“Expose to the right” vs “JPEG Histogram to the right”

[BJL]

Considering we can pick between sRGB which doesn't have a gamma curve but a TRC and Adobe RGB (1998) which has a gamma curve, one could assume (and I hate to do so), there would be small differences. Another reason why the JPEG Histogram, a lie, isn't all that useful:

Everything you thought you wanted to know about Histograms
Another exhaustive 40 minute video ...

Maybe I'll find the time for 40 minutes of exhaustion later!  For now my goal is to find a JPEG histogram that lies "consistently". If I know that the JPEG histogram on a particular camera at particular settings overstates exposure at the right end by around 2/3 stop, I can exceed its exposure recommendations by a bit less — say 1/3 stop — and bracket a bit where possible.

Or maybe I just go by the histogram when enough bracketing is possible: "Trust, but verify".

[digitaldog]

Bracket, YES!

[Ray]

Those are some rather old cameras, form before Canon adopted column-parallel ADC, so here are some more up-to0-date comparisons. I use the "print" graphs, rescaled to compensate for different pixel counts. (This is even after I momentarily put aside DXO's outright error in the horizontal scale (through its confounding the upper limit of raw files with Exposure Index), and the roughly counteracting adjustment needed to discount measures of highlight headroom and instead assess shadow handling.)

Wow! I didn't realize that Canon had improved it's DR so much in its latest 5D Mk IV, it's now only about 1.25 EV behind the latest Nikon FX, at base ISO, and only 1/2 a stop below the D850 at ISO 400. Well done, Canon. 
However, I don't think I'll be buying a 5D Mk IV, despite still having a number of Canon lenses sitting on the shelf, because the pixel count of the D850 is about 50% greater, which improves the resolution of all lenses used.

There is still the dubious relevance of the bottom few stops of that "engineering" dynamic range, going down to a photographically useless SNR=1. One way to dig our some information about SNR in the deep shadows is to look at DXO's "SNR 18%" graphs.

Not at all. SNR at 18% is approximately the level of the average skin tone in a face. It doesn't tell you anything about the shadows. Most cameras of the same format size have very similar 'SNR at 18%' values. MF digital cameras tend to have slightly better 'SNR at 18%' than FF 35mm, and FF 35mm formats have better SNR than APS-C or cropped formats, whether Nikon or Canon.

However, if one compares cameras of significantly different ages, the progression of technology can result in a smaller format having as good an SNR as a larger format. For example, the cropped format Nikon D7200 has 'SNR at 18%' values at least as good, and slightly better at base ISO, than the FF Canon 5D. Of course, the DR of the smaller format D7200 is massively better than the Canon 5D, almost 3 & 1/2 stops better at base ISO.

The 'SNR at 18%' values of the Nikon D850, at base ISO, are a significant 6.5dB higher than the Olympus E-M1 MkII.

[kirkt]

...
However, I don't think I'll be buying a 5D Mk IV, despite still having a number of Canon lenses sitting on the shelf, because the pixel count of the D850 is about 50% greater, which improves the resolution of all lenses used.

..

The resolving power of your lenses is fixed by their optics - compared to lower res sensors, the D850 sensor will provide a better chance of exposing the limits of those optics and the flaws in focus technique.

kirk

[DP]

The 'SNR at 18%' values of the Nikon D850, at base ISO, are a significant 6.5dB higher than the Olympus E-M1 MkII.

SNR @ 18% is where shot noise already dominates other noises sources (those affecting deep shadows), so there sensors of similar generation & technology differs just by a die size... FF shall be 2 stops better than m43 being twice bigger (when equalized to the same mp and comparing more or less similar tech generations - which D850 and E-M1mkII are, both being Sony Semi tech)

[DP]

For now my goal is to find a JPEG histogram that lies "consistently".

the answer to that question is known for a long time... you need to use UniWB, then only you have either consistent lie or better yet with some additional OOC JPG parameters you actually can nail clipping in raw and with zebra and/or blinkies see where it happens with <= 1/3 EV precision (which is good enough - because quite a lot of cameras might have ill effects near clipping in raw, so pushing further might not be worth the trouble).

[BJL]

Not at all. SNR at 18% is approximately the level of the average skin tone in a face. It doesn't tell you anything about the shadows.

You have completely overlooked the EI shift part of my comparison. Yes, the "SNR18" report for EI=200 is what you get at midtones when following the camera's metering set at that EI of 200. If you instead want to judge SNR in parts of the scene that are, say, four stops darker, you need to measure with photosites receiving four stops less light. This is the illumination received in the mid-tones at EI four stops higher, so you can get an approximate reading of that by looking at the SNR18 for that higher EI: 3200 in this example. It is not an exact match, because although the photosites are receiving the right amount of light for this test, that "EI=3200 midtone signal" is then being amplified by a greater amount than the  "EI=200 -4 from midtone signal", affecting the noise introduced after amplification—which in modern sensors is mostly from the ADC process itself.

