Histogram: Luminosity and Color

[Colin Southern]

I hate to weigh in at this point, after all the detailed discussion, but I actually think this can be simplified. Let me try:

1. Moving the histogram to the right increases the S/N ratio. That is often a good thing to do, but it is not always costless.
2. Whether one wants to move it, and how far, is a matter of judgment.
3. One factor that ought to influence #2 is how bad noise is likely to be if you don't shift the histogram. That will depend on a bunch of things, but in particular, the camera body, the ISO, and how much of the histogram is far to the left.

E.g., Colin, your case is low noise (base ISO, good camera, studio lighting), mass in the midtones, and concern with rendering of the midtones, right? So you don't shift. In other cases, noise in the shadows may be a big deal, and shifting makes sense. And, of course, if the histogram shows that the image subsumes most of the dynamic range of the sensor, you can pretty much forget about it unless you intend to combine exposures.

Re camera bodies: I have two, a 50D and a 5D3. The 50D is a much noisier camera, so I worry much less about the location of the histogram with the 5D3, particularly at low ISOs and when the mass of the histogram is not too far down from the right. Consistent with what Colin says, with the 5D3 under good lighting conditions, being somewhat below the top of the histogram doesn't seem to make much if any practical difference. But that is exactly when it shouldn't.

I don't personally care whether people use ETTR to mean "shift to the right" or "shift all the way to the right," except that people may need to clarify their use just to communicate.

For the great bulk of the photography I do, I don't think it needs to be any more complicated than this.

Am I missing something?

"A" for effort Dan, but I still think that there's assumptions creeping in like "that is often a good thing to do"; I don't want to split hairs, but I would argue that it's more a case of "most often it's not needed because the advantages promised are only theoretical, not real-world. Real world benefits of ETTL are more the exception than the rule with modern cameras. The founding article Manfred referenced is now 11 years old - and sensor technology has changed a lot in that time. That's my biggest gripe (in-fact my only gripe) with it; people keep recommending it because the theory sounds good, not realizing that the world has turned quite a few times since, and it's no longer as relevant as it once was, most of the time).

These days we don't need to ETTR often - nor have a person in front of our vehicles with a red flag to warn people of a vehicle approaching at "high speed" - nor cut-off broken limbs. Times have changed!

[Mike Buckley]

Does anyone know what Michael Reichman says about ETTR now?

[Mike Buckley]

Ahhhhhhh. I see that as recently as three years ago, he was not only still advocating ETTR but also complaining that it wasn't being included automatically in the cameras.

I hope the technically minded folks participating in this thread and elsewhere understand why the non-technically minded people such as myself and Maria, the OP of this thread, don't have a clue about who to believe. If you folks don't get that, you really don't "get it."

As an example, you can talk forever about gamma this and non-linear that, but that talk to the non-technical among us is no more convincing than looking at a graph without understanding the math, assumptions, and perspective behind it. That's because even the non-technical among us understand that graphs can deceive, intentionally or otherwise. Either way, it's deception.

[Colin Southern]

Ahhhhhhh. I see that as recently as three years ago, he was not only still advocating ETTR but also complaining that it wasn't being included automatically in the cameras.

And for what it's worth, an "ETTR Metering Mode" has been one of the many ideas that I've forwarded to Canon's Chuck Westfall over the years. Personally though, I can understand now why they wouldn't want to include it. At the end of the day it serves somewhat of a niche purpose, and used incorrectly creates more problems than it solves. Additionally, as it essentially says to the camera "throw away your usual 2-stop safety margin" - and with the imprecision of metering zones - it's likely that we'd end up with many more clipped highlights (and the ensuing collective moaning from the great unwashed "I didn't pay all this good money for a camera that clips highlights - Canon camera's are sh*t").

I hope the technically minded folks participating in this thread and elsewhere understand why the non-technically minded people such as myself and Maria, the OP of this thread, don't have a clue about who to believe. If you folks don't get that, you really don't "get it."

People say that about a lot of things, but I think it only goes so far. If you're not understanding some technical aspect then just ask - I'm sure many will be glad to help. One things for sure - the more you know the better the decisions you can make.

