me, working at the edges of comfort, any extra guidance is worth considering. Perhaps using a Nikon D3x instead of the Canon 5DII would obviate necessity for nifty tricks like this! That's an elegant but expensive solution. I'm hoping that a 5DIII would bring us that extra dynamic range we need for this low light high contrast application.

In most of our work, we control the lighting by choosing when, where and how to take pictures or we bring the light with us.

So Alain, it is about artistry, just artistry in difficult light!

Asher
 

Alex Weiss

New member
Can we see some of your work where such intimate knowledge of the camera's electronics made a decisive difference in your success?
Ken,

sure, of course you can. Because you asked so nicely I'll even do you one better: just for you, I grabbed my camera today and took a couple of pictures to illustrate my points. I'll walk you through.

ISO100.jpg

ISO400.jpg"

These first two pictures were shot with the same photometric exposure (f/2.2 @ 1/80), but with different ISO settings; the first one was taken with ISO100, the second one with ISO400 and then pulled back -2 in ADR to match the photographic exposure of the first shot (as you can see when you compare both images). Both shots were processed with the standard settings of ACR (except for the aforementioned exposure adjustment in the second shot). They were neither sharpened nor was noise reduction applied, so please excuse the somewhat mediocre quality - the point here is not to show off my conversion skills but to compare noise levels.

The immediate question here is why one would even consider shooting the ISO400 shot, as the ISO100 shot is well exposed, and higher ISOs are said to increase noise. I present you with a pair of 100% crops as an answer:

iso100_close.jpg

iso400_close.jpg


I have two other pairs of crops, but the forum doesn't allow me to link to more than 4 images per post, so I'll post the other ones in the following post.

The results, I believe, speak for themselves: the ISO400 shot has less noise than the one shot at ISO100. How is that related to my measurements? I found out that absolute noise levels DO increase with increasing ISO, but scaled to match relative brightness, the actually go down (this, by the way, is something people before me have discovered, I just confirmed it for my camera). This debunks the myth that increasing ISO increases noise per se; what increases noise is that you usually let in less light when you increase ISO (i.e. less photometric exposure).
In the next picture and its accompanying crops, the effect is even clearer: again, both shots were shot with the same shutter speed and aperture, but one was shot with ISO400, the other with ISO800.

Now, of course, the question remains whether these differences in noise levels are important in a snap shot like this; they probably aren't. Are they important in high-quality work? You bet. Are they important in film and video, where we don't always have the luxury of having RAW files at our disposal? Absolutely. Are these differences important in low-light situations? Of course.

The next picture shows the effect even clearer; again, both were shot with the same photometric exposure, but one with ISO400, the other with ISO800. Please compare for yourself.

Even though you asked me how these measurements are related to image quality, I'll give you another, unrelated, example of why such measurements are beneficial. You are undoubtedly aware of the race for increasing bit depths in sensors (the move from 12 to 14 bits was one such example). Often, marketing departments even use these increased bit depths to advertise their products. But do we actually need them, and do they make a difference? At least for the 550D (and I've read similar things about different cameras), I believe the sensor might as well be 12 bit, as the extra bits are essentially "eaten up" by noise - the absolute noise floor even at ISO 100 is well above 1 or 2 ADUs. Does this matter for image quality? No, it doesn't, but it allows me to make informed decisions instead of simply buying into the bit depth hype.
 
I'd benefit from an example using the known DR and a reading of the highlights to see how this is used in practice. It would help me understand how it might be applied to my own work in low light with the 5DII. I can take spotmeter readings of the brightest areas, the foreheads of the performers under stage lights. If this has some convenience/accuracy value compared to looking at the histogram, I'm interested!

Asher,

It's not so hard to figure out, but the result differs a bit between camera models.