The 'SNR at 18%' values of the Nikon D850, at base ISO, are a significant 6.5dB higher than the Olympus E-M1 MkII.

I am not sure what that comparison has to do with either the subjects of this thread (comparing "highlight based exposure strategies) or of my post (how to assess photographically relevant DR and SNR levels) — but it is about what is expected if the noise levels are dominated by photon shot noise and the sensor's well capacity is about the same per unit area, which would predict 6dB ("one stop") for a quadrupling of sensor area. Half of that 0.5db (1/12 stop) gap goes away once one corrects for the fact that EI indices used in DXO's comparison differ between the cameras and the EI for each camera differs from its ISO SSat speed ("base ISO speed") by a different amount. So then the measurements are within about 1/24 stop of a simplistic prediction based on photon shot noise and sensor area alone.

P. S. I just noticed that DP already said the same thing far more succinctly: "... photon shot noise ... die size ..."

[Ray]

The resolving power of your lenses is fixed by their optics - compared to lower res sensors, the D850 sensor will provide a better chance of exposing the limits of those optics and the flaws in focus technique.

kirk

I think it would be clearer and more truthful to say that the resolving power of a lens is limited by its optics. Do you see the difference?

The resolution of a photograph, using the best techniques such as a tripod, or sufficiently fast shutter speed when hand-held, and accurate focusing of course, is always a combination of sensor resolution and lens resolution.

If you increase the resolution of either one, you increase the final resolution in the photographic image. I tested this for myself years ago when Canon introduced the 15mp 50D, which has 50% more pixels than its previous model, the 10mp 40D.

Using the same Canon prime lens with both cameras, I took many shots of a test chart, at various apertures, using a tripod and manual focus. Whilst I can't remember the precise details without digging into my old records, I recall in general terms that the resolution differences in the centre of the image were approximately equal to, and sometimes greater than, a change in aperture by one F/stop.

In other words, the 50D at F8 was at least as sharp as the 40D at the lens' sharpest aperture of F/5.6, and the 50D at F11 would be at least as sharp as the 40D at F8, and so on.

Now, it's also true that certain weaknesses in a lens will become more apparent through comparison with the higher resolution in the center. For example, if a lens has rather poor resolution at the edges and corners, a higher-pixel-count sensor will not improve that resolution at the edges as much as it will improve the resolution in the center, so the edges only appear worse in comparison with the significantly better resolution at the center.

However, if resolution at the edges is an important consideration in a particular shot, the higher-resolving sensor allows you to stop down to improve edge performance, whilst still maintaining the same, or even slightly worse, center resolution that the lower-resolving sensor provides at the sharper aperture.

If you don't believe me, check out the lens tests at DXOMark. There are numerous examples of the same lens being tested on different camera bodies. The body with the higher pixel count always results in the same lens having a higher score. Here's one example in the attached image, comparing the performance of the Nikkor 14-24mm on the Nikon D700 and D810. The jump in performance is huge, but so is the jump in pixel-count, of the D810. 

[Ray]

You have completely overlooked the EI shift part of my comparison.

No I haven't. Did you not read my previous comments about Canon's DR being equal to Nikon's DR at higher ISOs? Nikon's DR advantage tends to be greatest at base ISO and 2 or 3 of stops higher, say to ISO 800. Beyond ISO 800 DR tends to be very similar. This is something useful to know.

The point I am making is that one cannot get any reliable information on the quality of the shadows by looking at the graph of 'SNR at 18%' values. Period.

To demonstrate this fact, I've attached an image comparing the SNR and DR values of the FF Canon 5D and the 'cropped format' Nikon 7200. Amazingly, SNR is about the same for both cameras, across the entire ISO range, despite the difference in format size.

However, it would be a huge mistake for anyone to deduce from the SNR graph that the noise in the shadows, from both cameras, is approximately the same. The difference in DR is so huge, it would have to slap you on the face. 

Perhaps we should start a new thread, maybe in the Coffee Corner, titled, "The Denial of the results of scientific camera tests."
I find it amusing that so many posters, especially at Dpreview, have claimed the DXO tests are fraudulent, or inaccurate, or a  misrepresentation of the situation. But such posters never show their own valid comparisons to debunk the DXO results. The best they can do is sometimes show out-of-camera jpeg images, which only demonstrate that Camera A has better 'in-camera' jpeg processing than Camera B.

[DP]

SNR @ 18% is where shot noise already dominates other noises sources (those affecting deep shadows), so there sensors of similar generation & technology differs just by a die size... FF shall be 2 stops better than m43 being twice bigger (when equalized to the same mp and comparing more or less similar tech generations - which D850 and E-M1mkII are, both being Sony Semi tech)

being 4 times bigger of course, was typing too fast ...

