Expose to right, it is as simple as

[madmanchan]

Exactly.  Once any significant non-linearity has been introduced into the system (and clipping is certainly a big one!), then clearly a simple linear operator like white balance is no longer correct.  (This is why it's hard to apply white balance correctly to JPEGs.  They have undergone potentially many non-linearities already in their processing.)

[Daniel Browning]

Is there any benefit at all to ETTR for a jpeg file?

Yes, but only if you correct the brightness in-camera before the JPEG is written. For example, if you are shooting a JPEG at ISO 100 and notice that you have 1 stop of unused highlight headroom, your camera might have an "ISO 50" (or "L") mode which is really just a digital "-1 EC" behind the scenes (in the camera). If so, you can enable that to get the benefits of ETTR in your JPEG.

[01af]

"Expose to right, it is as simple as that?"

No, it isn't.

Agree with the conclusion, we should to expose to the right and the cameras should support this.

No, they shouldn't. Cameras must not try to be smarter than photographers ... at least photographer's cameras shouldn't. Exposing To The Right (or, as we used to say in the pre-digital era, exposing for the highlights) is the right thing to do when the subject's dynamic range is about equal to or moderately greater than the dynamic range the camera can handle—usually. Exceptions can, and will, happen. And it usually is not the right thing to do when the subject's dynamic range is significantly smaller or vastly greater than what the camera can handle. So no, it's not as simple as that.

People bluntly advocating ETTR usually are ignoring two things. First, some highlights are not worth preserving—but it's you who's got to decide, not the camera. Second, ETTR will optimize separation of tones but sacrifice separation of colours. So if the subject's dynamic range is small then don't push it up all the way to the right. Instead, leave it centered ... or push it up half-way between center and right.

... if you are shooting a low-contrast scene, using your exposure meter and centering the image data in the center of the histogram is a suboptimal idea ...

No, that's not necessarily true. For the optimal exposure, you'd need to balance dynamic range versus colour range. Where the best balance is depends on your image content. If you plan to convert your capture to black-and-white then by all means push the tones as far to the right as you possibly can. But when you plan to capture a wide range of pastel and saturated colours then better back off a little.

That's why an automatic ETTR exposure mode would make a lot less sense than most people think. There are situations where it would be useful, but then there others where it's not.

[hjulenissen]

Exposing To The Right (or, as we used to say in the pre-digital era, exposing for the highlights) is the right thing to do when the subject's dynamic range is about equal to or moderately greater than the dynamic range the camera can handle—usually. Exceptions can, and will, happen. And it usually is not the right thing to do when the subject's dynamic range is significantly smaller or vastly greater than what the camera can handle. So no, it's not as simple as that.

I dont understand the last sentence. Scenes of low dynamic range woud be candidates for "ETTR" in my book.

People bluntly advocating ETTR usually are ignoring two things. First, some highlights are not worth preserving—but it's you who's got to decide, not the camera. Second, ETTR will optimize separation of tones but sacrifice separation of colours. So if the subject's dynamic range is small then don't push it up all the way to the right. Instead, leave it centered ... or push it up half-way between center and right.

What do you mean by "separation of colors"?

That's why an automatic ETTR exposure mode would make a lot less sense than most people think. There are situations where it would be useful, but then there others where it's not.

What is usually requested is for cameras to provide better feedback in order for the photographer to do the right choices wrgt exposure - for instance ETTR if that is deemed appropriate.

-h

[ErikKaffehr]

Hi,

ETTR is about optimizing dynamic range. The inclusion/exclusion of highlights is not a part of ETTR. ETTR essentially says that it is preferable to maximize the number of photons detected by the sensor.

I have not seen any evidence that color separation would be affectedby ETTR, on contrary, noise may reduce color separation and ETTR minimizes noise (it is all it does). Actually it does not minimize noise but maximizes signal to noise ratio, to be correct.

Color separation is in my view only a function of the spectral charectiristics of the color grid array and the color transformation matrix, both of which are independent of exposure.

Best regards
Erik

People bluntly advocating ETTR usually are ignoring two things. First, some highlights are not worth preserving—but it's you who's got to decide, not the camera. Second, ETTR will optimize separation of tones but sacrifice separation of colours. So if the subject's dynamic range is small then don't push it up all the way to the right. Instead, leave it centered ... or push it up half-way between center and right.

