Why is everyone calling the mk3 sensor the same as mk2?

[LetTheRightLensIn]

You know what is really silly?

Not believing your own eyes.

this is the most accurate summary IMO and why i give DXO a bit fat care factor of 0

Keep in mind he switched it to screen mode from print mode for some reason and that many users have independently measured ISO100 DR and gotten the same results. And that if you compare photos even using DPP and other maker's own custom software the DR differences in the numbers seem to be pretty apparent.

Also keep in mind that when Canon had universally the better sensors almost every Nikon owner I talked to readily admitted it (even if pointing to their often better body specs at the same time). Why do like 50% of Canon users, seemingly, have to make stuff up? Canon has 6fps vs 4fps in FF (vs 5fps in 1.2x and 6fps in 1.6x), 1-2 stops better SNR in video, video without moire, top AF, but it doesn't have better or even close to as good DR at lower ISOs and maybe if half of Canon users didn't pretend it did Canon might have paid attention and maybe the 5D3 wouldn't have been so far behind.

 

[LetTheRightLensIn]

That's because you are looking at the (print) DR figure, which is where DxOMark really cooks it up.

It's called normalization. Would you compare the times of two cars one to go 0-60 miles per hour and one to go 0-60 kilometers per hour directly without conversion factor?

You can't directly compare 22.3MP against 6MP as if the noise was all at the same power scale.
Or maybe you thought the D700 had much better SNR than the 5D2?

 

[TTMartin]

That's because you are looking at the (print) DR figure, which is where DxOMark really cooks it up.

It's called normalization. Would you compare the times of two cars one to go 0-60 miles per hour and one to go 0-60 kilometers per hour directly without conversion factor?

You can't directly compare 22.3MP against 6MP as if the noise was all at the same power scale.
Or maybe you thought the D700 had much better SNR than the 5D2?

DxOMark claims to be testing the sensor, yet they come up with Print results?

The reason I changed to screen is Dynamic Range doesn't change with the number of megapixels. Normalizing for Dynamic Range simply hides the fact that DxOMark can't fully decode the CR2 files, by making it look like their Dynamic Range reading is changing when it isn't.

A more accurate analogy would be comparing the 0-60mph times of a two seat car and a four seat car, would I 'normalize the results' with some conversion factor because the four seat car can carry twice as many people to 60mph in the same time? You and DxOMark might, but, I wouldn’t.

Whether the D700 or 5D Mk II is better is an extraneous argument that has nothing to do with the fact that DxOMark cannot fully decode the CR2 file.

You keep talking about all the tests that have shown that the DxOMark dynamic range scores are accurate. The only tests I’ve seen that ‘prove’ DxOMark focus on the shadow end and show that using a 3rd party RAW decoder you can pull more shadow detail out of D800 or other Sony sensored cameras. Shadow recover is shadow recovery, and only addresses half of the Dynamic Range of the camera, which includes highlight retention too. I’ve seen no tests that compare the dynamic range of a 5D Mk III using DPP as the RAW converter. The closest thing is DPReview’s Dynamic Range Tests which use the cameras JPG engine for RAW conversion. Those tests show that without any electronic enhancements on that the 5D Mk III has more Dynamic Range than the D800, and with D800s ADL on that it has about a 1/3 of a stop more Dynamic Range than the 5D Mk III with HTP on. This is certainly not the 2.7 stop advantage indicated by DxOMark (D800 – DxOMark Maximum Dynamic Range (screen) 13.23 vs .5D Mk III – DxOMark Maximum Dynamic Range (screen) 10.97)

 

[Aglet]

The reason I changed to screen is Dynamic Range doesn't change with the number of megapixels.

I agree, this gives us the best, and IMO, proper way to test low level sensor performance. It is the closest to the original data
What happens after normalizing to some standard size print is a different result, not necessarily meaningless or irrelevant, but certainly of no use when comparing just sensor performance, an electronics and processing issue.

