Why are the tests of IQ250 against CCD-backs taking so long?

[bjanes]

Good test/comparison.  The noise characteristics in your Canon 5D MK2 are so familiar to me.  The blue box is covered with red noise and if you zoom in (best done from an ipad) you can see some vertical banding in the Canon also.  I found this in all canon files until the 6D, which has a much more even noise, almost like film grain.   The 7100 shot also is not surprising to me, as all of the 24MP crop sensors I have used/demo'd show pretty harsh noise somewhere.  The Sony Nex-7 was even worse than the 7100 here. 

I agree in your test the Aptus is cleaner in the way the noise is displayed. 

I don't agree. The comparison has serious flaws. The patches are unequal in size among the cameras, so I presume the images are 100% views of the rendered files. This is what DXO would call Screen mode. It is not surprising that the Aptus has better per pixel performance than the D7000, since the pixel sizes are 8.98 and 4.89 microns respectively. For a valid comparison, one would have to resize the images so that the sizes of the patches are equal, analogous to Print Mode in DXO parlance. Although the per pixel performance of the D7000 and D800 may be similar, the D800 collects nearly twice as much light due to its larger sensor size and the print performance is better.

Furthermore, the raw converters may have applied noise reduction. One should really look at raw files. Also, one should confirm that sensor saturation is similar among the shots (equally exposed to the right).

Bill

[gerald.d]

Forgive my ignorance, but why...

"It is not surprising that the Aptus has better per pixel performance than the D7000, since the pixel sizes are 8.98 and 4.89 microns respectively."

... is it necessary for anything to follow that statement?

What are you implying with that statement?

That for the same resolution, medium format will have a clear advantage over 35mm because its pixels are larger?

Is it really as simple as all that?

[torger]

I don't agree.

I have the complete test here if you're interested: http://www.ludd.ltu.se/~torger/photography/noise-test.html

I explain there why it's pixel per pixel, but in short -- the reason I get a back with more resolution is to be able to make detailed prints of landscapes that look nice up close. I want 80 megapixels to print twice the size as 40 megapixels. If I must reduce the size to equal size as the 40 megapixel print to smooth noise I could just have kept the 40 megapixel back. Sure if you want to have the same print sizes you're free to scale down, and I discuss that in the test too.

The raw converter is RawTherapee, there's no noise reduction or different curves applied, just a standard matrix.

It's a straightforward test to verify some noise characteristic basics I was interested when I first got my back in 2012. I'll do a similar test when/if I upgrade to Dalsa 6um technology, but then look more at gradients and color stability in pushed shadows. I like to make some basic technical test to establish a baseline so I know what I got. Say if this old back would have had really bad noise performance I would have tried to get a newer back, but I could see here that it did perform very well for its age (which is not surpising in the world of CCDs).

I'm not particularly interested in having a back that performs best, I can't afford the latest anyway, but I want to verify that it's fairly competitive with the state of the art. The reason I use a digital back is for camera (tech camera), lenses and upgradability in terms of resolution. As a landscape photographer I couldn't care less about skin colors, but of course understand it's a main driver in the professional segment. To me resolution and dynamic range are the most interesting aspects. I'm becoming more interested in color separation in shadows though. I've noticed that cameras can lose it there, and I'm not sure it's a linear relationsship to noise.

[Fine_Art]

Forgive my ignorance, but why...

"It is not surprising that the Aptus has better per pixel performance than the D7000, since the pixel sizes are 8.98 and 4.89 microns respectively."

... is it necessary for anything to follow that statement?

What are you implying with that statement?

That for the same resolution, medium format will have a clear advantage over 35mm because its pixels are larger?

Is it really as simple as all that?

Not the pixels are larger, the sensor is larger so more total light data is recorded. Pixel size is a tradeoff between better location info vs better DR/ lower noise.

[Theodoros]

An IQ260 achromatic with a filter wheel should be the ultimate multi-shot back, right? For reproduction it could be feasible. Filter wheels have been used before.

Concerning CCD vs CMOS etc raw files shot side by side on the same subject would be needed for a proper comparison. Differences is probably going to be so small that you'd have to make A/B testing to draw any reliable conclusions. It would be great to see that with IQ260 vs IQ250 and throw in a D800 or A7r for fun.

It wouldn't be able to do 16X… and it wouldn't improve detail in 4X… only colour would benefit.

[ErikKaffehr]

Hi,

I may agree with Anders Torger, but I just looked one more time at Chris Barrets images and I don't think the IQ 260 stands a chance against the A7r in his exposures (which were exposed for highlights).

Chris used 100 ISO, as this is what he regards to be practical. Both images processed the same way. The second image downscaled A7r vertical resolution.

The DR advantage of MF CCDs is clearly a myth.

Best regards
Erik

I have the complete test here if you're interested: http://www.ludd.ltu.se/~torger/photography/noise-test.html

I explain there why it's pixel per pixel, but in short -- the reason I get a back with more resolution is to be able to make detailed prints of landscapes that look nice up close. I want 80 megapixels to print twice the size as 40 megapixels. If I must reduce the size to equal size as the 40 megapixel print to smooth noise I could just have kept the 40 megapixel back. Sure if you want to have the same print sizes you're free to scale down, and I discuss that in the test too.

The raw converter is RawTherapee, there's no noise reduction or different curves applied, just a standard matrix.

