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

[neuroanatomist]

Ok, but the sensor is 24*36mm, so whatever light hits the sensor is your image.  The sensor is not round.  Again, correct me if I'm wrong. 

Nope.  As jrista just confirmed, there are intentional 'dead/inactive' pixels at the edges of the sensor, which is why the specs for every Canon dSLR list a certain number of millions of total pixels, and a somewhat smaller value for 'effective pixels' - usually 3.5-5.5% less than the total pixels on the sensor.  The number that is advertised for the camera (e.g. 18 MP for the 7D, 21 MP for the 5DII) are the effective pixels (those cameras have 19 MP and 22 MP total pixels, respectively).

[Daniel Flather]

Thanks neuroanatomist.

[Orion]

I LOVE CANON!

I LOVE PHOTOGRAPHY!

how's that for dynamic range!?

Go out and enjoy your camera . . . whether you are a Nikonian or Canonite! Both make wonderful cameras, and it is a waste of time (with all due respect) to 'worry' about what I consider trivial things. In the end, if trivial is the game, then many Pulitzer Prize winning photographs have no place in this thread or any other. . . .

LONG LIVE ART in purest form.

[KeithR]

All canon sensors have two areas of "black masked pixels" to the left and right edges of the sensor.

Some converters seem to be able to get past the masking, too - 7D conversions in (say) Lightroom are 5184 x 3456, but Raw Therapee conversions of the same files are 5194 x 3457.

[LetTheRightLensIn]

Does your method involved the use of a specialized test device designed just for the purpose of testing DR, that is finely calibrated to produce consistent test images containing tonal swatches of exact known intensities, lit by a very specific kind of light bulb that is capable of emitting a very specific amount of light, with a very specific color balance, at a very specific intensity? Do you take dozens of sample shots of said test device, evaluate each individual shot to produce very accurate measurements into a data sample set? Do you aggregate the measurements of that data sample set to produce a scientifically accurate mean, standard deviation, ratios, etc.? Do you have your results verified by multiple parties for accuracy? Is your final conclusion based on those verified results? Are all of your test cases for every brand of equipment you test subjected to the EXACT SAME TEST, done in the EXACT SAME WAY, to minimize deviation between the results of various brands, models, and even samples of specific equipment? Do you have a reputation as someone who does these things...a track record of verifiable accuracy?

DXO doesn't just take a sample image and perform some little procedure on it to come up with the dynamic range numbers they publish. Claiming that your results are as accurate as theirs comes with a certain responsibility, and I'm sorry...but you repeatedly claiming that your as accurate and reliable as DXO, well its actually rather sad at this point. As it stands, all we really have at this point is your personal word...but someones word only has meaning to those who have reason to trust it.

Wow, did you even read the procedure? Do you even know how DxO measures lowest ISO DR?? How about reading that stuff first before spouting nonsense. It has nothing to do with wedged and special light bulbs and so on. Trying to find out what gain they use for a given ISO or the SNR is trickier but I am not doing that. I, and the others, are simply measuring ISO DR!

And as for millions of samples, obviously not, but all the same, as I said, just using one black frame from my 5D2 and I ended up actually getting THE EXACT SAME numbers for 5D2 ISO 100 DR as DxO did TO THE TENTH OF A STOP (the error bars are probably a bit larger than a tenth of a stop anyway, but by chance, it ended up being exactly the same).

Have you read anything I or any of the others have said?

If you think all half dozen of us are lying as some sort of dirty paranoid conspiracy against you then you have the files, you have the procedure, you have the software to do it, go do it yourself and see what you come up with.

You've done some "fiddling", and you have your personal conclusions. As far as I can tell, this supposed "engineer" your a fan of has also done some "fiddling", and also has his own personal conclusions. Speculation and hearsay, but personal conclusions nevertheless. You've been repeating your conclusions all over this forum for a few days now. Its certainly your prerogative to continue, but personally, I would stop comparing my own little home-grown, ad-hoc DR evaluation method to the precision testing performed by DXO...save yourself a little face.

