Expose to right, it is as simple as

[jrsforums]

One can use the dynamic range vs ISO plot on DXO to determine this point of diminishing returns. When the curve becomes linear, increasing ISO will have minimal effect on read noise. As shown on the plot below, this occurs at about a measured ISO of 700 for the Nikon D3 (the ISO indicated by the camera will be higher with this camera). With this camera there is really no point in increasing the ISO above 800, since you will only decrease highlight head room. Expose as much as possible at a camera ISO of 800 and use the raw converter to increase exposure.

With the D7000, the curve is linear from base ISO and there is no real need to increase the ISO to obtain a "better" histogram. One can expose as much as conditions allow and make up the difference in the raw converter. The appearance of the histogram is irrelevant. The scientific explanation for these considerations is given by Emil Martinec. The article was written before the D7000 was available.

Bill,

It is not clear to me if you are saying that the D7000 or the D3 is the better sensor.

My initial thought would be that I would rather be able to increase ISO, with less increase in noise....even if it were not the entire range of ISOs available.

Thanks, John

[Anders_HK]

However, if the area giving the individual channel ETTR exposure is a saturated colour, the exposure is likely to be very different from that obtained using a spot-meter.

A spot meter will measure any individual RGB value or combination thereof as plain tonal values. Thus we need to be aware also of zones 0, I or II (around saturation point) for the individual RGB colors.

[ErikKaffehr]

Hi,

The D7000 has better dynamic range at minimum ISO. If it is the best sensor depends on your needs. For high ISO work the D3 is probably better. If you need good shadow detail at low ISO the D7000 would be a better choice.

The major differences is that the D3 sensor is bigger, so it collects more photons. It may also have better quantum efficiency or less sharp color grid array. On the other hand the Nikon D7000 has a better signal processing path. So D7000 picks up less noise in the shadows. Noise in midtone will be mostly affected by "shot noise" which is linked with the number of photons captured, putting the D3 to advantage.

The D3 needs 50% less enlargement than the D7000 for a given print size, which means that it has lower demands on the lens. Another way of saying this is that it will transfer more contrast on fine detail than the smaller sensor camera.

The best camera is probably the D3X, which has very similar characteristics to the D7000 but having a larger and more high resolving sensor. For high ISO shooting the D3S may be better.

I include "Density Range" and "Tonality Range" plots from DxO for the D7000, D3 and D3S. Density range is essentially looking at shadow noise, while "Tonality Range" is more like how smooth midtones will be.

Discussing the "best camera" needs to relate to requirements but also to cost. What do you need and what you are willing to pay? The situation may be different if you acquire a new set of tools or if you are adding to an existing one? Let's assume that your favorite is a 24-70/2.8. It would have a field of view of 36-105 lens on a D7000, perhaps not you would regard to be optimal. On the other hand, Nikons 16-85/3.5-5.6 may be a decent standard zoom for the D7000, but not the lens for low light shooting.

Best regards
Erik

Bill,

It is not clear to me if you are saying that the D7000 or the D3 is the better sensor.

My initial thought would be that I would rather be able to increase ISO, with less increase in noise....even if it were not the entire range of ISOs available.

Thanks, John

[ErikKaffehr]

Hi,

I see some more complexity. The raw image will contain three color signals with luminance probably corresponding to the spot meter. Color balance is applied in post processing and the question is also how that processing is done? The raw signal will probably contain at most 14-bits except those MFDBs having 16 bit signals. Raw conversion is probably done in 16 bit. If we assume that raw conversion would be done in low bits we would have a headroom of 2 bits (that is two stops) for handling color balance without clipping.

Best regards
Erik

A spot meter will measure any individual RGB value or combination thereof as plain tonal values. Thus we need to be aware also of zones 0, I or II (around saturation point) for the individual RGB colors.

[Anders_HK]

I see some more complexity. The raw image will contain three color signals with luminance probably corresponding to the spot meter. Color balance is applied in post processing and the question is also how that processing is done? The raw signal will probably contain at most 14-bits except those MFDBs having 16 bit signals. Raw conversion is probably done in 16 bit. If we assume that raw conversion would be done in low bits we would have a headroom of 2 bits (that is two stops) for handling color balance without clipping.

The data we can retrieve (find useful) in post and which corresponds to our choice of zone 0, I or II as "bright point tool" is important (before and/or after recovery is our own preference). That does not limit to doing so mere for a neutral bright point, does it?

[ErikKaffehr]

Hi,

The data are scaled when doing white balance. So data that is unclipped in the raw file can be clipped when doing white balance.

