sarangiman said:
so even 25 electrons of read noise would mean you have to pull EXTREME shadows to get any real problems with non-random noise forms given Canon's higher black point (2048 rather than 1024 as it was with the 5D II.)
jrista: Yeah about that -- do you understand the black points of 1024 & 2048 vs. Nikon's 0? I just don't get it. BTW the Canon S100 has a black point of 128. Seems like it's always some power of 2.
Also, while we're on this subject (I know we talked about this before in a different forum, but no conclusion was drawn): I find it highly suspicious that when I use the method of LetTheRightLensIn & 'bronxbombers' (on dpreview's fora), i.e.:
DR = log[base 2] (max pixel value/stdev in black frame)
I get the same numbers as DXOMark... so often it's just uncanny & suspicious. It makes me wonder: is DXOMark using the same methodology, or are they actually measuring DR by also metering the objects in the scene? I get 11.2, 11.6, 14 for the 5D2, S100, D7000, respectively, using the simple equation above & black/white frames from those respective cameras. Which is almost exactly what DXO gets. Don't you find that suspicious?
Well, we have actual black-frame samples from the 5D III now, so I'm more trusting of the results of those tests at this point. My problem with those tests originally is it was exclusively using noise reads from the masked off border pixels of CR2 files. There was no way to know for sure what, exactly, those masked pixels might contain, especially given that the vast majority of the sample images available on the net were from pre-production 5D III samples. That made all the claims...which were stated as though they were undeniable fact...highly speculative at best. With a proper black frame from a production body, the results are certainly more realistic and verifiable, although they are still somewhat speculative (and counter Canon's claims about improved DR...regardless of how much improvement there may actually be.)
DXO does pretty much the same raw-file analysis, however they do it with images taken of a very specific and finely calibrated device that is supposed to ensure that DR, SNR, and ISO measurements are accurate. They sample light at many levels, not just between maximum saturation and standard deviation of black. I'm not sure about the DR calculation you specified though. That looks similar to DXO's SNR formula (which lacks the logarithm). DXO explicitly states their approach to determining DR as follows:
Given the option, I'd take a Canon 12 stop sensor with no noisy electronics over a Canon 14 stop sensor with noisy electronics.
Well, wouldn't the 14 stop sensor have to have relatively lower read noise in order to achieve the higher DR, according to the above equation?
I guess it depends on the characteristics of the noise. Fixed pattern noise is pretty fine in definition and usually forms boxes around less noisy areas (rather than entirely covering the whole image), and it could easily intrude into useful detail, without actually raising the lowest useful black level in every pixel. Thats kind of the problem Canon has had in the recent past...you could still extract
some useful detail from very deep shadows, its just that it was frequently intruded upon by sharp, very unnatural fixed pattern noise. Canon also has "rough" dark current noise that obliterates other detail in even deeper shadows, so between the two, you certainly don't have as much recovery power as with Sony sensors. However you still DO have some recovery power. Hypothetically, it should be entirely possible that dark current noise could be reduced to bare minimums...taking care of the "bitchiness", leaving fixed pattern noise behind. You could probably reach deeper into shadows at that point, but they would still be useless because you have this very unnatural crosshatch intruding into your picture.
From what I've read from Canon about the improvements to their sensor, it sounds like they worked to remove as much fixed pattern noise as they could, but don't appear to have improved dark current noise much. Even 5D III images appear fairly blotchy in deep shadows and very high ISO. However I'd take bitchiness over FPN any day, even if it interfered with the last stop out of 14 stops of DR.
Far better noise characteristics.
sarangiman said:
I know what you're saying about the moon -- it's a huge exposure differential. But in this case b/c I had a fighting chance (b/c of the clouds dimming the moon), I really was so close to nailing it in one shot after raising the shadows.
... images clipped ...
It does look like there was a measurable change in ambient there, the landscape contour in the first shot is much more defined, and not just because of the difference in noise. There is actually a window between "sunset"...the moment the sun passes the horizon, and the end of "astronomical twilight"...the point where no more light from the sun has any effect. During mid summer, that can be as wide a timeperiod as two hours (i.e. 8:30pm sunset, 10:30pm astrotwilight ends.) During the heart of winter, it might be as short as an hour and 15 minutes. Even under the last few minutes of astronomical twilight, the difference can be fairly significant on landscape exposures.