No, that's not necessarily true. For the optimal exposure, you'd need to balance dynamic range versus colour range. Where the best balance is depends on your image content. If you plan to convert your capture to black-and-white then by all means push the tones as far to the right as you possibly can. But when you plan to capture a wide range of pastel and saturated colours then better back off a little.

That's why an automatic ETTR exposure mode would make a lot less sense than most people think. There are situations where it would be useful, but then there others where it's not.

[PierreVandevenne]

Second, ETTR will optimize separation of tones but sacrifice separation of colours.

Could you elaborate a bit on that? I am not sure I understand.

No, that's not necessarily true. For the optimal exposure, you'd need to balance dynamic range versus colour range. Where the best balance is depends on your image content. If you plan to convert your capture to black-and-white then by all means push the tones as far to the right as you possibly can. But when you plan to capture a wide range of pastel and saturated colours then better back off a little.

You seem to imply increasing dynamic range hurts color range. By which mechanism? ETTR doesn't inherently increase the dynamic range of a system, it just exploits it optimally. If the captured scene has too much dynamic range for your taste, nothing prevents you from post-processing to limit it. OTOH, if you haven't captured enough of the DR of a scene, there's no way to get it back after the capture. This being said, although your suggestion is a bit imprecise, I don't dismiss it since "ETTR to the max" could very well end up on the upper non-linear response zone of the sensor.

[01af]

The inclusion/exclusion of highlights is not a part of ETTR.

Sure it is. As a matter of fact, including highlightsand not excluding them is the essence of ETTR. That's where ETTR got its name from.

ETTR essentially says that it is preferable to maximize the number of photons detected by the sensor.

You really want to recapitulate what ETTR actually is. Hint: Maximizing the number of photons detected by the sensor it is not. After all, a capture over-exposed by several f-stops will always detect far more photons than a properly exposed one.

... and ETTR minimizes noise (it is all it does).

Well, that's my point—it is not all it does.

Color separation is in my view only a function of the spectral charectiristics of the color grid array and the color transformation matrix, both of which are independent of exposure.

Your view is wrong.

Simply consider the extremes. The blackest black that an RGB image can produce is RGB(0, 0, 0). No colour separation at all; one colour only at this tone level: black. The same at RGB(255, 255, 255)—only one single colour, white, at the maximum tone level; no variations possible (that's for 8 bits per RGB channel—for higher bit depths adjust the numbers accordingly but the principle will remain the same). The most saturated red, for example, would be RGB(255, 0, 0)—that's a tone much brighter than black but also much darker than white. Colours that are as bright as RGB(255, 255, 255) but at the same time as saturated as RGB(255, 0, 0) simply cannot exist in an RGB system where each channel's range of values is finite. So the variation of possible colours is widest at medium tone levels and narrowest at extremely low or extremely high tone levels. To complicate matters even more, RGB(0, 255, 0) is not the same brightness level as RGB(255, 0, 0), and RGB(0, 0, 255) is yet another brightness level.

Of course, things still aren't as simple as that. First, you're right when you're saying that noise will hurt colour separation, so avoiding noise basically is a good thing. That's why I said, put your histogram's peak half-way between center and right when subject contrast permits and colours are important. Second, a camera's RGB channels usually won't clip all at the same brightness level—that further complicates matters and is another reason not to push your histogram to the farthest right when you don't have to.

So—ETTR sure is a useful rule of thumb generally but still needs some consideration. It is not the gold standard for all situations and circumstances. In most cases ETTR is a good idea but sometimes you must expose beyond ETTR, and sometimes it's better to back off from ETTR by one stop or two. It's just the same as everywhere else in real life: Simple rules aren't.

[MarkM]

The blackest black that an RGB image can produce is RGB(0, 0, 0). No colour separation at all; one colour only at this tone level: black. The same at RGB(255, 255, 255)—only one single colour, white, at the maximum tone level; no variations possible

…and RGB(113, 115,29) is just a single color too—so no variation is possible. Still not sure what you mean by color separation. Naturally, with a single color, whether it's pure black, white, or any other has no variation by definition, that's obvious. Color variation, or separation, or whatever you're calling it doesn't make any sense until you are speaking of more than one color. Since the sensor is capturing linear data, it shouldn't degrade as you push it to the right until you start clipping. Are you just explaining the long way not to clip individual channels?