It is not ad hoc. It's the same procedure people used on all of the past cameras. Heck the same procedure was used when he used to post in Canon forums, and is probably on his website,  of one of the four physicists who came up with Heterotic String theory which helped launch the first super-string revolution. Granted everyone can make mistakes and come up with incorrect procedures and make mistakes, but I haven't see anyone discredit this method and the results it gives seem to match what DxO gets (who BTW are not perfect anyway, their optical tests have some bizarre results: 16-35 2.8 has sharper corners at f/2.8 than at f/5.6; at 200mm f/2.8 they say the 70-200 2.8 IS is sharper than 2.8 non-IS is sharper than 2.8 IS II; 70-300 non-L is sharper than the L at 300mm wide open, etc.) and as far as it seems it sounds like DxO is using the exact same procedure (although at higher ISO they plot the points at effective ISO and they do a series of measurements to come up with that, but that is irrelevant here).

[neuroanatomist]

Ok, let's go back to the beginning.

Assuming the side masking area can be used for the measurement then:

I contend that the above assumption is invalid, and it's the assumption on which the conclusion is based.  Do you have data to support the assumption that the side masking area's signal value and standard deviation are representative of the portions of the sensor that are actually used to record the image?  I've got an open mind about this, but I must confess, I get really suspicious when someone 'measures' numbers in a system where a log₂ scale is commonly used, and the resulting values are exact integer results of 2ⁿ. 

EDIT:  You've got a 5DII, right?  So, can you shoot a black frame (lens cap on) and compare the ISIS data from the masked edge with the center of the frame?

[LetTheRightLensIn]

We are all spending hours and hours on numbers.
Very sad. Very sad.

Some time ago I saw an image taken from a MF and I was blown away.
I checked dxo and the DR was worse than the Sony Nex 10.

There are two things :
Numbers and what your eyes see.

I am going out now to take some pictures.
Or should I stay home to do some more calculations.

Please guys use the camera for where it was made for.

Nice article here:(eyes vs. numbers)

http://www.luminous-landscape.com/essays/eyes-vs-numbers.shtml

Just for the record some of the pixel-peepers in the forums actual do go out and shoot too. It's not an either or question. 

And sure you can take amazing shots with this or that, no question at all so you don't need this or that other thing. But there may be types of amazing shots you can't take with cam at all or not with a high take rate that you might on a different one.

[LetTheRightLensIn]

I listed my entire procedure in detail! I wasted like ten minutes typing it out. It's the same thing DxO does.

Really?  It doesn't sound like you did this:

We perform noise measurement using a transmission target placed on top of a uniform light box. The transmission target, designed by DxO Labs, is made of a plate of thick black plastic with precision-drilled holes which are equipped with a range of neutral density filters designed to absorb the light identically for all wavelengths.

The filters are made of pure optical glass with no structures that can be measured as noise. (While other image quality measurement solutions make use of printed targets, we believe such targets are inappropriate for noise measurement testing, as the intrinsic noise pattern of the print paper may be recorded by the tested camera and then confused with the camera’s own noise pattern.)

We place high-density filters on neighboring positions to limit reciprocal illumination of the patches.
The light box (placed behind the target) is composed of two fluorescent daylight spectrum tubes with a diffusing sheet on top, achieving a perfect uniformity on each filter. The luminance is about 1500cd/m2.

We use filters having different light absorption levels ranging from 0% to 99.99% in order to test across a dynamic range of 4 density steps (= 13.3 f-stops — a dynamic range much greater than today’s digital cameras). When shooting such a chart, the sensor of the camera being tested sees a wide range of light levels, with a 1/10,000 ratio from minimum to maximum. For comparison, a printed target dynamic is typically 2 density steps (6.65 f-stops), which is inadequate to simulate high dynamic range or back-lit scenes.

Each uniform zone on the chart (a “patch”) is measured for luminance (cd/m2) with a certified luminance-meter; then all the values are input into DxO Analyzer software.

Once we measure the target and calibrate the DxO Analyzer software, the selected camera shoots an image of the noise target at different ISO settings, and we measure the noise for each color channel of the target image (R, Gr, Gb, B). We compute the mean gray level and noise values for each patch and for all images shot at different ISO settings. We then interpolate these numerical values for all gray levels to calculate and plot signal-to-noise ratio (SNR) curves, from which DxO Analyzer extracts the SNR 18%, the dynamic range, and the tonal range.

Perhaps the 10 minutes you spent typing out your 'procedure' wasn't the only time you've wasted on this 'analysis'...