Back a few years ago, when I first started photography, I was hiking around some 14ers here in Colorado. I stayed up above 13,000 feet (where your angle of view to the sun is better than if you were at a lower altitude) loong after sunset, and kept taking photos. Over an hour and a half after sunset, the upper parts of the peaks were still exposing brighter than the rest of the mountains...and a hard shadow edge separated the bright peaks from the rest (only really visible with longer exposures, if you are ever inclined to experiment.) It was kind of like aplenglow, only with the gray light of astronomical twilight rather than the brilliant red of the first/last rays of sunrise/sunset.
So, IMHO, 2 stops extra DR would've really helped here. Do you disagree?
Certainly, more DR is always useful, however how you get that extra DR and where it comes from matters. When it comes to shadow range, its most useful when coupled with very low-noise electronics. Keep in mind, dynamic range in a digital sensor is allocated exponentially, such that it favors the highlights (more a consequence of base 2 math than anything, really.) Additionally, Canon's sensors are technically 14-bit sensors...its just that they are losing the
least significant stops to noise. Gaining two additional
usable stops of DR on the bottom end isn't really going to help the picture. If we break down the 14 bits of information a modern digital sensor has available, you get the following levels per stop (highlights at the top):
Stop 14: 8192
Stop 13: 4096
Stop 12: 2048
Stop 11: 1024
Stop 10: 512
Stop 09: 256
Stop 08: 128
Stop 07: 64
Stop 06: 32
Stop 05: 16
Stop 04: 8
Stop 03: 4
Stop 02: 2
Stop 01: 2
A sensor with very low read noise might consume Stop 1 and interfere with stop 2. Canon cameras probably consume the first couple stops with noise, and interfere with Stop 3 and to a small degree Stop 4. Lets say the moon exposure utilizes Stops 7-14, and the landscape exposure utilizes stops 1-4. Assuming you improve noise characteristics on a Canon such that it stops interfering with Stop 3 and 4...well, you haven't really changed anything. You aren't adding highlight room...so you can't push your exposure farther without blowing out highlights. You are making a few extra levels of shadow information useful...but there really wasn't much headroom there to start with. You may reduce non-random noise characteristics...but those shadows are still going to be noisy due to photon-shot noise. Instead of having maybe 11 levels to work with for the landscape...you now have approximately 14 levels. You can certainly bump the exposure of those levels now without much worry that you'll see unsightly noise patterns emerge...but your not gaining as much as you might think. Your simply
gaining access to what was already there.
Now, the story would be different if you only had 12-bits to start with, and actually increased the sensors capabilities to 14-bits. At that point, your highlight headroom increases by
12288 levels! You can push your exposure a
lot farther, and as a consequence of being able to push the moon exposure to a higher range of levels, you DO gain more room in the shadows (or rather, you don't have to underexpose the shadows as much, and therefor shadow SNR improves.) That would have a far greater impact on DR than simply making the least-significant levels of already-existing bit depth "usable". The same would go for a move from 14-bit to 16-bit sensors and ADC's. You would gain an
additional 49152 levels with those two extra bits (a full three times the grand total you had to work with before)! (Might also explain why 16-bit electronics capable of processing 6-14 high resolution photographic frames per second are currently beyond the reach of
reasonably priced DSLR's now.
)
Obviously, this is a bit of a contrived example, but that is indeed the reality of the DR story with Canon sensors. Were not lacking highlight range...were unable to fully utilize existing shadow range. Exact distribution of levels may not be as "neat" as my example, and there are numerous levels of indirection between the signal on the sensor and the data in a CR2 file. We know that Canon's black point (at least as measured by the masked pixel border) is fairly high...however how that plays into real pixel data in non-masked areas when a CR2 file is properly processed is something I can't speak to. It does seem like Canon is only capable of achieving about 15760 levels in a fully saturated pixel (if previous numbers floating around are indeed accurate), so the exponential breakdown of levels is going to differ, and the bottom two stops are largely (or maybe entirely) inaccessible (regardless of how many levels they may support.)