[DP]

6dB ("one stop") for a quadrupling of sensor area.

except for this matter a "stop" here will be 3dB = you get twice better SNR (6dB) by getting 2 stops more exposure = either by 4 times bigger sensor area (let us not get into vignetting, angle of light to sensels away from the center, etc) or 4 times longer exposure time or 2 stops wider aperture or combination of ...

3dB = 20log(1/sqrt(2)) = 1 stop more exposure (light collected by bigger die size,longer exposure time, wider aperture or combo)

[DP]

The point I am making is that one cannot get any reliable information on the quality of the shadows by looking at the graph of 'SNR at 18%' values. Period.

true

[BJL]

The point I am making is that one cannot get any reliable information on the quality of the shadows by looking at the graph of 'SNR at 18%' values.

I partially agree, but let me put it this way: both the "SNR 18%" and the "DR graphs give limited partial information about usable DR, in the sense of what subject brightness range can be handled with adequately high SNR in the darkest relevant parts of the image. Some of the limitations are:
- SNR 18% assesses only shot noise and "upstream" noise introduced by the camera before the variable amplification for ISO speed adjustment is introduced, so n particular omits quantization noise from the ADC process.
- DXO's DR overestimates by going down to SNR=1 instead of stopping at a more appropriate level. The number of stops to subtract to get a "photographically useful DR" are not known to be the same for different cameras (or even different ISO setting on the same cameras) so it is not jus a matter of "a higher DXO DR score means more useful DR by roughly the same increment".

Aside: the DXO graphs suggest that
(a) the D850 has an unusually low true SSat (aka "base ISO speed": 44, vs 70 for the Sony A7R III and 87 for the Olympus OM-D E-M1 Mk II), and
(b) this is due to lower QE rather than deeper wells.
This may be a design decision weighed more to color accuracy (with narrower pass bands for the CFA) than low light performance, compared to other cameras.
To see the comparison go to DXO comparison and select Measurements -> SNR 18% -> Screen, for  a per pixel comparison. The surprising thing is that the D850 and EM1MkII have the same SNR curve, indicating equal photo-electron counts at equal ISO speed, despite the larger D850 photosites.

Or else something is weird with the way that DXO measures for those SNR graphs.

[BJL]

except for this matter a "stop" here will be 3dB = you get twice better SNR (6dB) by getting 2 stops more exposure = either by 4 times bigger sensor area (let us not get into vignetting, angle of light to sensels away from the center, etc) or 4 times longer exposure time or 2 stops wider aperture or combination of ...

3dB = 20log(1/sqrt(2)) = 1 stop more exposure (light collected by bigger die size,longer exposure time, wider aperture or combo)

I was using "one stop" loosely, to mean a doubling, so I think we are saying same thing: "one stop more light per photosite" meaning a doubling of light received by the photosite (whether by doubling photosite area or doubling exposure duration or increasing aperture by one stop) will increase SNR by sqrt(2), which is 3dB in the scale used by DXO, so it takes two stops more exposure to double the SNR.

[David Eckels]

Not to change the subject but I have read the earlier parts of this thread with new understanding and went about trying some experiments to where I cross the overexposure line with my D850. At ISO 64 with a test image image including a bright sky, I was able to overexpose by two stops and still had some headroom to fully recover highlights in Lightroom. This seems to be corroborated in RawDigger (Fig 1), where I see for the NEF version 27.9% avg G, 5.9% B, and 0% R O/E; while the Max stat values for G/B channels are both around 15.8K, R channel only comes to 11.2K. I realize this could vary per image. Upon converting to DNG during import (zeroed), the RawDigger histograms look exactly the same to me (Fig 2), the O/E stats are a little different only in the G channel average (27.4%). OK, I am not too concerned at this point and my results seem to be close to those discussed earlier by others.

However, when I compare the NEF and DNG histograms in FastViewer, I get slightly different results for each (Fig3&4, respectively).

My questions: Practically, this maybe makes no difference, or does it? I have read that Adobe makes some hidden compensation adjustments, but again the histograms are from zeroed files and shouldn't the two programs from the same vendor yield the same result? I didn't go above +2EV, but perhaps there is even a bit more overhead room? Also, if on average, I am shooting 2 stops underexposed, that means I am throwing away 75% of my data, if I understand the discussion!!!

Andrew, went through your histogram video and 40 minutes seemed like 10!!! Lot's of great info on the Libraw site. Thanks for the discussion folks!

[DP]

shouldn't the two programs from the same vendor yield the same result?

FRV does rough calculations by design - RawDigger does precise ... when in doubt believe RawDigger

PS: plus FRV has an option "Apply Adobe hidden exposure correction" - check if it is set or not

[Ray]

The surprising thing is that the D850 and EM1MkII have the same SNR curve, indicating equal photo-electron counts at equal ISO speed, despite the larger D850 photosites.