As I said I was not measuring SNR, not measuring integrated tonal response, not measuring color fidelity, not measuring for actual ISO ratings, etc. They are measuring tons more stuff, lots of more complicated stuff. I was just measuring maximum DR at base ISO which is vastly simpler than doing all of that other stuff. It is possible that their software extracts the dark current patch SNR ration and the blown channel and does actually what I and others have done. And I mean I did end up with the exact same ISO 100 results for the 5D2 DR.

I just hope that the main sensor area turns out to not act like the masked area this time. That would be awesome. I'd love to be wrong. But also note that all of the talk from Canon about the 5D3 has carefully left out low iso, they always mention improved for mid and high iso.

And while this next post below is a bit speculative, his info is a bit down the line and speculative but how about this quote from someone not from Canon but with ties to them (I don't know who he is but he has been right about everything Canon the last few years on larger scale issues, he certainly might not have all the itty bitty tech stuff correct, etc.), quoting his post on another forum:

"I told you so, the 5DmkIII sensor is mainly a "polished" version of the former mkII sensor. Not that this is not a real improvement, but the step forward in real image quality is in fact marginal if you shoot RAW mainly. The 1Dx however has a sensor with a complete redesign. Expect it to have some real native characteristics that are above the 5DmkIII. .... The 1Dx performs some better in almost every department..... ISO, DR.... except for a few mp. Still the 5DmkIII..... is mature enough to completely satisfy the needs of enthusiast amateurs, semi pro's and professionals who prefer the compact body size over the 1Dx's additional qualities."

Off case take that with a grain of salt, but it sounds like, as some on DPR speculated, that maybe (again this is very much speculation now and may be prove to be entirely wrong) they may have decided to use the old fab for the 5D3 to save money again and only used the new fab for the 1DX sensor.

And I think he may be playing down the high iso improvements of the 5D3 a little bit. And maybe the 1DX sensor won't even be that different and it was just marketing spiel.

Anyway this guy has hardly been a Canon basher, in fact, he had long been a very 100% rah-rah cheerleader on everything they ever did.

[LetTheRightLensIn]

(Having an understanding of the CR2 format, sounds a bit sketchy to me as the masked pixels are intended to be used to set black level by RAW converters, and they are purposely fixed at around 1024, rather than around 0 (don't ask me why Canon does that...its just what they do)...I don't believe they are a valid basis to use for measurement of DR.)

I hope that will turn out to be the case this time, that the masking area is not a valid basis this time, and that all my measurements will not match what happens when you measure the main sensor area. But it has proven to be a valid basis for use in the past. I don't think it was tried for all off the different bodies, but it seemed to give identical or quite close results when it had been tried. I was initially a little skeptical of it myself but the results were same ballpark completely.

But yes, might it be different this time? It is always possible. Judging by the past it is probably grasping at straws, but maybe this time.

I'm starting to think if they did improve low ISO DR it might only be with the 1DX this time. I hope the 1DX at least has it improved since then we know 5D4 and so on will get it and that Canon is back on track. But it would be pretty annoying to 5D3 users, if they used the cheap old fab and yet still charged $500 more the Nikon (produced on a new fab) and saved the new fab for the 1DX only.

Taking the 1D APS-H series as an example using DXO data, the II had 11.1 stops, the III had 11.7 stops, and the IV had 12 stops. Read noise for those cameras was 29.7, 22.3, 16.6 e-, respectively. Canon has shown a notable ability to improve their low ISO dynamic range with reductions in read noise, and they have made some significant claims about improved DR that could bite them in the ass if it turns out to be false, or worse, the 5D III exhibited worse DR than the 5D II.

OTOH they have yet to have any low ISO DR improvement since the 1Ds3. The 5D2 was no better nor was the 1D4.

I expected the 5D3 and 1DX to be a lot better, but it seems quite likely that the 1DX is the last hope now.

I have no idea what kind of lighting was used to illuminate any sample image I might get my hands on, whether the camera was a pre-production model or not, whether the exposure settings were set exactly correct to measure DR for the type and intensity of the bulb used to illuminate the scene.

You don't use any lighting at all for a dark current reading. And for a blown channels max raw levels it doesn't matter the lighting either, if every channel is blown out it is blown out.

Again we are not talking trickier stuff like measuring middle tone SNR or what the ISOs really are or tonal response or CFA array response and so on, that stuff is trickier but note that I am not mentioning having measure that sort of stuff either. I even said I await the DxO results for that.

[LetTheRightLensIn]

...something to say that will help us evaluate this method of DR data gathering.