Best regards
Erik

The data we can retrieve (find useful) in post and which corresponds to our choice of zone 0, I or II as "bright point tool" is important (before and/or after recovery is our own preference). That does not limit to doing so mere for a neutral bright point, does it?

[Stefan.Steib]

After having read through the article and now through this thread (which was interesting, otherwise I wouldn´have spent that much time!) some thoughts come to my mind:

1.: Usability of the exposure concept
it is agreed that using all space on the histogram- with a preferrence to "expose to the right" is a valuable and useful method to achive optimum SNR.
The problem is only that scene lighting changes, sometimes rapidly. If an automatic process is not able to prevent  full clipping (full loss of information - whereas using the histogram fully is also not "linear" so there are parts of the  available steps lost when doing this which can be overall usefull as dicussed here ) then I think it is more useful to compromise and give it a certain amount of reserve.

2.: Performance of differing Chip Concepts in comparison
As we have heard the channels do not react the same, further it may push the Dynamic range to a limit using it fully. How is a concept like Foveon reacting on these finetunings -and -
wouldn´t it be better if a chip was made which sacrifices the luther condition but instead collects SNR by all available tricks for the channels and then postprocess (either by firmware or in the Rawconverter), something like Fuji made with the 2 different sizes of pixel onchip to improve dynamic range or maybe using other color schemes (also Fuji Hexagon "EXR") or maybe going a step further (and more crazy costwise) - why not using a 3 chip concept as it is in Video cameras but for Photography purposes (using differing sensitivities for the channels or differing exposure times-or even better using differently sensible silicon for each channel ?) ? Wouldn´t this eliminate the need of about any of these discussions and maybe there is a way to use this in future cameras ( Imagine a cheap and small sensor highspeed 3Chip video camera using superexposure and future incamera processing power to resolve high resolution in variable resolutions together with variable sharpness like the Lytro camera....?!)

I´d be interested to hear how the forum people would sort these thoughts to this thread.

TIA and greetings from Munich

Stefan  Steib - HCam.de

[NikoJorj]

1.: Usability of the exposure concept

For me, that's exactly Michael's point : for now, we must rely on Ruth-Goldbergesque contraptions to be able to achieve ETTR (and the number of threads about UniWB, and the number of people willing to stand the horrendous green previews it gives, could be a reasonable proof that this is not exactly a small niche need), but a simple firmware upgrade (at least on cameras with live view) could probably give us a much more accurate automated exposure, for those times when we don't have time to fiddle with trial-and-error and still need an optimal capture to maximise DR.

2.: Performance of differing Chip Concepts in comparison

Ach, diese Bayerischen, die sind immer nach alles komplizieren und... Ooops, sorry, I think I have eaten one too many Prussians at breakfast today.  
For me, what works, works ; and for my need of wide dynamic range (the classical landscape need of bright cloudy sky and darker land beneath) a Bayer sensor the thing that works, ie the better compromise. It just would be even better with an ETTR expoosure mode and a raw histogram.
If you got some spare time, the strength and weaknesses of Foveon sensors are abundantly exposed in this thread (and ETTR and raw histograms would also be very useful for those, because they have less total DR and still give WB-corrected and tone-corrected jpeg previews - and histograms AFAIK).

[Guillermo Luijk]

It always amazes / annoys / amuses me when I publish an article that has been painstaking peer-reviewed by some of the brightest minds in the industry – people who design sensors and write raw software (in this instances) and then "experts" whose credentials are unknown tell me (us) why the information in the article is wrong.

Michael, those brightest minds in the industry came to tell you what you could have read from your own forum's "experts" more than two years ago. The questions exposed in your article were already discussed here:

Camera manufacturers PLEASE: when RAW histograms and an ETTR mode?.

The no-reason for not having an automated ETTR mode in a digital camera, a user setting indicating % of allowed clipped pixels, RAW-based histograms and clipping warnings,...

PS: 2^12=4.096, not 4.098

[bjanes]

Hi,

The data are scaled when doing white balance. So data that is unclipped in the raw file can be clipped when doing white balance.

Best regards
Erik

Quite true, but one can handle clipping with white balance by setting the white balance multipliers all less than or equal to one. Guillermo Luijk explains this topic quite well in his DCRaw tutorial--see the section on white balance.

As en example, here are actual camera histograms taken from the Nikon D3 of a saturated yellow flower. The camera was set to Adobe RGB, the widest space available on the camera. Shot 10 was at the exposure indicated by the meter, but I saw clipping in the red and green channels, so I reduced exposure until the green no longer clipped (shot 12). Note that the luminance histogram (black and white) shows no clipping. See below for an explanation.