[ErikKaffehr]

Hi,

There has been a lot of dicussion about this and it seems that there is no consensus about it. I got the impression that most color cast observed comes from clipping one or more of the RGB channels. Another impression I got is that sensors are seen as mostly linear devices. On the other hand it may be that sensors have some kind of shoulder, too. This may also depend on the make of sensor. A shoulder would introduce some color shift.

The included density plot of a 41 step wedge doesn't show much of a shoulder. The photo was made on a simple lightboard so some unlinearity may come from illumination rather than from sensor/wedge.

Another view may be that todays sensors are so good that noise seldom is a large problem at base ISO, so the need to keep noise down may be less important. On the other hand, exposing to the right is the only way of keeping maximal shadow detail. Overdo it and there will be clipping.

Best regards
Erik

This being said, although your suggestion is a bit imprecise, I don't dismiss it since "ETTR to the max" could very well end up on the upper non-linear response zone of the sensor.

[madmanchan]

Colors is absolutely not a problem with ETTR as long as you have not clipped the raw channels.  If you have clipped a raw channel (any one, or multiple ones) in an area of the image that you care about, then all bets are off. 

[hjulenissen]

Your view is wrong.

Simply consider the extremes. The blackest black that an RGB image can produce is RGB(0, 0, 0). No colour separation at all; one colour only at this tone level: black. The same at RGB(255, 255, 255)—only one single colour, white, at the maximum tone level; no variations possible (that's for 8 bits per RGB channel—for higher bit depths adjust the numbers accordingly but the principle will remain the same). The most saturated red, for example, would be RGB(255, 0, 0)—that's a tone much brighter than black but also much darker than white. Colours that are as bright as RGB(255, 255, 255) but at the same time as saturated as RGB(255, 0, 0) simply cannot exist in an RGB system where each channel's range of values is finite. So the variation of possible colours is widest at medium tone levels and narrowest at extremely low or extremely high tone levels. To complicate matters even more, RGB(0, 255, 0) is not the same brightness level as RGB(255, 0, 0), and RGB(0, 0, 255) is yet another brightness level.

Actually, I think that your view is wrong. Can you provide us with images that support your controversial theory, or is it only based on thought-experiments? I think that the arguments presented above indicates that you should think things over once more.

If ETTR meant "expose the image as hot as possible", then you could hypothetically be right: if you manage to clip every single sensel, then all color information will be lost, and no highlight information will be present (in fact, no information will be present at all). I dont think that anyone is advocating such a practice, and if "ETTR guides" seems to indicate so, they may be poorly written.

The lesson that I take from ETTR is to expose as hot as possible _while_ avoiding clipping wherever I find it perceptually necessary. If bright, small highlights are uninteresting to me, then I will clip them (and both their brightness and color will suffer more or less as a consequence). If a larger bright patch is interesting to me (for instance, the sky), I will keep it within the linear region of my sensor, thereby no loss of "color separation" will occur in that part.

The consequence of separating "what exposure looks good on print, given an ideal camera" and "what exposure allows me to capture the stuff that I am interested in with the best possible precision" is that one can (if the situation allows) increase exposure somewhat, and have a better "desired signal"-to-noise ratio. After all, digital raw files can have their exposure adjusted for subjective targets after-hand, but adjusting the SNR after-hand is a lott harder.

-h

[Bart_van_der_Wolf]

Your view is wrong.

Simply consider the extremes. The blackest black that an RGB image can produce is RGB(0, 0, 0). No colour separation at all; one colour only at this tone level: black. The same at RGB(255, 255, 255)—only one single colour, white, at the maximum tone level; no variations possible (that's for 8 bits per RGB channel—for higher bit depths adjust the numbers accordingly but the principle will remain the same). The most saturated red, for example, would be RGB(255, 0, 0)—that's a tone much brighter than black but also much darker than white. Colours that are as bright as RGB(255, 255, 255) but at the same time as saturated as RGB(255, 0, 0) simply cannot exist in an RGB system where each channel's range of values is finite.

Hi 01af,

As others have mentioned, your reasoning doesn't hold because the conceptual model is flawed. You are not alone, I've seen others with similar reasoning, so it's important to address and set straight. RGB(255,0,0) is by definition less bright (~22% Luminance) than RGB(255,255,255) is (100% Luminance), it's a different color.

