5D3 same max dynamic range as the 5D2???

[LetTheRightLensIn]

Numbers are hard to interpret, there are so many factors that don't show in a number. No measurement so far I've seen take noise quality (pattern etc) into account, which has a huge impact on subjective DR.

Even if 5Dmk2 and 3 measure the same, if pattern noise is less than usable DR will be better.

High ISO measurements often miss out on quantum efficiency and can be misleading due to that.

A test that I would like to see is mk2 and mk3 shooting the exact same test scene perfectly ETTR, and then push a shadow area (preferably containing a color checker) 3 stops and show the crops side by side, resolution is almost the same so no sacling is required. I suspect/hope that the mk3 will fair much better in such a test than these measurements indicate.

This is true and there are hints that at high iso the noise of the 5D3 looks nicer than the noise of the 5D2 so the usable improvement might be a little bit more than whatever the measured improvement in SNR turns out to be and for scenes having extensive areas that are very dark there is a chance the 5D3 might visually do noticeably better than the plain measure SNR difference suggests. We don't really have great samples to compare that yet though.

The same might be true at low ISO but it's too hard to tell from the tiny little masking area. Pulling IR shadows you still do seem some vertical streaking though. It's really hard to say what the difference will be. Maybe it won't help at all. Maybe you will get an extra 1/2 stop of usable dynamic range down there or something for many shots even if the measured value is almost the same.

 

[LetTheRightLensIn]

I saw this chart in another thread, and it looks like the D800 is the new DR king for sure.


The 5D3 and D800 are pretty much comparable, but only at ISO800 and above.. Below ISO 800, the D800's DR is heading sky high, setting a new record.

If this chart is true and not a result of some miscalculation, that would indicate that the D800 applies some sort of artificial curve to balance the shadows and highlights. An aggressive HTP sort of thing.

actually the linear plot for DR of that type is the expected if you had a perfect read from the sensor at a certain stage of the readout, so the D800 is not applying any sort of weird curves (nor is the Canon, it curves over, which is NOT the expected, but it is not because of weird tone curves either but because part of the read out noise stage that doesn't come into play much at high ISO is not so hot so it shows up a lot at low ISO and you get diminishing returns compared to what you'd expect looking at the high iso numbers as you go to near base ISO since you get worse read performance at lower ISOs)

 

[LetTheRightLensIn]

The 5D3 and D800 are pretty much comparable, but only at ISO800 and above.. Below ISO 800, the D800's DR is heading sky high, setting a new record.

Bill Claff, whose site this is from, is a long-time fervent Nikon guy; and these graphs are based on estimated values for both the D800 and the 5D Mk III.

Big pinch of salt in order here...

And a bigger pinch of salt for the postings of a long time 100% rah-rah Canon no matter what poster such as yourself .

BTW his charts follow the same pattern of 5D2 vs the other Exmor sensors on DxO, etc. so his results don't seem all that shocking for the 5D3 and D800. But yes, as he says too, they are preliminary estimates. We can hope the final results will change. But if I had to, I'd place money against it. While there is no saying patterns can't break, it's a safer bet that things work as they have in the past rather than suddenly change.

 

[qwerty]

Question for those who may know.

Do these test (the OP's method, those in links or DXOlabs' stuff) include sensor level processing?

The Sony A580 and Nikon D7000 use the same sensor, yet the Nikon is evidently capable of an extra .6 stops of DR. I assume that's due to Nikon's Active D Lighting processing, but maybe it's due to generally better front-side optics.

Any ideas?

http://snapsort.com/compare/Nikon_D7000-vs-Sony_Alpha_DSLR-A580

The A580 has a Pellicle mirror ( http://en.wikipedia.org/wiki/Pellicle_mirror ), so the sensor has less light to work with. That gives slightly lower performance at a given ISO. The Nikon has a standard mirror that flips out of the way before the shutter opens.

Edit: looks like I may have been confusing the 580 with another Sony camera. Nevermind me.