As an example, you can talk forever about gamma this and non-linear that, but that talk to the non-technical among us is no more convincing than looking at a graph without understanding the math, assumptions, and perspective behind it. That's because even the non-technical among us understand that graphs can deceive, intentionally or otherwise. Either way, it's deception.

OK - good case-in-point. In this context just think of gamma as being "getting the midtones looking right" (no more, no less). Gamma adjustment is what it's technically called, adjusting midtones is what it actually does.

[Inkanyezi]

So I might have misunderstood Chuck Gardner's idea then? If I move the bulk of the histogram right, by exposing more, I will blow those tiny bits of blank white even more, although like there is no such thing as overexposing a highlight more, just as romance languages don't accept comparatives of too much (much too much).

I thought of it as a method to preserve highlights, just as is done with Nikon's "Active D-Lighting". That the brightest highlight where rendition is important is the determinant for what has to land in the lower right corner of the histogram. Obviously my "misunderstood" method will accept that brighter highlights clip, for example specular highlights or light sources.

And I came to wonder about whether it could be a better idea, as I found that my Panasonic Lumix G1 exposed up to two stops more than Sunny 16 in bright sunlight. Measuring the highlights is a Good Thing, when an important part of the subject is very bright and needs to be held back in exposure. When I tried it out first, I used my balcony as a subject, where a white plastic chair became partially clipped, showing those colour fringes of unlinearity when trying to bring them down to show structure and detail. I was disappointed then by what's called "intelligent automatic exposure", a feature that the Lumix G1 was said to have.

Concluding that iAuto was not so intelligent after all, I went out for a stroll taking images according to the Sunny 16 rule. Lo and behold, they came out perfectly exposed to my taste, although a bit dark in standard conversion, needing PP adjustment (gamma, brightening shadows and midtones). Only some small specular highlights were blinking.

After that, I tried out just where a spot reading would place the measured spot right there, down in the lower right corner. The caveat with it however, is that I don't know exactly how large the spot is in the camera. The camera does not tell me that, as the Pentax or Minolta hand held meters do.

And there are occasions where I do adjust exposure down, contrary to the ETTR idea as first presented. Maybe it is that I don't like the fringing at the edges of specular highlights, which occurs more conspicuously when exposure is taken further to the right. So in my images in the snake pit, I have used matrix metering and compensated with minus, which is the twentieth century method in the twentyfirst century, effectively applying the zone system. And I don't give a damn about noise in almost pitch black areas. The highlights in those images are right there, in the lower right corner of the histogram, and the bulk of the histogram is way left.

Do not open the link below if you suffer herpetophobia and would be disturbed by seeing a snake that eats:
http://uploads.ifokus.se/uploads/f39...3/snok-003.jpg

The histogram of the image, which in my opinion is exposed to the right, looks like this:

[DanK]

The histogram of the image, which in my opinion is exposed to the right, looks like this:

that's how I would use the term as well.

[ajohnw]

I would be more inclined to call it a correct exposure and the scene fits the dynamic range available. Actually if you look at the histogram with the GIMP there are slight signs of clipping when the view is enlarged. So slight it's hardly worth even mentioning and seems to be mostly in the green channel and some red. These tiny bits could probably be bought in from raw. Once it's in a jpg there isn't really any way of telling if an image is clipped other then the bright end petering away to zero and that needs a large histogram or seeing that it actually falls short of clipping.

Perhaps an example that most definitely isn't ETTR might help. On this one I think it would be entirely the wrong thing to do as there would be a lot more difficulty in pp'ing to get the colours correct. This is too dark yet I only have 40 odd bits of highlight left for brighting and contrast adjustments. In principle 1 stop down from full house sits at 128 bits - Not going to try to work out what 40 is log wise. This touches on a point Colin made some time ago - OK in theory adjust for the full colour space but in practice it's best to leave some space for further work. Worth thinking about when developing from raw with what I use anyway. Automatic brightening might make life easier or adjusting to use up all of the colour space might seem fine but may not be the best thing to do with everything.



