Since the first test images were shown in september, I have made some attempts to measure the system gain, or in other words the photon collection ability of this camera. I never got consistent results. Now I discovered som unusual properties. It seems that the system gain not only is different for the different channels, but also changes with the light. For instance, the blue channel shows a higher system gain indoors on a white sheet, than outdoors taking the clear blue sky.
This is very preliminary, and it may well be something else I'm seeing. However, if this really is the case, the noise and dynamic range will not be static properties, but vary with the type of light, differently than other cameras.
If someone wants to chime in, maybe we can work together. I don't know how much effort I'm going to put into this alone.
Pixelogy is amuzing
To all photographers: Sorry for hijacking your forum with this quasi-science. I have learned alot about photography from reading here, and feel obliged to give something back to the community when i can.
alundeb wrote:
Since the first test images were shown in september, I have made some attempts to measure the system gain, or in other words the photon collection ability of this camera. I never got consistent results. Now I discovered som unusual properties. It seems that the system gain not only is different for the different channels, but also changes with the light. For instance, the blue channel shows a higher system gain indoors on a white sheet, than outdoors taking the clear blue sky.
mfurman wrote:
Would you prefer that manufactures stopped listing the number of pixels in camera specifications? Would you like to use subjective comparisons similar to what is common for High End audio reviews?
Not at all, but I would like to see reviewers ensure equal magnification when comparing detail and noise between various cameras. It should be a rule of basic reviewing competence.
And it's important to understand that cropping 8 Mpixels off the 7D file, and not doing so with a 40D file is a method of comparison that, while it's instructive for you, Mike, and your intended use of the camera (probably in a minority of situations), would be a long way from normal usage for the great majority of people who will buy this camera, and who will use the same lenses that they used on their 40D's, and who will not routinely discard 45% of the captured image just because it's a higher rez camera.
First, I need a target without any image detail. With the clear blue sky, I find a direction with little tonal gradient, and use a telephoto lens (400 mm). I stop the lens down to reduce vignetting, and make the image out of focus, however not at infinity.
With a white sheet indoors, i illuminate as even as I can, and focus at infinity.
To get away from the read noise, I expose near 50% from raw clipping level.
I know that frame subtraction is advised to remove all image detail, but I'm afraid that any alignment error will spoil it, so I use one frame only.
After CFA separation, I make small crops near the middle. With small crops I mean 1000-10000 pixels. I use different crops, and discard if it is off the others.
These crops should now as far as possible contain the mean value and dominantly photon shot noise.
According to the mean-variance method, the system gain should be
mean/(sigma^2)
All this, i do at different ISO's, and if the quality of the first steps is high, the results are consistent over a broad range of ISO's.
brainiac:
And it's important to understand that cropping 8 Mpixels off the 7D file, and not doing so with a 40D file is a method of comparison that, while it's instructive for you
It is not, as you may imply, an exercise in crippling one camera just to prove the point. I think that if one needs a 10 Mpixel image, one could settle on certain IQ of this image size as adequate and intend to use the higher Mpixel sensor to "improve reach". If the resolution at a given ISO is proportional to the number of pixels, it should be a straightforward process. Unfortunately those pixels are not of the same quality for different cameras.
alundeb wrote:
The simple answer is: Somewhere in between.
With the 7D you can crop into a 20% narrower Angle of View and maintain the same noise level in the resulting image. You will still have more detail, as you now have a 14 MP image.
Error margin +/- 10 %.
Thanks alundeb. That is the kind of thing I think people are trying to figure out. So, to subjectively answer mfurman, the 7D cropped to 14 MP is approximately a 1.2x teleconverter compared to the 40D at equivalent IQ. I guess that is not too bad since mfurman's 10 MP crop would be equivalent to a 1.33x teleconverter with unknown IQ deterioration.
This is all a very subjective continuum from a 10 MP crop to a 10 MP down sample. I really like the IQ and MP my 5D classic gives me when it is not reach limited. So before it is worth it to spend $1700 on a new crop body, I would want to be able to downsample to the 5D's 12.8 MP, achieving the 5D's overall IQ, and still have some magnification gain on my 40D when viewing the resulting images. Is this too much to expect?
Jeff Nolten:
So before it is worth it to spend $1700 on a new crop body, I would want to be able to downsample to the 5D's 12.8 MP, achieving the 5D's overall IQ
This is probably a much more important question than my 10 Mpixel cropping exercise would answer. Let's forget 40D and see if the crop camera reached the quality of a four year old FF. If it did, it would mean that $1700 is a good price to pay (of course others would say that the focus and video alone are worth it ) for 7D
mfurman wrote:
It is not, as you may imply, an exercise in crippling one camera just to prove the point. I think that if one needs a 10 Mpixel image, one could settle on certain IQ of this image size as adequate and intend to use the higher Mpixel sensor to "improve reach". If the resolution at a given ISO is proportional to the number of pixels, it should be a straightforward process. Unfortunately those pixels are not of the same quality for different cameras.