The Rawnalize histogram of image 10 shows that the red and green channels are just at clipping, but the red channel will be strongly clipped when white balance is applied (the red white balance multiplier is about 1.5).



In shot 12, the red and green channel is about 0.3 EV below clipping and the green channel about 0.5 EV below clipping. The camera histogram clipping is likely due to saturation clipping in AdobeRGB, the color space of the preview. Unfortunately, ProPhotoRGB is not available with this camera (or any other cameras I know about).



ACR will not allow setting the WB multipliers all <= 1, but one can reduce the exposure to achieve this effect and use ProPhotoRGB to reduce saturation clipping. I allowed slight clipping of the red channel and increased brightness to give a better appearance.



Finally, one should note that luminance histograms displayed by the camera black and white histogram may not show clipping when it is present in the color channels. This is because the luminance histogram looks at all pixels but does not keep track of their location in the image. This is well demonstrated in the Cambridge in Color tutorial on histograms.

Here is the Photoshop luminance histogram and the color histograms of the same image. Note that the luminance histogram shows no clipping, but clipping is present in the red channel. One can get a better estimation of the color channels by using UniWB, but saturation clipping can still occur in the AdobeRGB space used for the camera color histogram. I don't usually use UniWB do keep it in one of the data banks of my camera, so I can use it when needed. Modern cameras are good enough that slight underexposure (short of ETTR) can be tolerated. Remember that SNR increases as the square root of exposure, so doubling of the exposure increases the SNR only by a factor of 1.4. Raw histograms would be nice, but one can still get good results with the available tools if you know how to use them.



Regards,

Bill

[stamper]

I had the chance to try out what was the main thrust of the essay. For a high contrast scene I use the digital zone method which I will keep using but on a lower contrast scene which imo falls within the dynamic range of a scene I used centre weight  and raised the EV on some scenes by 1.0 and 1.3 above what I would normally shoot. Imported them into ACR and I had the flashing red overlay on a lot of highlights which would normally horrify me. Lowered the exposure slider which was about -1 and then processed in my normal manner and liked what I saw. In one particular scene I had auto bracketed it on the camera using + 1.0, two shots. I processed both. One didn't have the flashing red and the other had to be lowered by -1 exposure. Processed both and used the side by side method in Photoshop to compare them and they looked very similar. The +1 had lighter dark areas to begin with which was the object of the exercise. The overexposed areas which I feared burnt out initially weren't once the exposure was lowered. I will definitely try again on a suitable day and so far my verdict is .... worth trying.

[Anders_HK]

Hi,

The data are scaled when doing white balance. So data that is unclipped in the raw file can be clipped when doing white balance.

Best regards
Erik

Erik,

You make a valuable point. However if WB does not change by much, can we assume there will not be much influence?

On another note, in above I misstaken referred bright point to around zone 0, I, II , instead should be VIII, IX, X, sorry about that.

Regards
Anders

[ErikKaffehr]

Anders,

I don't know! What happens is that the channels are shifted quite a lot in postprocessing . To achieve correct white balance, multipliers are applied to each channel. Which multipliers depend on color temperature. These things can be checked out using DxO data and perhaps using "raw-analyzer". I have not really done that, because I presume that it is easy to draw false conclusions.

Much may also depend on raw processor. They may have different strategies, algorithms and priorities.

Best regards
Erik

Erik,

You make a valuable point. However if WB does not change by much, can we assume there will not be much influence?

On another note, in above I misstaken referred bright point to around zone 0, I, II , instead should be VIII, IX, X, sorry about that.

Regards
Anders

[marvpelkey]

I have read through this thread (although my eyes began to glaze over with some of the technical stuff  ) and apologize if I missed this point. There is a bit of a discussion on another site about the use of ETTR for a jpeg image (especially in light of Michael's comment under the "Caveat" heading). Can anyone advise if ETTR is raw-specific or if there is any benefit at all to ETTR for a jpeg file?

Thanks,

Marv

[Guillermo Luijk]

Can anyone advise if ETTR is raw-specific or if there is any benefit at all to ETTR for a jpeg file?

ETTR means post processing (exposure correction)
JPEG is a finished output format, never intended for post processing
So conceptually JPEG-ETTRing is crazy.