One takes a series of Raw images with increasing exposure of a uniformly lit surface. It's best to use an aperture of e.g. f/8 to reduce vignetting as much as practical, and slighly defocus the lens to avoid picking up any surface detail. So everything is on manual, the lens and the camera shooting mode. Now increase the exposure time in 1/3rd stops starting at, say, a metered exposure +2EV up to say metered exposure +4EV. If you are testing spotmetering, then use spot metering for the baseline. Normally i'd expect the exposures of 3 and 1/3rd stops above the (spot) metered settings to fully clip (Raw ADU values do not increase with increasing exposure).

Now use a tool like Rawnalyze (or IRIS which is a bit more challenging to use) to inspect the maximum Raw values of the different exposures, and there is your answer; how many EVs above highlight metering does the camera offer.

Two possible complications:
First with high ISO settings the noise will be higher so some noise will clip, while the other half of the highlight noise distribution is not yet clipped. So use your judgement wether you may want to lower your maximum EV above metering by a third.
Second, the highlight in the scene may be smaller than the metering spot. Again use you judgement and reduce the suggested exposure with an additional amount.

In addition, my advice to relatively 'underexpose' high ISO shots by one or two EVs, still stands. It will reduce the output noise level and give extra highlight detail.

Cheers,
Bart

P.S. it appears that the Rawnalyze website has vanished into cyberspace, a pitty. Well one could also use DCRaw with non-demosaiced Raw linear gamma output, but I'd have to look up again how that works.
 

Doug Kerr

Well-known member
Hi, Alex,

Very nice piece.

These first two pictures were shot with the same photometric exposure (f/2.2 @ 1/80)
For the benefit of the readers, let me point out that what you directly mention as being the same is the photographic exposure (f/2.2 at 1/80 s). However, since this is the same scene (the same luminance at each spot) in both shots, the photometric exposure for each spot is the same as well (which is of course your real point). (But I don't want people to think that shutter speed and aperture define photometric exposure.)

This debunks the myth that increasing ISO increases noise per se; what increases noise is that you usually let in less light when you increase ISO (i.e. less photometric exposure).
Indeed. As always in information theory and related fields, energy is the key! (One can even deduce relationships related to depth of field from an energy outlook!)

Thanks again for this nice presentation.

Best regards,

Doug
 

Alex Weiss

New member
For the benefit of the readers, let me point out that what you directly mention as being the same is the photographic exposure (f/2.2 at 1/80 s). However, since this is the same scene (the same luminance at each spot) in both shots, the photometric exposure for each spot is the same as well (which is of course your real point). (But I don't want people to think that shutter speed and aperture define photometric exposure.)

Ah, yes, that is indeed exactly what I meant. I always confuse the two terms. Thanks for pointing that out.

Thanks again for this nice presentation.
You're most welcome. I hope it will convince some readers.
 

Alex Weiss

New member
Asher,

It's not so hard to figure out, but the result differs a bit between camera models.

One takes a series of Raw images with increasing exposure of a uniformly lit surface. It's best to use an aperture of e.g. f/8 to reduce vignetting as much as practical, and slighly defocus the lens to avoid picking up any surface detail. So everything is on manual, the lens and the camera shooting mode. Now increase the exposure time in 1/3rd stops starting at, say, a metered exposure +2EV up to say metered exposure +4EV. If you are testing spotmetering, then use spot metering for the baseline. Normally i'd expect the exposures of 3 and 1/3rd stops above the (spot) metered settings to fully clip (Raw ADU values do not increase with increasing exposure).

Now use a tool like Rawnalyze (or IRIS which is a bit more challenging to use) to inspect the maximum Raw values of the different exposures, and there is your answer; how many EVs above highlight metering does the camera offer.

Thanks for that succinct explanation - much better than I could have explained it. I used a very similar procedure.

Second, the highlight in the scene may be smaller than the metering spot. Again use you judgement and reduce the suggested exposure with an additional amount.

This, I believe, is one of the most important, yet also one of the most overlooked aspects of this method.
 