You probably are thinking of the 3D hull representation of an RGB colorspace which seems narrower at the top and bottom of the 3D shape, but that is a transformation of the RGB coordinates into a colorspace where luminance is (more or less) separated from the chromaticity, while an RGB colorspace records coordinates in an orthogonal 3 axes system where luminance is not one of the axes.

With that hopefully out of the way, low contrast scenes are prime candidates for ETTR because S/N quality can be improved with minor risk of clipping highlights. Color gamut is not an issue, assuming the Raw converter doesn't introduce Hue-twists or unequal blackpoint clipping and such when pulling the exposure in postprocessing.

Cheers,
Bart

[01af]

… and RGB(113, 115, 29) is just a single color too—so no variation is possible.

Are you trolling, or don't you understand the difference between colour and tone? There are no other RGB colours that are as bright as RGB(255, 255, 255)—but here are thousands of RGB colours that have the same brightness as RGB(113, 115, 29).

... because the conceptual model is flawed.

No, it's not.

RGB(255, 0, 0) is by definition less bright (~22 % luminance) than RGB(255, 255, 255) is (100 % luminance), it's a different color.

It is both a different colour and a different tone. There is only one RGB colour that has 100 % luminance but many thousands of RGB colours that have 22 % luminance (or any medium level of luminance). The closer the luminance level is to 0 % or 100 %, the smaller the number of RGB colours at that level of luminance is. Just how hard can it be to grasp this?

You probably are thinking of the 3D hull representation of an RGB colorspace which seems narrower at the top and bottom of the 3D shape ...

It doesn't just seem narrower at the top and bottom—it actually is narrower at the top and the bottom.

... but that is a transformation of the RGB coordinates into a colorspace where luminance is (more or less) separated from the chromaticity, while an RGB colorspace records coordinates in an orthogonal 3 axes system where luminance is not one of the axes.

You probably are thinking that hings are having either colour or luminance. However actually they always have both. In the orthogonal three-axes system—one axis for each of the three colour channels R, G, and B—the luminance axis basically is the space diagonal from the (0, 0, 0) corner to the (255, 255, 255) corner. And near the corners, the RGB colour space is pretty pointy.

[PierreVandevenne]

complicate matters even more, RGB(0, 255, 0) is not the same brightness level as RGB(255, 0, 0), and RGB(0, 0, 255) is yet another brightness level.

OK - yes, we are aware of that. One shouldn't clip anything, everyone agrees. Yes, it can be difficult at times, everyone agrees and some have tried to develop methods to make sure clipping doesn't happen.

About linearity

There has been a lot of dicussion about this and it seems that there is no consensus about it.

Yes, I am afraid this is true. In scientific applications, the linearity of the sensor's response is important, analyzed, documented, etc... In cameras, whyen they don't use sensors that are available otherwise, there's essentially zero information available..

Edit: yes, your chart show a nice response linearity. Since all sensors (either CMOS or CCD) show a certain level of integral nonlinearity, I assume manufacturers take this into account and one shouldn't run into trouble. BTW, those discussions are always interesting in terms of personal education: since I try not to put my foot in my mouth too often (one can always hope ) I usually take the time to review the topic a bit before posting and have learned that in some applications, the variability of the non linearity over time leads to recalibration in some data acquisition protocols. Yet another fascinating example, with the one that Bart posted a while ago from the Dutch professor, that even when one thinks one knows something, there are always mountains of hidden things to learn!

[hjulenissen]

You probably are thinking that hings are having either colour or luminance. However actually they always have both. In the orthogonal three-axes system—one axis for each of the three colour channels R, G, and B—the luminance axis basically is the space diagonal from the (0, 0, 0) corner to the (255, 255, 255) corner. And near the corners, the RGB colour space is pretty pointy.

Who is telling you that you have to expose any more than [255,0,0]*) when doing ETTR?

-h
*)no native sensor output looks like this, but I like your analogy to display-referred color spaces.

[PierreVandevenne]

You probably are thinking that hings are having either colour or luminance. However actually they always have both. In the orthogonal three-axes system—one axis for each of the three colour channels R, G, and B—the luminance axis basically is the space diagonal from the (0, 0, 0) corner to the (255, 255, 255) corner. And near the corners, the RGB colour space is pretty pointy.