Still not sure I'd call it data gathering.  As for consistent results, the method guarantees them.  As for results that match proper empirical tests, a broken analog clock is correct twice a day. But whatever.

Here's a question: when 'measuring' the dark noise level in the 'side masked area', what is the probability that the numbers provided for the dark signal are meaningful?  He reports average values of 1024 for the 5DII and 2048 for the 5DIII - exact 2ⁿ values.  Personally, I find that suspicious, and it seems more likely that those values are not actual signal, but rathet result from the camera firmware setting those pixels, which are outside the image area, to an arbitrary value as it writes out the .CR2 file.  In that case, both the absolute value of the dark signal and the noise of that signal are not valid for DR determination.

take black frame with you 5D2 and measure the blackpoint and it is right around 1024 in the center of the imaging area and in the masking area, above ISO1600 I started getting slightly higher values 1025 and then eventual up to 1034 by ISO 25,600.  They were not all to the decimal point, but whether it is 1023.8 for 1024.1 for a patch doesn't matter.
 
You can find the black level of 1024 for the 5D2 (most canon cameras have used 1024 or 2048) listed on tons of different sites or ask one of the raw processing software guys in one of the forums what level he set his raw converter to use for the 5D2.

And it's not like this MY procedure anyway. Lots of people have used it and was hardly the first.

[LetTheRightLensIn]

Here's a question: when 'measuring' the dark noise level in the 'side masked area', what is the probability that the numbers provided for the dark signal are meaningful?  He reports average values of 1024 for the 5DII and 2048 for the 5DIII - exact 2ⁿ values.  Personally, I find that suspicious, and it seems more likely that those values are not actual signal, but rathet result from the camera firmware setting those pixels, which are outside the image area, to an arbitrary value as it writes out the .CR2 file.  In that case, both the absolute value of the dark signal and the noise of that signal are not valid for DR determination.

Good point. Do you reckon the darkest meaningful parts of an image (blown-out black) would have a significantly lower value?
Also, if(?) you can take the blown-out highlights to be the same value every time (i.e. blown-out white for a specific sensor will always be the same value for any picture with that sensor in RAW files), is it not strange that he actually got the same DR number as DxO predicted (for the Mk2)?

the backpoint has stayed about the same across all ISOs for canon cameras (sometimes rising a bit towards the upper end for whatever reason)

the max saturation value stored differs by ISO setting
for instance to list just a few I found on the 5D2 at ISO:
100 - 15760
125 - 15761
160 - 12810
800 - 15761
1000 - 15762
1250 - 12812
12800 - 16383

it tends to be the same for base iso and 1/3 up and then be lower for 2/3 up stops and for very high isos it can go to a uniform really high value like say 16383

the 50D had the same pattern

the 40D was a little different since base iso, +1/3, +2/3 all had a different storage

[neuroanatomist]

You can find the black level of 1024 for the 5D2

Based on the calculation you indicated, the black level is not nearly as relevant as the noise...  Is the noise the same in the center as the masked area?

[jrista]

@LetTheRightLensIn: Perhaps we are not referring to the same sources of information and data. I can no longer tell, based on statements about low ISO DR having not improved from 5D II to 1D IV (which as far as I can tell from DXO data, it did. 5D II screen DR 11.16, 1D IV screen DR 11.46; 5D II print DR 11.86, 1D IV print DR 11.95. Given how bad low ISO read noise has been on Canon sensors, its not surprising they have trouble getting more than 12 stops.) First off, having re-read your previous posts, no...you did not say you were measuring SNR. Given that you are solely measuring DR, I'll certainly grant that is a much simpler test. But thats never really been my complaint.

Regarding the 5D II, I believe you can get more accurate results, because you can actually get your hands on one and create your own test images. Regarding the 5D III, unless I've missed something (posts regarding this topic are spread across several threads now), the best we have to go with are sample photos posted to the net (the best of which seem to be from IR at this point.) We can't be sure one way or another whether the darkest pixel of an IR sample shot is actually indicative of cap-on dark current read noise. At best, we can make assumptions. It seems you came to that same conclusion, which is why you used the black masked pixels in the CR2 file as a basis for minimum read level. I am pretty sure I have some bookmarks at home (and if I get the chance here at work, I'll see if I can search for them sooner) that have useful details about what those pixels are, what data they contain, and how to interpret that data. To put it simply, those pixels are specialty cases either for setting black point and/or calibration, and cannot reliably be used as a basis for minimum signal (there are some statistics about deviation around the base 1024 value they tend to have that you might be able to use to deduce a true dark frame minimum signal level...but even so, any conclusions based on such a measurement would not be particularly reliable as I'm not sure anyone fully understands those pixels.)