I'm not saying that you're trying to cripple the camera to prove a point. I absolutely understand why you want to find out how much worse 10 million 7D pixels are than 10 million 40D pixels, through your inadequate lenses (if your lenses were adequate you wouldn't be forced to crop down to 10 Mpixels). If your lenses are good enough, then you may find that 10 million 7D pixels are pretty close to 10 million 40D pixels at low isos. I agree that reach is one of the big advantages of the 7D over the 40D. My point was just that most photographers in typical use will be aiming to use as much of the frame as possible, and will use appropriate lenses to do so. The 40D and 7D have the same angle of view, so normal usage won't be represented at all by the comparison that you want to do. It only applies when lenses aren't long enough. Having the right length of lens probably accounts for more quality than a 7D over a 40D or a 5D2 over a 7D. Who knows, even you may one day take an image which uses the majority of your frame ;-)
alundeb wrote:
First, I need a target without any image detail. With the clear blue sky, I find a direction with little tonal gradient, and use a telephoto lens (400 mm). I stop the lens down to reduce vignetting, and make the image out of focus, however not at infinity.
With a white sheet indoors, i illuminate as even as I can, and focus at infinity.
To get away from the read noise, I expose near 50% from raw clipping level.
I know that frame subtraction is advised to remove all image detail, but I'm afraid that any alignment error will spoil it, so I use one frame only.
After CFA separation, I make small crops near the middle. With small crops I mean 1000-10000 pixels. I use different crops, and discard if it is off the others.
These crops should now as far as possible contain the mean value and dominantly photon shot noise.
According to the mean-variance method, the system gain should be
mean/(sigma^2)
All this, i do at different ISO's, and if the quality of the first steps is high, the results are consistent over a broad range of ISO's. ...Show more →
Frame subtraction isn't just to remove the effects of image detail, it helps remove the effects of pattern noise, hot/dead pixels, issues of pixel response non-uniformity (which may be an issue, eg skibum's 7D seems to have an issue with varying gains for different columns) etc. But most importantly, even if there is no detail but just a modest gradient across the image, that will affect the std dev. I have used a method similar to yours (blue sky, telephoto lens defocussed but I still use differences of pairs of frames.
I still find it puzzling that you are finding more accurate determination of the gain at high ISO than at low ISO; that should never be the case, due to read noise effects.
Are you seeing the difference of green subarray outputs reported by TheSuede?
To brainiac (as an additional explanation of my reasons):
My constant back and force about 7D comes from the fact that I do want and have to keep and use 40D (long story). The decision then is if I should get 7D or 5D mkII to supplement (no matter how silly it sounds) 40D. With 5D mkII, it is easy to justify but the reasons to get 7D and keep 40D are more difficult for me to find
alundeb wrote:
However, if this really is the case, the noise and dynamic range will not be static properties, but vary with the type of light, differently than other cameras.
The 7D has a "completely new" dual-layer exposure meter which recognises different colours of light (one layer sensitive to red/green, the other layer sensitive to blue/green) and adjusts the exposure to account for the proportions of these different spectra it receives - I wonder if what you describe is a function of that?
(And - thinking about it - I wonder if this dual-layer meter might account for the apparent imbalance in the green channels?)
Apologies if I'm late to that party!
The camera also uses this information in its AF processing too, using it to adjust "for any CA in the AF system" (says Canon) to improve AF accuracy in (for example) artificial lighting, so it's clearly making a lot of use of this apparent ability to recognise different kinds of light...
I completely understand your noise equivalence theory.
Continuously repeating yourself does not make your point more valid.
There are two issues with it.
First of all, it works in theory but not in practice.
That's because in practice, pixel S/N ratio does not decrease linearly as you shrink pixels.
In other words, a 2x smaller pixel does not necessarily have a 2x worse S/N ratio - as your theory would predict.
It could be 3x-5x worse or 2x better - this is entirely technology dependent at this time and has nothing to do with theory.
Since Canon doesn’t publish the exact technical specs of their sensors, your theory is pure speculation at this time because it is based on the theoretical assumption that there’s a linear relationship between shrinking pixels and decreasing S/N ratios – which is simply not true in practice.
The second issue with your theory is that even if we assume that it holds true, there’s a practical limit of how much it makes sense to shrink pixels.
According to your theory (or should I say speculation), the current 21mp sensor in the 5DII is no better than a 100mp sensor with much worse pixel S/N ratio, right?