 

[sarangiman]

part of the read out noise stage that doesn't come into play much at high ISO is not so hot so it shows up a lot at low ISO and you get diminishing returns compared to what you'd expect looking at the high iso numbers as you go to near base ISO since you get worse read performance at lower ISOs

LTRLI: Does the amplification based on the ISO setting occur before charge read out (before noise injection due to read)?

 

[DavidRiesenberg]

actually the linear plot for DR of that type is the expected if you had a perfect read from the sensor at a certain stage of the readout, so the D800 is not applying any sort of weird curves (nor is the Canon, it curves over, which is NOT the expected, but it is not because of weird tone curves either but because part of the read out noise stage that doesn't come into play much at high ISO is not so hot so it shows up a lot at low ISO and you get diminishing returns compared to what you'd expect looking at the high iso numbers as you go to near base ISO since you get worse read performance at lower ISOs)

As far as I remember, I never saw a sensor that exhibits a kink in the curve like the D800 one. It doesn't happen with the D7000 or the Phase One (wide DR examples) at DXO. That's why in my technical-limited-knowledge logic it seems to me that Nikon felt that below ISO 800 the noise is low enough to justify some post capture adjustments.

 

[sarangiman]

I think they say records and captures because the read electronics can mess up what the sensor itself captured. I believe it has been shown that the 1D4 sensor, for instance, actually captures close to 14 stops but then the readout loses a few stops and that is what gets recorded in the RAW file and it's only 11.5 stops instead of almost 14.

I'm not sure I'm explaining myself properly. My concern is that by just looking at the RAW file, are you really measuring the number of stops recorded in the actual scene that was being shot? Seems to me that to be able to do that you'd have to have a wide range of subjects with different luminosities (i.e. the multiple ND filter set up that DXO says it uses), measure those luminosities using a reliable spot meter, then look at the RAW data and see the range of luminosities for which the camera is able to:

[list type=decimal]
[*]Give a reasonable SNR for the darkest luminosity recorded
[*]Differentiate one luminosity vs. the next on the higher (brightest) end
[/list]

By just looking at the RAW file, with no actual scene information, the only way the test would be valid would be if there was a directly linear relationship between the scene luminosity & the luminosity recorded in the RAW. Which your statement about the 14 stops going to 11.5 itself seems to show isn't the case.

 

[LetTheRightLensIn]

actually the linear plot for DR of that type is the expected if you had a perfect read from the sensor at a certain stage of the readout, so the D800 is not applying any sort of weird curves (nor is the Canon, it curves over, which is NOT the expected, but it is not because of weird tone curves either but because part of the read out noise stage that doesn't come into play much at high ISO is not so hot so it shows up a lot at low ISO and you get diminishing returns compared to what you'd expect looking at the high iso numbers as you go to near base ISO since you get worse read performance at lower ISOs)

As far as I remember, I never saw a sensor that exhibits a kink in the curve like the D800 one. It doesn't happen with the D7000 or the Phase One (wide DR examples) at DXO. That's why in my technical-limited-knowledge logic it seems to me that Nikon felt that below ISO 800 the noise is low enough to justify some post capture adjustments.

Oh that kink, I was ignoring that, I thought you meant something else. I think the said that he made a small mistake with the ISO400 reading on the camera, that would explain some of it. Or maybe in how the ISO got rated, once you go above base ISO and compare things it's a little trickier because of that. I doubt it's anything in the camera itself.

The 7D does get a bit of a kink in the SNR charts because it has an unusual form of gain, as opposed to offset, vertical banding that can add a little to the SNR and it has more influence at a certain point along the ISO scale, I think.

 

[LetTheRightLensIn]

I think they say records and captures because the read electronics can mess up what the sensor itself captured. I believe it has been shown that the 1D4 sensor, for instance, actually captures close to 14 stops but then the readout loses a few stops and that is what gets recorded in the RAW file and it's only 11.5 stops instead of almost 14.