[fdisilvestro]

ETTR works based on the linear behavior of digital sensors and corresponding RAW files

Jpegs are far from linear (gamma encoded, color space encoded, white balanced, to name a few)

The best you could do with Jpegs is to aim for precise exposure and correct white balance, not needing post processing, just as Guillermo said.

[Peter_DL]

As en example, here are actual camera histograms taken from the Nikon D3 of a saturated yellow flower. The camera was set to Adobe RGB, the widest space available on the camera. Shot 10 was at the exposure indicated by the meter, but I saw clipping in the red and green channels, so I reduced exposure until the green no longer clipped (shot 12). Note that the luminance histogram (black and white) shows no clipping. See below for an explanation.



The Rawnalize histogram of image 10 shows that the red and green channels are just at clipping, but the red channel will be strongly clipped when white balance is applied (the red white balance multiplier is about 1.5).



...
ACR will not allow setting the WB multipliers all <= 1, but one can reduce the exposure to achieve this effect and use ProPhotoRGB to reduce saturation clipping. I allowed slight clipping of the red channel and increased brightness to give a better appearance.



Finally, one should note that luminance histograms displayed by the camera black and white histogram may not show clipping when it is present in the color channels. This is because the luminance histogram looks at all pixels but does not keep track of their location in the image. This is well demonstrated in the Cambridge in Color tutorial on histograms.

... One can get a better estimation of the color channels by using UniWB, but saturation clipping can still occur in the AdobeRGB space used for the camera color histogram. I don't usually use UniWB do keep it in one of the data banks of my camera, so I can use it when needed. Modern cameras are good enough that slight underexposure (short of ETTR) can be tolerated. Remember that SNR increases as the square root of exposure, so doubling of the exposure increases the SNR only by a factor of 1.4.
Raw histograms would be nice, but one can still get good results with the available tools if you know how to use them.

Interesting example, Bill.

So the reduction of exposure was not necessary
while the initial shot # 10 was already quite well exposed (to the right) – if I get you right.

First, the single R/G/B camera histograms may suggest a bunch of clipping due single channel "amplifications" resulting from in-camera processing i.e. white balance and saturation clipping, but then all of this is "averaged" with the luminance histogram - which finally correlates with the results from Rawnalize (no relevant clipping).

Peter

--

[bjanes]

Interesting example, Bill.

So the reduction of exposure was not necessary
while the initial shot # 10 was already quite well exposed (to the right) – if I get you right.

First, the single R/G/B camera histograms may suggest a bunch of clipping due single channel "amplifications" resulting from in-camera processing i.e. white balance and saturation clipping, but then all of this is "averaged" with the luminance histogram - which finally correlates with the results from Rawnalize (no relevant clipping).

Yes, that is how I interpret the situation. The clipping of the color histograms occurred with white balance. UniWB would give a more accurate RGB histogram, but the color space would still be AdobeRGB with this camera.

Regards,

Bill

[ejmartin]

Quite true, but one can handle clipping with white balance by setting the white balance multipliers all less than or equal to one. Guillermo Luijk explains this topic quite well in his DCRaw tutorial--see the section on white balance.

This method will not always work.  Suppose the WB multiplier on the R,G,B channels are 2.0,1.0,1.0.  We could normalize so that they are 1.0,0.5,0.5, but now suppose there is a clipped green patch that should have been at 1.5 times the clipping point of the raw data; it will register as 0.5 times the clipping point after the normalization when it should have been 0.75.

[Bart_van_der_Wolf]

This method will not always work.  Suppose the WB multiplier on the R,G,B channels are 2.0,1.0,1.0.  We could normalize so that they are 1.0,0.5,0.5, but now suppose there is a clipped green patch that should have been at 1.5 times the clipping point of the raw data; it will register as 0.5 times the clipping point after the normalization when it should have been 0.75.

That is the challenge for realistic HL recovery routines, but your example involves a clipped (Green) channel, not a proper ETTR situation. A proper ETTR image can be White Balanced without loss of highlight accuracy by using a proper scaling of all 3 channels in linear gamma space.

Besides allowing the user a manual control over the assumed true level of the clipped pixels, one could employ various heuristics for making an educated guess. Part of the heuristics could be an assumption of common differences in WB scaling factor in other parts of the image that have slightly lower average luminosity level. Areas with presumed (e.g. with an assumed D50 illuminant) neutral color should have higher weighting. Another assumption could be used for clipped specular highlights. Specular highlights often also have a high spatial frequency and could be treated as a special case of mostly reflecting the illuminant's color and a bit of their own color. A high pass filter combined with a high luminosity threshold could be used to select such areas.

Cheers,
Bart