Doug Kerr

Well-known member
Hi, Alex,

Bart wrote (someplace):
One takes a series of Raw images with increasing exposure of a uniformly lit surface. It's best to use an aperture of e.g. f/8 to reduce vignetting as much as practical, and slighly defocus the lens to avoid picking up any surface detail. So everything is on manual, the lens and the camera shooting mode. Now increase the exposure time in 1/3rd stops starting at, say, a metered exposure +2EV up to say metered exposure +4EV. If you are testing spotmetering, then use spot metering for the baseline. Normally i'd expect the exposures of 3 and 1/3rd stops above the (spot) metered settings to fully clip (Raw ADU values do not increase with increasing exposure).

That would simplistically tell us that the "calibration" of the exposure control system (using spot metering) is such that for a metered exposure of a uniform-luminance scene the resulting photometric exposure would be about 0.1 Hsat (that is, 10% of the saturation photometric exposure).

For comparison, if we consider the calibration of the exposure control system to be that prescribed by the ISO standard for such, and assume that the exposure index is consistent with the ISO definition of ISO speed, we would expect a photometric exposure in such a test case of about 0.125 Hsat (12.5% of the saturation photometric exposure).

However, Canon normally rates the ISO speed of its cameras in terms of ISO SOS (ISO Standard Output Sensitivity), a value that will be about 1/2 stop lower than the ISO speed.

In a camera with an "ISO standard" exposure control system calibration, using as the exposure index the ISO SOS value for the sensor, we would expect, for our uniform-luminance test subject, a photometric exposure that was about 0.18 Hsat (18% of the saturation photometric exposure).

That translates to a "headroom" of about 2.5 stops above the average luminance of the metered portion of the scene.

The photometric exposure that relates to the result Bart tells us to expect is about 0.85 stop lower than that. Said another way, we would expect to see the hallmark of about 0.85 stop greater headroom than that (that is, about the 3.33 stops that Bart's test directly reveals).

Best regards,

Doug
 

Asher Kelman

OPF Owner/Editor-in-Chief
In addition, my advice to relatively 'underexpose' high ISO shots by one or two EVs, still stands. It will reduce the output noise level and give extra highlight detail.

This is under-exposure to get extra highlight details and less noise. Alex, (who has not specifically addressed highlight detail preservation), seems to be suggesting a different strategy, but just to reduce noise and seemingly get an otherwise identical image:

For the same amount of light, Alex says that the one stop extra higher ISO will give lower noise after pulling back the over-exposure in RAW processing.

So for my stage shots in poor overhead light, what's your current recommendation based on the histogram on the LCD. We want highlight details and reduced noise.

OTOH, if I use a handheld spotmeter (or the camera set to spot as you have described), directed to the forehead hands of a performer, then how can I best use that reading to optimize detail and have the minimum of noise using data on the dynamic range of the camera at different ISO settings? The highlight measured in the spotmeter will be presumably matched with an exposure recommendation for 18% grey so we have headroom to overexpose.

Asher
 

Doug Kerr

Well-known member
Hi, Asher,

I haven't followed all the suggestions here and tried to reconstruct the "theory" behind them.

So, working from such a benighted condition, my simplistic view is that if one's principal emphasis is on good shadow detail (with best noise performance in the shadows), one should shoot with the hottest photographic exposure that is usable without blowing out highlights.

Bart's tests seem to suggest that one would fully press this (which one certainly shouldn't do) by spot metering on a representative highest-luminous spot in the scene and using an exposure compensation of +3.33 Ev.

So if we reserve 1/3 stop of margin, that would suggest shooting at the metered exposure (metering off the representative highest-luminous spot) with an EC of +3 Ev.

Now if in fact the overall Canon "scheme" places the metered spot at 0.18 Hsat (as evaluation of some other data suggests), and allowing about 1/3 stop of margin, we should shoot at the spot-metered exposure with an EC of about +2.33 Ev.

What will actually work best for you I wouldn't try to guess from here.