I suppose it could be argued that the same result for adjustment could be obtained from raw by using ETTR and then developing it darker but then there is the risk of loosing highlights given the histograms available when people actually shoot.

One of Dcraw's FAQ asks about something that relates to this subject..

Why is 16-bit output dark / unreadable?
If you want pretty pictures straight out of dcraw, stay with 8-bit output. 16-bit linear output is the best raw material for professional image editors such as Photoshop and CinePaint, but it's no good for most image viewers.

John
-

[Glenn NK]

It seems (to me at least) that Michael Reichmann's articles on ETTR has been interpreted in different ways.

Much of the discussion in posts in this thread discuss noise in the darker parts of the image. Although there will be some noise in the image at the right side of the histogram, the S/N ratio is large enough that noise isn't often a problem in the bright areas - at least I don't find too much in the bright areas, but I digress.

When I read the first ETTR article published in 2003, what I took away from it was the following (copied from the most recent one http://www.luminous-landscape.com/tu...exposure.shtml

A typical consumer DSLR recording 12 bits per sensel is able to record up to 4,098 separate tonal values.
If we assume a 10 stop dynamic range this is how this data is distributed...
The brightest stop = 2048 tonal values
The next brightest stop = 1024 tonal values
The next brightest stop = 512 tonal values
The next brightest stop = 256 tonal values
The next brightest stop = 128 tonal values
The next brightest stop = 64 tonal values
The darkest stop = 32 tonal values

As can be seen, each stop from the brightest to the darkest contains half of the data of the one preceding it.

What I took away from reading the articles is that the brightest stop contains twice the number of tonal values as the next brightest stop. And it seems to me (correct me if I'm wrong), having more tonal values might be a good thing. At least is makes me feel good.

A potential trade-off comes at the very right edge of the histogram where "non-linearity" of the sensor comes into play. If one wants absolute colour accuracy, then it seems wise to stay away from this area and not use strict ETTR. If absolute colour accuracy isn't the most important, then expose more to the right and catch more tonal values. By the way, what is colour accuracy anyway - considering that the human eye/brain connection has so much variability, and hence we all see things differently?

In any event, this topic is recurrent, it will recur again, and it will be argued some more. And few, if any, will be convinced to change their minds. So it isn't really worth arguing about - unless one insists.

Have a good day - or night - whichever comes first.

Glenn

[Colin Southern]

What I took away from reading the articles is that the brightest stop contains twice the number of tonal values as the next brightest stop. And it seems to me (correct me if I'm wrong), having more tonal values might be a good thing. At least is makes me feel good.

In THEORY it's good, but in PRACTICE, it depends; if you're dealing with tones that are 6 stops down and only represented by 32 levels, it becomes a moot point if all of those 32 potential levels are 0. Therein lies the "fine print"; if you take a photo of a reflective scene - 12 bit ADC - 2 stop safety margin, you end up with the following

- 2048 levels - wasted in safety margin

- 1048 levels - wasted in safety margin

- 512 levels - Describes "Bride's dress"

- 256 levels - describes high mid-tones

- 128 levels - describes low-midtones

- 64 levels - describes shadows ("grooms suit")

Can the human eye differentiate between 64 levels of shadow detail? No.

All of the above is somewhat over-simplified as I'm not taking things like gamma into account, but the thing I wanted to point out it that (a) a reflective scene (and print) typically only contains a maximum of 4 stops and (b) a monitor typically only displays a maximum of around 6 stops. Additionally, the human eye can only decerne around 200 levels on a monitor. or put another way - when you're dealing with objects that are so dark they appear black, it doesn't matter how few bit represent them.

Where things get more interesting is when we want to "distort" the data to reveal detail that's not normally apparent eg you shoot a building in full sun - revealing about 4 stops of information by default, but you also need to reveal detail of areas that are in shadow and are typically another 3 stops or so down. In the above example these would be represented by only 8 levels, and may (or may not) start to look pretty blocky (due to fewer bits to represent them) and noise (due to getting closer and closer to the noise floor), but even then, if they're so far down that they're only represented by 8 bits then they're probably going to want to be kept pretty black in the final image (because they'll still be dark dark colours).