Normalizing for Dynamic Range simply hides the fact that DxOMark can't fully decode the CR2 files, by making it look like their Dynamic Range reading is changing when it isn't...
Whether the D700 or 5D Mk II is better is an extraneous argument that has nothing to do with the fact that DxOMark cannot fully decode the CR2 file..

I still don't understand why and how you conclude 3rd party raw converters are missing some ability to properly decode Canon's CR2 files. Please explain

The big difference I've found between them is more to do with de-Bayering and Noise Reduction. There, DxO and Lightroom used to do a slightly better job compared to DPP from my experience. Latest DPP actually seems to provide more image (color) detail than DxO but DPP still lags both in NR performance.

Also, the CR2 file produced by a Canon G11 or G12, which, AFAIK, is a SONY sensor equipped camera, exhibits a DR curve more like a Nikon than a Canon, complete with excellent base ISO DR about equal to any current Canon DSLR. Pretty good for tiny pixels!

The only tests I’ve seen that ‘prove’ DxOMark focus on the shadow end and show that using a 3rd party RAW decoder you can pull more shadow detail out of D800 or other Sony sensored cameras. Shadow recover is shadow recovery, and only addresses half of the Dynamic Range of the camera, which includes highlight retention too.

I somewhat agree, highlite retention ability is part of the equation.
However, total DR is a ratio of the highest recordable EV (hilite) to the lowest recordable EV (deep shadow). Since the lowest recordable EV depends on the Signal to Noise Ratio being set at some arbitrary, but useful limit, below which noise obscures image data, this is the more important end of the DR range. SNR at hilites is very high so not a factor unless you're evaluating IQ of very small pixel sensors, which we're not.

These deep shadows are where Canon's sensors have trouble with read noise intruding on their signal. It's an electronics design issue. They chose a particular compromise which works very well except it loses at the shadow end.

I’ve seen no tests that compare the dynamic range of a 5D Mk III using DPP as the RAW converter.

I see the same deep shadow pattern noise in my Canon raw files no matter what raw converter I use, and I normally use DPP for basic processing.

The next difference I find is that Canon's DPP doesn't provide the same tone-curve controls Lightroom or even DxO have, so it's not as simple to bring up the shadow areas in DPP to see the noise - but it's still there. DPP seems to process the shadow end a little darker than the other converters, which minimizes the appearance of the noise.
When I process for printing I like to bring the shadow areas up a bit more to retain some shadow detail in the final print, and that's when the patterned and banded noise structures of Canon's raw files show up and sometimes cause problems. The more you have to bring them up (sunlit landscapes = more DR) the more likely the problem with the shadow areas.

E.G. I recently used my 5D2 with a 580EX II flash for fill to do some outdoor family portraits, mix of sun and cloud day. 3 of 4 people in the group wore pants of varying shades of dark gray.

In DPP, I added +1 EV to correct for overall exposure on the subjects because a strongly backlit sky threw off my basic exposure more than I'd planned. (I don't normally shoot outdoor portraits in rapidly changing sky conditions) I still had to blow out the sky to get good exposure on the subjects.

That's only ONE stop of push, and even in DPP there's clearly banding noise visible on 2 of the darker pants, the noise exhibiting both vertical and horizontal banding and plenty of chroma noise. This isn't even processing the dark levels up to where I'd normally prefer them yet.

I export those files to 16-bit TIFF to continue to work in Photoshop to do the usual touchups.
When I try to lift the dark tones of those pants slightly more I cannot, the banding noise becomes objectionable. So that's a tiff file, exported by DPP, which is the de-facto decoder of Canon's raw file, still exhibiting strong noise in the shadows with only +1 EV of push.

Does this matter?
Well, i can make a shot like this work for a family portrait that the customer finds acceptable at the print sizes they want.
If I were to print it 30" wide then the pattern noise on the pants would be something I would see.