Note that I don't ever say "18% gray" (a graphic arts term referring to a reflectance, or relative luminance, of 82%). Most discussions of metering concepts or exposure strategies that use that term (even granting that the speaker means "18% reflectance" or "18% of maximum luminance" or something like that) do not actually describe any specific technical situation.​

Best regards,

Doug
 

Ken Tanaka

pro member
Thanks very much, Alex, for your detailed treatise in response to my inquiry. Interesting stuff. You're obviously enjoying this undertaking in a thoughtful, earnest manner. So I tip my hat and wish you well.
 

Doug Kerr

Well-known member
Hi, Alex,

In the wake of your work, I thought I might do a little,probing around in similar regions. I decided I had better refresh my understanding of Rawnalyze (which I hadn't used for some while) and, in the course of that, I thought I had best confirm that I have the latest version (which I suspect I haven't).

But it seems as if the source site for the program has evaporated.

Do you know anything about that?

Thanks.

Best regards,

Doug
 

Alex Weiss

New member
This is under-exposure to get extra highlight details and less noise. Alex, (who has not specifically addressed highlight detail preservation)
Yes, I'm sorry. I just answered Ken's question.

For the same amount of light, Alex says that the one stop extra higher ISO will give lower noise after pulling back the over-exposure in RAW processing.
Yes, but the fact is most noticeable when you compare with lower ISOs. At higher ISOs (e.g. ISO1600 vs ISO3200), you might reach a point of diminishing returns; if the effect is still there, it will be much less noticeable. Additionally, there seems to be a cut-off point (after which ISO is implemented differently and the effect thus isn't there anymore). That is camera-dependent.

I don't have my measured data here right now, but I'll check later. IIRC, I even stopped measuring after ISO1600, so I might have to take a couple of extra exposures.

Your other question regarding proper exposure of highlights has already been answered by Bart, I believe. I'd be curious to hear if you've tried it out already and if so, if it worked.

Bart_van_der_Wolf said:
In addition, my advice to relatively 'underexpose' high ISO shots by one or two EVs, still stands. It will reduce the output noise level and give extra highlight detail.
That is very interesting. So you underexpose and then push it in post?

Ken Tanaka said:
Thanks very much, Alex, for your detailed treatise in response to my inquiry. Interesting stuff. You're obviously enjoying this undertaking in a thoughtful, earnest manner. So I tip my hat and wish you well.
Thank you, and same to you!

Doug Kerr said:
But it seems as if the source site for the program has evaporated.

Do you know anything about that?
Yes, unfortunately it seems that vanished some time ago. Somebody sent it to me, let me know if you want. However, I used Iris for my analysis anyways. It has more features than Rawnalyze, is also free, and, best of all, still around :) You can find it here: http://www.astrosurf.com/buil/us/iris/iris.htm
 

Doug Kerr

Well-known member
Hi, Alex,

There is much useful information in the "help" file for Rawnalyze. (If by any chance you don't have it, let me know, and I'll send it.)

Here are some interesting excerpts (emphasis in blue is added here):

The pixels of some cameras can reach the maximum value representable with the given bit depth; for example the pixel values of the Canon 20D can reach 4095. Other cameras' pixels may clip at a lower level; the clipping level may depend on the ISO, on other factors, even on the particular copy of the camera model. It often depends on the channel (pixel color); even differences between the two green positions within the color filter array are not seldom. Some cameras' pixels (sometimes only certain channels) do not clip at a single level, but in a level range; this is the range of non-linearity. The pixels have to be regarded as clipped from the lower limit of this range, because the values within the range do not represent meaningful information about the image.​

This sounds like "soft clipping".

Note, that the numerical pixel value range is not always the same as the numerical range of the bit depth. Some cameras create raw data, which fully occupies the range of the bit depth, while the numerical pixel value range with some other cameras (for example Canon DSLRs) can be substantially smaller than the range given by the bit depth. In some cases the range depends on the ISO and/or on the channel. The actual range used in the histogram is declared in the camera's description, which is hard coded in Rawnalyze; it is equal to or slightly greater that the initial white point, and it is always the same for all three channels, even if the saturation levels of the channels are different.​

On the one 40D raw file I have yet looked at, the histogram is labeled: Range 0:16383 (which is 2^14-1, the highest 14-bit value).