Yes. That is surprising and is also good news. If Olympus were to get a full frame sensor manufactured, using the same quality and size of pixels as the E-M1 Mk II, it would be approximately an 80mp sensor (depending on aspect ratio), with better 'SNR at 18%' than the D850. That's technological progress.

Unfortunately, the DR of the smaller E-M1 pixels is significantly worse, as shown in attached image of the DXO graph.
An example of the practical implications is, if the D850 image were cropped to 20mp with a 4:3 aspect ratio, skin tones or similar levels would we just as good as a full frame shot with the E-M1 MkII at its base ISO, but the DR of the cropped D850 shot would be about one stop better, resulting in better shadow detail and lower noise in the shadows. Agreed?

[Guillermo Luijk]

SNR @ 18% is where shot noise already dominates other noises sources (those affecting deep shadows), so there sensors of similar generation & technology differs just by a die size... FF shall be 2 stops better than m43 being twice bigger (when equalized to the same mp and comparing more or less similar tech generations - which D850 and E-M1mkII are, both being Sony Semi tech)

Wrong. For the same exposure, the FF sensor collects 4 times as much light as the M4/3 sensor, but the SNR advantage (for equal technologies) is NOT 2 stops but just 1 stop, and the advantage FF vs APS is not 1+1/3 stops but just 2/3 stops. I.e. the improvement of collecting 4 times as much light over four times the surface (1 stop), is lower than the improvement of collecting 4 times as much light over the same surface with +2EV extra exposure (2 stops).

This is about noise statistics, just compare the Canon 7D II vs Canon 5DS (I take this example since it's the only one in the market where 5DS pixels are exactly the same technology and size as 7D II pixels):

SNR18%: 4dB gap (2/3 stops):


DR 2/3 stops gap:



Since most M4/3 and APS cameras usually have denser sensors than FF cameras, i.e. smaller photocells than in the previous 5DS vs 7D II comparison, those gaps can be slightly increased for technological and QE reasons, but the mathematical starting point of reference is: 1 stop noise advantage FF vs M4/3 (not 2 stops), and 2/3 stops noise advantage FF vs APS (not 1+1/3 stops).

To understand the reasons for this just look at how 4 pixels bin into one: noise is added in quadrature, so SNR is doubled (1 stop of advantage), not multiplied by 4 (which would if there were 2 stops of noise improvement):

http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p3.html#pixelsize


That is why sensor technology is far more important than sensor size for noise performance. The following infography shows the crop on a A7R III sensor that would provide the same DR as the entire sensor of the most popular FF cameras. Look at the tiny sensor (0,9Mpx) with A7R III technology that would provide the same DR as the whole Canon 6D II sensor:




Thanks to this statistical fact smartphone cameras are achieving 'unexpected' great performance with minuscule sensors.

Regards

[DP]

Wrong. For the same exposure, the FF sensor collects 4 times as much light as the M4/3 sensor, but the SNR advantage (for equal technologies) is NOT 2 stops but just 1 stop

you missed my next post in that thread with the math... let me help you with the quote from above

you get twice better SNR (6dB) by getting 2 stops more exposure = either by 4 times bigger sensor area (let us not get into vignetting, angle of light to sensels away from the center, etc) or 4 times longer exposure time or 2 stops wider aperture or combination of ...

at no point in time above I claimed that 4 times bigger/longer/wider means 4 times more SNR-wise...

[BJL]

Yes. That is surprising and is also good news. If Olympus were to get a full frame sensor manufactured, using the same quality and size of pixels as the E-M1 Mk II, it would be approximately an 80mp sensor (depending on aspect ratio), with better 'SNR at 18%' than the D850. That's technological progress.

It might also be good news now for the D850's target users: trading away a bit of a modern sensor's excellent low light performance for improved color accuracy could make many "landscape/enough light" users very happy.

Unfortunately, the DR of the smaller E-M1 pixels is significantly worse, as shown in attached image of the DXO graph. ... the DR of the cropped D850 shot would be about one stop better

Of course! Have we not all agreed long ago that when a sensor four times larger (or a crop to equal pixel count with photosites twice as large) can be given full exposure (not constrained to substantial "photosite underexposure" by the need for enough DOF and high enough shutter speed limits), it will give better image quality? Larger formats are certainly some times for some uses worth their greater cost and bulk! (True even when comparing 36x24mm to 44x33mm to 54x40mm.)

But Ray, you should know by now that you will not impress me with DXO's "total engineering dynamic range" graphs, with an unknown number of a camera's lowest stops having uselessly low SNR, and that number of "junk stops" of DR unlikely to be the same for all cameras. Are there any good up-to-date photographic dynamic range graphs for recent cameras, from Bill Claff or Jim Kasson or other sources?