At the end of the day, how much of a real-world problem does this represent? Answer - it's not normally a big issue because (a) most cameras are now resolving to 14 bit (4 times as much info - but not necessarily meaningful info - as 12 bit ADC), which ironically, is about what ETTR typically gave anyway, and (b) modern sensors have lower noise floors anyway (thus higher dynamic range capability).

Case in point - during the week I got a call from a client to shoot a building in full morning sun. Worst possible time of day to shoot, but that's what the client needed. Hard shadows everywhere. To cover my but I shot a 5 stop bracket - and what did I end up using in EVERY case? ... the normal exposure because (a) the highlights weren't blown and (b) there was sufficient clean data to be able to give the images a healthy dose of fill light to lighten many of the areas that were literally in shadow. This was THE classic scenario for ETTR - and yet I didn't need it.

[ajohnw]

It seems (to me at least) that Michael Reichmann's articles on ETTR has been interpreted in different ways.


A typical consumer DSLR recording 12 bits per sensel is able to record up to 4,098 separate tonal values.
If we assume a 10 stop dynamic range this is how this data is distributed...
The brightest stop = 2048 tonal values
The next brightest stop = 1024 tonal values
The next brightest stop = 512 tonal values
The next brightest stop = 256 tonal values
The next brightest stop = 128 tonal values
The next brightest stop = 64 tonal values
The darkest stop = 32 tonal values

As can be seen, each stop from the brightest to the darkest contains half of the data of the one preceding it.

What I took away from reading the articles is that the brightest stop contains twice the number of tonal values as the next brightest stop. And it seems to me (correct me if I'm wrong), having more tonal values might be a good thing. At least is makes me feel good.

Glenn

NO. When you look at these numbers you are seeing a linear representation of stops and having 1024 tonal steps at some level is pointless as out eyes wouldn't be able to detect the changes. Loosely speaking we see in stops. In other words if asked to grade equal steps in light intensity each step up in intensity doubles the previous one. Modern science has made this a more complicated law but the doubling is fairly close.

One thing that can be used to illustrate it is stars. Long ago they were split into magnitudes - equal steps in brightness as seen via the eye Later the brightness was measured and found to double for each step just as stops do. At these brightness levels it's doubtful if we can split of 1/10 of a stop and the brightest stars as you may have noticed are rather bright. As measured the sun is about 400,000 time brighter than the moon but in terms of magnitudes it's 14 times brighter - ie 14 stops on the scale that was chosen. As the brightest star needed numbers brighter than the original scale could represent it seems that we aren't so good at seeing differences at the bright end of things.

Starting at 32 in the table as far as a camera is concerned may be sensible or it might not be. I'd guess it needs to be higher actually once it's processed but even then

32,64,128,256,512,1024,2048,4096 so 12 bit "12 stop" cameras actually capture 8 stops. 14 bit cameras could be exactly the same. No reason why they shouldn't be. The advantage is that having a count of 64 at the 1st sop means that it can be brightened and have the contrast increased more before posterisation occurs. Hopefully noise levels will be lower too on these cameras so that these extra steps can actually be used. Pixels get smaller and lower ISO settings disappear or are faked in some cases but that's another story.

None linear sensors - carry on by all means.

John
-

[Colin Southern]

Starting at 32 in the table as far as a camera is concerned may be sensible or it might not be. I'd guess it needs to be higher actually once it's processed but even then

32,64,128,256,512,1024,2048,4096 so 12 bit "12 stop" cameras actually capture 8 stops.

I understand what you're saying John, but saying it like that is just asking for trouble; in reality no (DSLR anyway) camera made recently or even not so recently would stop at 8 stops DR at base ISO; in reality, dynamic range and sensor resolution aren't really related. The first is "the height of the staircase" whereas the 2nd is "the number of steps on the staircase"

The reality is - for DSLRs anyway - that these days we're dealing with 12 or more stops of DR and 14 bit ADC resolution.