With that file loaded (and Rawlalyze working in the context of its "hard-coded" parameters for the 40D, the default white point is 16220 and the default black point is 950 (these automatically appear in the control boxes when the file is loaded.

I have to figure how the Canon "offset" fits in with this. It may be that all these values are from an origin at that offset (but see below). (Interestingly enough, notwithstanding the general gigantic level of detail in the "help" file, so far I don't find any recognizable reference to that matter. It may be there but just in some other frame of reference. But we all know how easy it is to write "everything" about a machine except where the power switch is.

Perhaps the offset situation is different in the 40D. (The "range" of the histogram suggests that the values are not from an origin at the offset.) In the one file I am looking at, there are values as low as 821 on the scale in use (as discerned by manipulation of the black point setting)

Well, that's some thoughts for now.

I will do some playing with the 40D later today. I need to make some test CR2 files with "interesting" distributions of H ("all black", "all gigantic", "about 2/3 black and 1/3 gigantic", and so forth, plus a set with a range of exposures "above metered" by different fractions of a stop.

Best regards,

Doug
 

Doug Kerr

Well-known member
Hi, Alec,

I note that Martinec in his paper refers to an offset of 1024 in Canon 14-bit raw files.

Best regards,

Doug
 

Alex Weiss

New member
Hi, Alex,

There is much useful information in the "help" file for Rawnalyze. (If by any chance you don't have it, let me know, and I'll send it.)

I don't have it, so I'd be grateful if you could send it to me.

The pixels of some cameras can reach the maximum value representable with the given bit depth; for example the pixel values of the Canon 20D can reach 4095. Other cameras' pixels may clip at a lower level; the clipping level may depend on the ISO, on other factors, even on the particular copy of the camera model. It often depends on the channel (pixel color); even differences between the two green positions within the color filter array are not seldom. Some cameras' pixels (sometimes only certain channels) do not clip at a single level, but in a level range; this is the range of non-linearity. The pixels have to be regarded as clipped from the lower limit of this range, because the values within the range do not represent meaningful information about the image.
Very interesting. I suspected something like that. However, there still seems to be something off, as I did not get values within a certain range, but every single time only two distinct values.

On the one 40D raw file I have yet looked at, the histogram is labeled: Range 0:16383 (which is 2^14-1, the highest 14-bit value).
Hm, even more interesting. My values have never reached the max (and because of Canon's offset also not 0). Are you sure that this is the range of the actual RAW data and not the result of some scaling Rawnalyze performs with the white and black point setting (if it indeed does that; see below)?

Perhaps the offset situation is different in the 40D. (The "range" of the histogram suggests that the values are not from an origin at the offset.) In the one file I am looking at, there are values as low as 821 on the scale in use (as discerned by manipulation of the black point setting)

I have never fully understood what the black and white point settings do in Rawnalyze (but then again, I didn't have a help file). Do those settings simply operate as thresholds of sorts and discard everything below and above them and then scale the data? Uncertainties like that made me move on to Iris...

Couldn't the low values simply be noise at very dark tones that fluctuates around the offset?

About the offset: you are undoubtedly aware of Emil Martinec's site. In it, he also gives a brief description of the offset. You can easily determine your offset with a couple of black frames. In my case, it was 2048, I assume it will be 1024 in yours.

I will do some playing with the 40D later today. I need to make some test CR2 files with "interesting" distributions of H ("all black", "all gigantic", "about 2/3 black and 1/3 gigantic", and so forth, plus a set with a range of exposures "above metered" by different fractions of a stop.

Yes, please do! That should be very informative.

P.S. For closer analysis of histograms, I recommend ImageJ, as it also prints the histogram as a list.


EDIT: I just saw your post about Martinec's papers. So you're obviously aware of his site :) However, I believe that the offset is also dependent on the model; at least for the 550D, it is 2048.
 
Top