I'm not sure I'm explaining myself properly. My concern is that by just looking at the RAW file, are you really measuring the number of stops recorded in the actual scene that was being shot? Seems to me that to be able to do that you'd have to have a wide range of subjects with different luminosities (i.e. the multiple ND filter set up that DXO says it uses), measure those luminosities using a reliable spot meter, then look at the RAW data and see the range of luminosities for which the camera is able to:

[list type=decimal]
[*]Give a reasonable SNR for the darkest luminosity recorded
[*]Differentiate one luminosity vs. the next on the higher (brightest) end
[/list]

By just looking at the RAW file, with no actual scene information, the only way the test would be valid would be if there was a directly linear relationship between the scene luminosity & the luminosity recorded in the RAW. Which your statement about the 14 stops going to 11.5 itself seems to show isn't the case.

Unlike film, the digital sensors have linear capture from the scene.
The 14 to 11.5 is because read noise makes the bottom few stops meaningless mess of randomness.

 

[sarangiman]

It means that SNR would be better up in the realm where photon noise dominated read noise. But it does relatively little, in comparison, for the performance in the deep shadows compared to the read noise. So it can be that they have the same SNR in the deepest shadows near black but that one has better SNR say at semi-dark gray and brighter.

Ah, all you're saying is that the absolute difference in SNR will be less for low signals vs. high, as is clear in this simple linear plot here of the SNRs for QE 1 vs QE of 0.5, assuming a read noise of 6e-:

This always begged the question in my mind: if you increase QE dramatically for a sensor without increasing full well capacity, don't you run the risk of blowing out the sensor for any given f-ratio & exposure combination compared to an older sensor with lower QE & same full well capacity? I realize that the increased QE sensor would have a lower gain for any f-ratio/exposure combination, but blown out is blown out... this always confused me.

Thanks.

 

[Loswr]

This always begged the question in my mind: if you increase QE dramatically for a sensor without increasing full well capacity, don't you run the risk of blowing out the sensor for any given f-ratio & exposure combination compared to an older sensor with lower QE & same full well capacity? I realize that the increased QE sensor would have a lower gain for any f-ratio/exposure combination, but blown out is blown out... this always confused me.

Sure. 'Full well capacity' is measured in electrons, after conversion from photons. If QE goes up, you get more electrons per photons input, and the well fills faster. It's probably not a practical problem, though, since in most normal shooting situations a lower gain (ISO) would solve the problem.

 

[sarangiman]

neuroanatomist, or anyone: Though I asked this previously, I think it got lost in all the back-and-forth -- is 'gain' at the pixel-level applied during charge reading? I.e. does the ISO setting determine the 'conversion gain' (µV/e-)? Meaning, if you have a very low signal, is increasing this gain advantageous b/c it will help offset the read noise to be injected & b/c of the limited precision of reading?

 

[sarangiman]

jrista:

Do you mean "charge reading" or "conversion of charge to a digital unit"?

Not sure what happened to your post, but I'm answering anyway... I mean during 'charge reading'. If it's just during 'conversion of charge to digital unit', then LTRLI's statement "part of the read out noise stage that doesn't come into play much at high ISO is not so hot so it shows up a lot at low ISO and you get diminishing returns compared to what you'd expect looking at the high iso numbers as you go to near base ISO since you get worse read performance at lower ISOs" wouldn't make much sense (if I understood what he was trying to say correctly).

I may be totally wrong.

But I'm basing my notion from this comment in the chapter on CMOS Image Sensors in "Image Sensors & Signal Processing for Digital Still Cameras" (assuming these are active, not passive, pixel sensors): "The active pixel concept in which a photogenerated charge is amplified in a pixel and the amplified signal is read out has its roots in the phototransistor array image sensor. One advantage of the active pixel is its suppression of noise generated and/or injected in the signal readout path..." (pg. 145).

Then again, no idea if this 'amplification' is the one you can set by ISO... someone more knowledgeable will have to chime in.

 

[jrista]

Not really sure what happened to my post either. ISO, at least with Canon cameras, is a pretty ugly beast when you really dig into it (its not really as simple as I stated before.) Its not handled the same way at all ISO settings, and there are big changes once you get to really high ISO settings. Its not really actually handled just at one point, it can be handled at multiple points along the read and processing pipeline.