[ajohnw]

I understand what you're saying John, but saying it like that is just asking for trouble; in reality no (DSLR anyway) camera made recently or even not so recently would stop at 8 stops DR at base ISO; in reality, dynamic range and sensor resolution aren't really related. The first is "the height of the staircase" whereas the 2nd is "the number of steps on the staircase"

The reality is - for DSLRs anyway - that these days we're dealing with 12 or more stops of DR and 14 bit ADC resolution.

Sorry Colin I wont accept that because they haven't. As far as looking at the images goes the bottom end is of limited use. If others want to accept it that's fine by me.

Someone who is being pedantic might say that the sensor does have 14 bits dynamic range and it does but the whole object is to view the image and if that could be done with a single 14 bit dynamic range image multiple exposure HDR in most instances would not be needed. Sensors also vary from one camera model to another. Yours for instance will be one of the best on the planet ( In my view probably the best ) My D7000 has a 14bit a/d - need I say more? And that's just one aspect. Well I wont say anymore anyway. People are always seduced by numbers and often don't consider the whole shooting match.

John
-

[Colin Southern]

Sorry Colin I wont accept that because they haven't. As far as looking at the images goes the bottom end is of limited use. If others want to accept it that's fine by me.

OK, let's compromise; 12 to 13 stops (not bits) DR by internationally accepted standards, which equates to 8 to 9 stops on the "AJW" scale.

[MariaMaria]

I should have titled this thread "To ETTR or not to ETTR…That is the Question!!"

[Glenn NK]

in any event, this topic is recurrent, it will recur again, and it will be argued some more. And few, if any, will be convinced to change their minds. So it isn't really worth arguing about - unless one insists.:d

q.e.d.

[Manfred M]

I should have titled this thread "To ETTR or not to ETTR…That is the Question!!"

Maria - that would have resulted in very similar set of comments. Perhaps "When to ETTR and when not to" might be a better question.

[Manfred M]

The last time I checked, John, a sensor is an analog device. Yes, your output is digital, that is true.

I find Colins argument far more compelling; the height of the staircase, versus the number of steps. 14-bits represents the colour range that the camera can output, rather than what the sensor itself is capable of capturing. Nicely said, output of a sensor that is 1-bit (0=black; 1 = white) can also be represented as a 14-bit data string; 00000000000000 and 00000000000001.

[benm]

(a) a reflective scene (and print) typically only contains a maximum of 4 stops and (b) a monitor typically only displays a maximum of around 6 stops.

Agreed. So why is there so much emphasis on cameras that have greater and greater dynamic range? There does not seem to be any current device that can actually display all that DR. And even if it could, if one creates an 8-bit jpg you get 8 stops DR maximum, probably less.

[Manfred M]

Agreed. So why is there so much emphasis on cameras that have greater and greater dynamic range? There does not seem to be any current device that can actually display all that DR. And even if it could, if one creates an 8-bit jpg you get 8 stops DR maximum, probably less.

Simple explanation - I think this really goes back to Colins other post when he looks at the height of a staircase, versus the number of stairs. The image capture reflects the height of the staircase, whereas the dynamic range shows the number of steps.

In capture, we try to get as much dynamic range as possible, from the brightest whites to the darkest blacks. This means our cameras need to have a large enough dynamic range to get us the range of values without blowing out the white values (too much) or blocking up shadow detail (again too much). When we display; the white and black values are still the contraints, what we end up compressing is the stuff in the middle.

[Colin Southern]

So why is there so much emphasis on cameras that have greater and greater dynamic range?

Because we compress the dynamic range that we capture into something we can print or display. That sounds complicated, but in practice, it's what occurs when we use something as simple as a fill light slider in post processing.

There does not seem to be any current device that can actually display all that DR. And even if it could, if one creates an 8-bit jpg you get 8 stops DR maximum, probably less.

That's a common misunderstanding - an 8 bit representation doesn't limit the image to an 8 stop dynamic range any more than limiting the number of steps on a staircase limits the maximum height of a staircase.