Given the lack of a sufficient set of test images, I don't think we (the internet community at large exploring DR, SNR, banding noise, etc.) can make enough conclusions to be concerned or relieved one way or another. I think doing so, and claiming that the tests and the results are as reliable as DXO's, is asking just a bit too much (and putting a certain responsibility on yourself that you probably don't want in the first place.) (Hence my rather rambling approach before to asking: Do you really believe you are being scientific in your measurements?) To put it more succinctly: Without known-good test images that we are certain contain patches of nothing but read noise in pixels other than the black masked pixel columns to the left and right of the sensor...we can't really come to reliable conclusions one way or another...about the 5D Mark III. Additionally, this is in contrast to conclusions about tests performed with the 5D Mark II, which we can be reasonably confident of given that we can actually get our hands on actual 5D II cameras in order to generate viable test images that contain the right data.

Apologies for being obscure before, I was not making my point clear. I'm not necessarily doubting the validity of your test method in its context (although I do think use of masked pixels is not reliable.) I'm doubting the validity of the test samples themselves as a means to calculate DR. As soon as someone gets their hands on an actual production (vs. pre-production) 5D III and can actually generate a cap-on dark frame and a cap-off channels blown frame, I'll be far more willing to listen to the conclusions as "reliable" and "similar to DXO".

[LetTheRightLensIn]

So... what are you doing? Blowing the highlights and measuring, blowing the shadows and measuring, and computing the magnitude between them? If you repeat the same calcs for every 5D3 sample do you get the same thing? If you repeat for every 5D2 image do you get the same 11.2? Or are you just measuring the DR of a single image?

I'm not doing anything. I'm using IR's files. Thankfully they blew the highlights on some specular highlights so that is where I got the raw saturation levels from. The dark current noise I measured from the masked area of the file that was cut off from light. it seems to be around +/- .1 stops for across three quick tries on files, doing the same thing my 5D2 values happen, by chance to match DxO exactly to the tenth. different copies might vary +/-.15 or so perhaps unless you got a real weird copy

Setting aside the log function to represent it in stops, I'm just trying to understand the calculation.

Correct me if I'm wrong: masked area = some part of the sensor that is physically blocked from light but still records brightness values. By definition, this would be the darkest part of any exposure.

So when you say you're measuring noise, are you reading random values from that black area (which in theory should be 0) and determining the minimum level at which noise no longer occurs?

Then a similar procedure is done in a blown out area?

Then the magnitude between dark and light is DR?

measuring a block in the masked off area where no light hits and the avg value is the black point, ideally every single value be exactly say 1024 or 2048 but there is noise from the read out electronics so the values bounce above and below the black point, i record the std dev, the read noise in ADU.

For the blown out channels, i put a big box around IR's image where there were bright specular highlights and recorded the highest value reported for the selected block, it is possible that they are not really fully 100% blown out, but since more than one highlight had the exact same peak value it seems likely they were, even if one channel was still a bit under or something and heck let us just assign the max 14bit value that can even be recorded it doesn't even change the results much, instead of 11.1 you get 11.23, so bid deal .13 stop difference, 1/8th stop, meaningless in any practical sense and if you chose the max 5D2 value then difference is even less than that, maybe 1/16th of a stop to what I reported.

If there is something to be had it would be in the dark current. We can just hope that they are doing something very different with the masking area this time.

[t.linn]

And who on Earth needs (emphasis intentional) that many stops of shadow recovery anyway?

Folks can argue back and forth about the validity of the tests and the sample size of one.  I get that.  This sentiment I don't get. 

As a landscape shooter I routinely expose as far to the right as I can and still encounter images where I need to draw considerable detail out of the darkness.  It isn't the least bit uncommon to encounter scenes where the dynamic range easily exceeds the capabilities of my 5D2's sensor.  This is, after all, why many photogs embrace HDR. 

If Sony can figure out a way to reduce the need to rely on HDR by providing greater dynamic range within a single capture, I expect Canon to be just as competent.  It's not like they don't have the resources to compete.