Generally speaking, a specific ISO setting is achieved via per-pixel amplifiers, the ADC, as well as post-ADC digital "gain", which is pretty much the same as adjusting exposure during post processing. In the case of pre-1DX/5DIII sensors, Canon also used a full-stop native/third-stop push\pull approach. Analog gain (which could be via amplifier or amplifier and ADC) ISO only affected full-stop ISO settings: 100, 200, 400, 800, 1600. Beyond ISO 1600, you start seeing both analog and digital gain. When it came to third-stop settings, full-stop ISO gain would be applied, followed by a 1/3rd stop push or pull exposure adjustment. ISO 125 would be ISO 100 with digital "gain" of +1/3rd stop applied. ISO 160 would be ISO 200 with digital "gain" of -1/3rd stop applied. Things get really complex when it comes to higher ISO settings. For example, ISO 5000 would be:

+4 stops "clean" analog gain (pixel amplifier, I believe...never found any specific details)
+1 stops "dirty" analog gain (modified ADC gain, I believe...never found any specific details)
-1/3 stops "nasty" digital "gain" (reduction of EV of post-ADC digital image)
+1 stops "pretend" "gain" (setting in metadata that tells RAW processor that more digital exposure increase is necessary)

So, the initial four stops of clean analog gain by the pixel amplifier would conform to what you've read in your book. Modern CMOS sensors usually include nose reduction circuitry to make sure that the direct pixel read is as clean as possible (usually CDS, or correlated double sampling, to limit dark current noise...used in both Canon and Sony sensors; sometimes additional measures are taken to minimize fixed pattern noise caused by deviations in the performance of each pixels transistors relative to each other, etc.), so that when the pixer amplifier jacks up ISO according to the camera setting, its jacking up an electron count read as close to the real photon conversion count as possible. That would only be the case with full-stop ISO settings up to ISO 1600 on a Canon sensor, however.

No one can say yet how Canon's new sensors handle ISO settings. One of the things I personally really want to know is whether analog gain, either at the pixel or via ADC or something else, is done for every ISO setting. The whole push/pull third-stop approach is rather annoying, as it always chops off/shifts at lest 1/3rd stop DR at those ISO settings. It also causes a very funky progression of noise levels. Full-stop noise is "normal", push stop noise is "excessive" (often worse than full-stop ISO two to three settings higher), and pull stop noise is "clean". ISO settings beyond 1600 (which, in the 1D X and 5D III, might actually be ISO 6400 or ISO 12800 now) pretty much get a double-whammy, with two kinds of analog gain as well as two kinds of digital exposure amplification, usually leading to a large jump in noise levels that really eat away at fine detail. It does sound like the new 1D X and 5D III sensors have incorporated new per-pixel pre-amp hardware noise reduction. Not sure exactly what kind of noise it reduces...but given that there appears to be practically no fixed patter noise in 5D III sample images, and at worst minor vertical banding noise...I'd venture a guess that it eliminates fixed pattern noise.

Clean analog gain (which I believe is best achieved via per-pixel amplifiers after fixed-pattern and dark current noise are mitigated during photodiode read) is how all ISO settings should be achieved in my opinion. I believe that is indeed how Sony sensors work...at least up to ISO 1600 on current camera models (I think they also do additional ISO adjustment in the ADC at higher ISO's.)

 

[KeithR]

As long as the methodology used is the same across all cameras then it is valid to compare cameras with.

Oh, I'm sure they'll be OK, but as I say, some caution is in order.

Must be honest though - I struggle to understand where Nikon/Sony can have got the thick end of two stops from.

 

[LetTheRightLensIn]

gotta get back to the game and then eagles but quickly, based on someone's black frame posts here (another thread):
disappointing results at ISO 100:

banding:
horizontal banding appears to be entirely removed, this is good, and will help make ISO look better and get less badly clumped chroma noise and all but vertical banding appears to be zero improved and it shows up enough that I don't think you gain any usable amount of low ISO DR back compared to the 5D2

so for low ISO I'm not sure the removal of horizontal banding alone will really help usable DR any since the vertical sticks out to the eye just as much, maybe here on there on parts of the image it might help (it ay be enough to help high iso look better though and make a difference)