Compare this to sunlit landscape shots I recently took with my D800 where I want to retain ALL the scene detail, from textures in bright white clouds to textural detail in the charcoal of a burned tree, in the shade. I manually expose using the old Sunny 16 thumbrule. 1/500s @ f/8 & ISO 100. Considering the D800's ISO is about a half stop less than what it says, I could have exposed at least another 1/3 EV and still retained highlite details.
The resulting image from shooting such a scene has the remainder of the landscape very dark and the charred area pretty much black.
In Lightroom or Photoshop (ACR) I can then lift those shadow areas NUMEROUS EV, to the point where I can now see detail in the shaded area of the charcoal! And guess what?... NO noise. No banding of any kind, not even a spec of noise and that's with the usual 25/100 default chroma NR setting set to ZERO.
I didn't have to recover lost hilite detail because I already exposed to retain it near the camera's maximum.

This means this camera produces FAR cleaner shadows than a Canon. So I'm not surprised it's DR is rated a few stops better using the existing testing methodology. There's no way in Hades I could process an image like this if I shot it with my 5D2!
Ergo, I have to agree with DxO that the D800 provides considerably more real world usable dynamic range than a 5D Mark II.

Anyone want to BUY a super clean, barely used 5D2?... I really don't know why I'm keeping it any longer.

 

[Aglet]

LTRLI - I recall when you were examining early raw files from the 5D3, you had some sort of software that allowed you to view individual pixel data levels numerically?..

Any way some stats could be run on that kind of data if one were to make test shots of a white sheet or gray card at various exposures and the variance/mean could be analyzed for different tonal values and yet another way of determining total dynamic range with shadow values set to some level of SNR ?

 

[LetTheRightLensIn]

The reason I changed to screen is Dynamic Range doesn't change with the number of megapixels.

I agree, this gives us the best, and IMO, proper way to test low level sensor performance. It is the closest to the original data
What happens after normalizing to some standard size print is a different result, not necessarily meaningless or irrelevant, but certainly of no use when comparing just sensor performance, an electronics and processing issue.

They are not literally making prints and comparing, they just decided to call it print results instead of normalized results since by print results they mean image shown at same scale.

The only tests I’ve seen that ‘prove’ DxOMark focus on the shadow end and show that using a 3rd party RAW decoder you can pull more shadow detail out of D800 or other Sony sensored cameras. Shadow recover is shadow recovery, and only addresses half of the Dynamic Range of the camera, which includes highlight retention too.

I somewhat agree, highlite retention ability is part of the equation.
However, total DR is a ratio of the highest recordable EV (hilite) to the lowest recordable EV (deep shadow). Since the lowest recordable EV depends on the Signal to Noise Ratio being set at some arbitrary, but useful limit, below which noise obscures image data, this is the more important end of the DR range. SNR at hilites is very high so not a factor unless you're evaluating IQ of very small pixel sensors, which we're not.

There really is no such thing as highlight recovery with a standard digital sensor. It's a hard cut-off and they have linear capture. You simply find the saturation point and then you measure the black point noise.

You can change what part along the linear sequence you declare to be middle gray and apply various tone curves but none of that comes into play measure the max engineering DR the sensor can deliver.

These deep shadows are where Canon's sensors have trouble with read noise intruding on their signal. It's an electronics design issue. They chose a particular compromise which works very well except it loses at the shadow end.

It is. The Canon sensors themselves have a LOT more dynamic range captured than the result that gets collected after all the reading and converting of the sensor. I forget the details but I think the 5D2 sensor itself grabs well over 13 stops at the screen level, closer to 14 while the RAW file, after all the electronics damage the signal, has like only 11 stops left at the screen level.

I’ve seen no tests that compare the dynamic range of a 5D Mk III using DPP as the RAW converter.

I see the same deep shadow pattern noise in my Canon raw files no matter what raw converter I use, and I normally use DPP for basic processing.