But let me ask you this: would you be happy if Canon decides to push down your throat a 100mp sensor, telling that it’s OK – all you have to do is downsample and you will get the same image quality as the 5DII?
This is extreme, of course, but illustrates the point that I’m making:
Even if downsampled images from an 18mp sensor are the same as 12mp images with better pixel level S/N, I’d still prefer the latter.
For my needs, 12mp is a good enough resolution.
So, I’d rather see that manufacturers maximize the S/R ratio for 12mp – rather than giving me more resolution (which I don’t need) with worse pixel level noise (which I don’t like) and telling me that I can downsample and I’ll be fine (which I find arrogant).
I’m a customer. The customer is always right and if they want to sell me something, they'd better make what I'd buy.
mfurman wrote:
This is probably a much more important question than my 10 Mpixel cropping exercise would answer. Let's forget 40D and see if the crop camera reached the quality of a four year old FF. If it did, it would mean that $1700 is a good price to pay (of course others would say that the focus and video alone are worth it ) for 7D
Yes, I think the 7D might possibly match the 5D's noise when downsampled to its resolution and still provide a bit of reach improvement on the 40D. The other new features, especially movies, AF and AE are very attractive. That is why I am paying the time to follow this thread. Besides dollar cost, however, there are also costs in added post processing to carefully noise reduce and downsample for IQ, and to learn to work and then remember to set all of the new AF adjustments to the current shooting conditions. I will probably have to upgrade my computer system to even use the movies. Lots to think about. Today I am leaning towards keeping my 40D which has been a solid performer.
keithreeder wrote:
The 7D has a "completely new" dual-layer exposure meter which recognises different colours of light (one layer sensitive to red/green, the other layer sensitive to blue/green) and adjusts the exposure to account for the proportions of these different spectra it receives - I wonder if what you describe is a function of that?
(And - thinking about it - I wonder if this dual-layer meter might account for the apparent imbalance in the green channels?)
Apologies if I'm late to that party!
The camera also uses this information in its AF processing too, using it to adjust "for any CA in the AF system" (says Canon) to improve AF accuracy in (for example) artificial lighting, so it's clearly making a lot of use of this apparent ability to recognise different kinds of light... ...Show more →
At least, they have made dynamic color gain adjustment possible.
About the green channel imbalance, I believe theSuede has solved that mystery now by proving different spectral responses. Bravo!
Again, that triggered my thought, and I found in one of my mean-variance measurements: Heck, the two green channels are balanced in this one, they have the same mean and the same variance. I cannot by now find any other explanation other than a dynamic system gain. This is a bit difficult to explain, but the ratio between mean and variance doesn't change for different exposure and color alone. It is closely linked to the ratio of collected photons per raw unit, or the system gain.
jorkata wrote:
For my needs, 12mp is a good enough resolution.
So, I’d rather see that manufacturers maximize the S/R ratio for 12mp – rather than giving me more resolution (which I don’t need) with worse pixel level noise (which I don’t like) and telling me that I can downsample and I’ll be fine (which I find arrogant).
I’m a customer. The customer is always right and if they want to sell me something, they'd better make what I'd buy.
I'm a customer too. I bought the 7D because it has 18MP. I would not have if it was 12.
What is most arrogant, to tell you to downsample, or to oversee my needs completely?
That's why in my previous posts I'm suggesting that Canon should make the 7D in two flavors - one with lots of megapixels for those who like that, and another one will less megapixels and better ISO.
Let buyers decide what works best for them (and not argue what works better for the others).
Canon will have the same problem if they decide to push the 5D resolution beyond 30mp.
Not all FF buyers will be happy with that for sure.
Guys rookie question, how much better (compared to old 1dmk2) can I expect the high ISO performance of this 7D? What if I throw 40d into the picture also.
Currently I have 1dmk2 and need another body. It is either 40d with 17-55mm f2.8 IS or just the bare 7d. If 7d can come close to my 1dmk2 as far as AF is concerned, I will be very happy. Looking at higher ISO performance also as shooting football at ISO3200 is not that good with my 1dmk2.
keithreeder wrote:
The 7D has a "completely new" dual-layer exposure meter which recognises different colours of light (one layer sensitive to red/green, the other layer sensitive to blue/green) and adjusts the exposure to account for the proportions of these different spectra it receives - I wonder if what you describe is a function of that?
(And - thinking about it - I wonder if this dual-layer meter might account for the apparent imbalance in the green channels?)
The metering system simply suggests an overall exposure time and f/ratio; it does not go and tune the exposure of individual pixels/columns/color planes.
If true, the difference is due to a use of different spectral transmissivity for the two "Green" locations in the Bayer array. IIRC (though perhaps I do not) Canon did something like this in some early P&S models, using a CRGB color filter array (ie one of the Green locations was replaced by Cyan).