DR:
5D2: 15760,1024,6.1 AU,3.8 8MP normalized ADU, DR - 11.2 , DR8MP - 11.9

5D3: 15309 (maybe the channels were still not quite blown though?),2048, 6.6 ADU, 6.4 normalized to 5D2 6.4 and then to 8MP 3.96, DR- 11.0 DRto5D2 - 11.0 DR8MP - 11.7
and let us say WP should be 15760 instead then DR8MP - 11.8
and let us even say it should be max 14bit 16383 - 11.8

so I actually get fractionally worse DR than for the 5D2, but it's for all intents, the exact same, within copy to copy variation and well within any difference you'd be able to notice

Sad to say but even the DX D7000 handily beats the latest Canon FF for DR, and not by a little. The D800, at 100% view, full 50% MP advantage may beat it by 2 stops, maybe even 2.5 usable stops.

I did not expect this at all. :'(

I think we just have to accept that it will do nothing for the ISO 100-400 non-action shooters. It's basically a big step up in AF and possibly some degree better image quality at high ISO (hopefully enough to matter, it looks like it might be, but can't be sure yet) only. Not that those are minor things, but still, not sure how that makes it $500 more than the 6D800. I am now really, really surprised they decided to price it above the D800.

 

[Abraxx]

Oh, what I feared, no real DR improvement.
If further tests confirm this, I'm not going to buy the 5D Mark III, especially not at the extreme price point Canon has chosen here in Germany.
DR improvement was crucial for me. I'm really disappointed if this turns out to be true. :'(

 

[sarangiman]

I forget who posted this .CR2 file (I think it was either here or on dpreview), so, sorry I'm missing credit for it, but, vertical banding is pretty poor. This is +100 Fill Light, ISO 100 (it's a crop, not full image):

That just looks like FPN that shoulda been removed in internal camera processing...

That's really going to limit usability of shadows in high dynamic range landscapes (& yes, I do shoot w/ Singh-Ray filters... but sometimes it's either not practical b/c of you're, say, shooting a cityscape without a simple horizon, or when even those filters aren't enough... e.g. when shooting long telephoto where effects of grad ND filters are minimal or, say, when shooting w/ the moon in the frame).

 

[sarangiman]

jrista:

One of the things I personally really want to know is whether analog gain, either at the pixel or via ADC or something else, is done for every ISO setting. The whole push/pull third-stop approach is rather annoying, as it always chops off/shifts at lest 1/3rd stop DR at those ISO settings.

It's kind of stupid that on the 5D2 you can set ISO steps at 1-stop increments, thereby never selecting any of those intermediate ISOs, but Auto ISO will still use the intermediate ISOs. Seems to me like it'd be helpful if Auto ISO also considered your setting of ISO increments. Along this train of thought, I wish Canon/Nikon/etc. would allow us full access to the hardware to modify the firmware ourselves... kind of like writing apps on iOS... would open up a world of possibilities for photographers with special interests. But I doubt we'll ever see that happen.

given that there appears to be practically no fixed patter noise in 5D III sample images, and at worst minor vertical banding noise...I'd venture a guess that it eliminates fixed pattern noise.

I don't know about that... please see my previous post. That particular model of the 5D III has as bad or worse vertical banding than my 5D II...

 

[jrista]

given that there appears to be practically no fixed patter noise in 5D III sample images, and at worst minor vertical banding noise...I'd venture a guess that it eliminates fixed pattern noise.

I don't know about that... please see my previous post. That particular model of the 5D III has as bad or worse vertical banding than my 5D II...

There are slight differences in banding vs. fixed pattern...I believe they are caused by different gains at different parts of the pipeline. The fixed pattern stuff is finer and cross-hatched, and pretty much always the same in every frame. Whatever causes banding is different, and the banding can change frame to frame. As far as I can tell, the fine, FIXED PATTERN noise is gone in the 5D III. There does appear to be some vertical banding, however I've seen inconsistent results. Some samples seem to exhibit it, while others don't seem to exhibit any at all. I have been curious if that was due to differences in pre-production samples or something like that? Anyway, I think banding and fixed pattern noise are separate types, caused by effectively the same thing, just at different parts in the read pipeline...and banding does not exhibit the same in every frame where as fixed-pattern (by definition) would.