You know, you don't care about that. You've just a set of photon counters that return a bunch of numbers. ETTR is about getting the most accurate photon count (succintly the more you collect, the more precise your count is). Whether there is a blue, a green, a red, a SII an OIII, an Halpha filter in front, whether the data will be turned into a pretty picture, a graph, noise, an inverted pyramid or a purple icosahedron or whathever doesn't matter. Obfuscating the topic by taking those numbers and moving a level or three up the conceptual chain isn't helpful. The goal is "obtain the optimal initial numbers" then do whatever you want with them.

If you mean to say "we need to be careful when doing ETTR because we can eventually overflow a set of counters whithout overflowing others thereby creating a color/tone/accuracy/whatever issue", well, I guess everyone agrees and is aware of the issue.

[Guillermo Luijk]

There has been a lot of dicussion about this and it seems that there is no consensus about it.

I have two reasons to think my camera's sensor is very linear in the highlights and therefore colour shifts caused by ETTR are a myth. First beacuse I measured my camera's response, and only when some RAW channel gets clipped, the survivor(s) seems to have some non-linearity:




Second because I have fused dozens of images for HDR, and if exposure is matched correctly between differently exposed shots, the seams are totally invisible, even with zero transition (the transition happens when the RAW data coming from the most exposed shot gets closer to saturation than 0,15EV, and I arbitrarily chose that limit which could have even been higher).

Try to find any visible seam in the image on the left:




So until someone comes to me with non-clipped RAW files displaying any colour shift, I'll stay thinking about it as a myth.

Regards

[MarkM]

The closer the luminance level is to 0 % or 100 %, the smaller the number of RGB colours at that level of luminance is. Just how hard can it be to grasp this?
…
It doesn't just seem narrower at the top and bottom—it actually is narrower at the top and the bottom.
…
And near the corners, the RGB colour space is pretty pointy.

You're pretty mixed up here. And you are using the word luminance incorrectly—it is not the diagonal in an RGB cube, but that doesn't really matter because it's much simpler that you are making it.

Look, imagine a low contrast scene with RGB values between [100, 100, 100] and [155, 155, 155]. It's taking a section right out of the middle of the cube where you say there are more colors. But how many discreet combinations can you make with these RGB numbers?

Now push everything up into the corner so you have RGB values between [200, 200, 200] and [255, 255, 255]. Now how many values do you have? The same, right?

See what's going on there. With a low contrast scene you are not taking a complete cross-section of the RGB cube, you are taking a smaller cube out of the middle of a larger one. In this case with it's own pointy areas around [100, 100, 100] and [155, 155, 155]. This smaller cube 'fits' just fine up in the corner of the big one—they have the same area and shape.

[01af]

You're pretty mixed up here.

Actually it's you who's mixing up things.

And you are using the word luminance incorrectly—it is not the diagonal in an RGB cube ...

The space diagonal in the RGB cube is not the same as luminance, that much is true ... but it basically is some sort of approximation thereof.

... imagine a low-contrast scene with RGB values between [100, 100, 100] and [155, 155, 155].

Interesting thought—you almost pulled me into your delusion. However ... why would a low-contrast scene be limited to to RGB values between (100, 100, 100) and (155, 155, 155)? That does not characterise a low-contrast scene. A limited range of luminance values does. So take all those colours from your triple-100-to-triple-155 sub-cube—and then add all the other RGB values that have the same luminance values as any of the points inside your sub-cube. You'll end up with many more RGB values that definitely do not fit ithe triple-200-to-triple-255 sub-cube.

So please understand: Pushing up the exposure reduces noise, yes, but it also reduces the colour gamut.

The reason why you mostly can get away with ETTR is that in real life, low-contrast scenes (foggy landscape, for instance) usually don't contain particularly saturated colours ... and all in scenes, be their contrast low, medium, or high, the highlights usually are less colourful than the mid-tones (to our eyes, at least). But in those scenes where the brightest tones also are colourful, ETTR needs to be applied with care—especially when you're looking at a luminance histogram (as opposed to a three-channel raw histogram). That's all I'm saying.

[Guillermo Luijk]

So please understand: Pushing up the exposure reduces noise, yes, but it also reduces the colour gamut.

Why should that happen?. Changing exposure is just a linear scaling of the captured RGB triplet, RGB={400,200,800} overexposed by 2 stops becomes {1600,800,3200}. If no clipping occurs, how can you say exposing more reduces colour gamut? it will simply have more resolution in defining the colours, and less noise, but not less colour definition.