This means this camera produces FAR cleaner shadows than a Canon. So I'm not surprised it's DR is rated a few stops better using the existing testing methodology. There's no way in Hades I could process an image like this if I shot it with my 5D2!
Ergo, I have to agree with DxO that the D800 provides considerably more real world usable dynamic range than a 5D Mark II.

exactly

 

[LetTheRightLensIn]

LTRLI - I recall when you were examining early raw files from the 5D3, you had some sort of software that allowed you to view individual pixel data levels numerically?..

Any way some stats could be run on that kind of data if one were to make test shots of a white sheet or gray card at various exposures and the variance/mean could be analyzed for different tonal values and yet another way of determining total dynamic range with shadow values set to some level of SNR ?

Yeah but it's awfully tricky if you want exact measurements out of that since you need a super stable light and power source and high-end equipment to measure output or some high end device to regulate the light and a detailed table of how much it outputs per various fed currents I think. I don't think many people have the setup to be able to do that very well.

Maybe if you had stable power in your area (I sure don't, at least not in summer) or used batteries to run the light you could use a monitor calibration probe to measure the light. I don't know.

I think it's tricky and time consuming and I'd rather go out and take pics.

EDIT: Actually I guess you could just use a high quality step wedge on a light table, that would be a lot easier and solve many of the problems anyway I don't own one. They probably don't cost too too much but who knows it maybe they do charge like $100-200 for a quality stpe wedge. Anyway I don't feel like wasting time with it.

It might be interesting to do a rougher visual test at some dark but not black level and see at what point banding starts making the engineering DR look unusable. That might be the real deal. But in the end I don't feel like bothering with the time that would take and just simply note that with some bodies banding affects usable measure DR more (5D2,5D3,7D,1D4) than with others (40D, 1Ds3, D800, etc.) whether by 1/4 or 3/4 stop or 1/3 or 1 n each case I don't know. And that some cams like 5D3/1DX have banding affecting usable DR less than the 7D less than 5D2/50D/5D/etc. when talking high ISO.

 

[Wilmark]

Because its basically almost indistinguishable.

 

[Aglet]

Yeah but it's awfully tricky if you want exact measurements out of that since you need a super stable light and power source and high-end equipment to measure output or some high end device to regulate the light and a detailed table of how much it outputs per various fed currents I think. I don't think many people have the setup to be able to do that very well.

i was thinking of single shot analysis, step wedge would be good but not necessary, just some varying reflectance targets from black to a few EV above.

The kind of analysis I'm thinking of would be to analyze the noise (variance) and try to determine the banding effect that way rather than averaging it or however DxO does it. in fact, I'd be most interested in Peak-to-peak variance at those various levels because that's what becomes the visible distraction.

not that this is even necessary!
We can see from a typical shot that the raw files contains this to an annoying degree.
I don't need to measure HOW annoying it really is. ;D

.. I don't feel like bothering with the time that would take..

+3!

I barely have time to go shoot for fun, let alone monkey in the studio testing stuff beyond what I've already done to decide I needed to get a better body for some of my work.

 

[Aglet]

There really is no such thing as highlight recovery with a standard digital sensor. It's a hard cut-off and they have linear capture. You simply find the saturation point and then you measure the black point noise.

I suspect RGB channels will saturate at different levels tho, so that's about the only "recovery" available but then color accuracy is lost.
Some cameras tone curves vary a fair bit as they approach cutoff. Canon seems to take it a bit more gently than Nikon from what little I've checked but I'm careful not to clip any color channel when taking a shot and that can be really tricky with some subjects.

It is. The Canon sensors themselves have a LOT more dynamic range captured than the result that gets collected after all the reading and converting of the sensor. I forget the details but I think the 5D2 sensor itself grabs well over 13 stops at the screen level, closer to 14 while the RAW file, after all the electronics damage the signal, has like only 11 stops left at the screen level.

I'd be very interested to know how Canon's sensors would perform if they used the same type of on-sensor processing as SoNi-kon's.