It's a straightforward test to verify some noise characteristic basics I was interested when I first got my back in 2012. I'll do a similar test when/if I upgrade to Dalsa 6um technology, but then look more at gradients and color stability in pushed shadows. I like to make some basic technical test to establish a baseline so I know what I got. Say if this old back would have had really bad noise performance I would have tried to get a newer back, but I could see here that it did perform very well for its age (which is not surpising in the world of CCDs).

I'm not particularly interested in having a back that performs best, I can't afford the latest anyway, but I want to verify that it's fairly competitive with the state of the art. The reason I use a digital back is for camera (tech camera), lenses and upgradability in terms of resolution. As a landscape photographer I couldn't care less about skin colors, but of course understand it's a main driver in the professional segment. To me resolution and dynamic range are the most interesting aspects. I'm becoming more interested in color separation in shadows though. I've noticed that cameras can lose it there, and I'm not sure it's a linear relationsship to noise.

[gerald.d]

Not the pixels are larger, the sensor is larger so more total light data is recorded. Pixel size is a tradeoff between better location info vs better DR/ lower noise.

Sorry, but the claim was... of course it will have "better per pixel performance", because the pixels are bigger.

Not my words.

And presumably, for the same resolution, better per pixel performance will translate into better performance at the sensor level when comparing larger sensors to smaller ones.

The question remains - assuming resolution is the same across both formats, is that fundamentally what it boils down to? bjanes would appear to be saying that you have to remove the fundamental advantage that a larger sensor brings to the table, precisely because it gives it an advantage.

[Fine_Art]

Sorry, but the claim was... of course it will have "better per pixel performance", because the pixels are bigger.

Not my words.

And presumably, for the same resolution, better per pixel performance will translate into better performance at the sensor level when comparing larger sensors to smaller ones.

The question remains - assuming resolution is the same across both formats, is that fundamentally what it boils down to? bjanes would appear to be saying that you have to remove the fundamental advantage that a larger sensor brings to the table, precisely because it gives it an advantage.

Ok, per pixel performance.

There is a fundamental assumption we all make, that is the bigger MFDB image circle starts with a bigger lens. The primary purpose of any lens is to gather light data. Bigger area, more data. If the starting lens is not bigger, there is no advantage to MFDB. Etendue is conserved, you cannot make something out of nothing.

So if (big if due to weight concerns) MFDB makers create big lenses (larger than the 77mm that is the common high end of 135 format) they will always be able to resolve it into a bigger image circle with more quality, assuming similar manufacturing quality.

The easiest way to make MFDB look bad is to saddle it with a 55mm diameter lens gathering no more data than what is slapped on a 135 D800 or A7r.

So the advantage of larger format will always be real if they make big lenses. If they don't they will die off.

I agree with you all images should be shown at normal resolution. Compare what it is, not a fake rendered apple for an apples to apples comparison.

[bjanes]

I have the complete test here if you're interested: http://www.ludd.ltu.se/~torger/photography/noise-test.html

I looked at the complete test and it does address my concerns and I retract my criticism.

Bill

[Doug Peterson]

Don't know where this thread drifted to. But here are some pretty pictures to fight about instead of just numbers and engineering theory:
http://www.luminous-landscape.com/forum/index.php?topic=87143.0

[bjanes]

Sorry, but the claim was... of course it will have "better per pixel performance", because the pixels are bigger.

Not my words.

And presumably, for the same resolution, better per pixel performance will translate into better performance at the sensor level when comparing larger sensors to smaller ones.

The question remains - assuming resolution is the same across both formats, is that fundamentally what it boils down to? bjanes would appear to be saying that you have to remove the fundamental advantage that a larger sensor brings to the table, precisely because it gives it an advantage.

Other things being equal, a larger pixel will have a better signal to noise ratio since it collects more photoelectrons. SNR is proportional to the square root of the number of photons collected, so doubling the number of photons collected by doubling the pixel area will improve the SNR by a factor of 1.414. However, for a given sensor size, increasing the pixel size will decrease the resolution. In comparing sensors with different resolutions, it is necessary to normalize the SNR and DR values. DXO does this by resampling to 8 MP. This does not remove the fundamental advantage of a higher resolution sensor. One could alternatively upres the lower resolution sensor. This is the concept of the DXO print DR figures. If you are printing at a given print size and resolution, the file has to be resized to match the resolution of the printer. If one has downsized the image with a higher resolution, the SNR will be improved by pixel binning.

Bill

[Fine_Art]

Other things being equal, a larger pixel will have a better signal to noise ratio since it collects more photoelectrons. SNR is proportional to the square root of the number of photons collected, so doubling the number of photons collected by doubling the pixel area will improve the SNR by a factor of 1.414. However, for a given sensor size, increasing the pixel size will decrease the resolution. In comparing sensors with different resolutions, it is necessary to normalize the SNR and DR values. DXO does this by resampling to 8 MP. This does not remove the fundamental advantage of a higher resolution sensor. One could alternatively upres the lower resolution sensor. This is the concept of the DXO print DR figures. If you are printing at a given print size and resolution, the file has to be resized to match the resolution of the printer. If one has downsized the image with a higher resolution, the SNR will be improved by pixel binning.

Bill

Bill, that may be standard procedure, it still makes one of the compared items into something it is not. I think most of us can look at images side by side and realize which has properties superior to our use. For some the extra resolution may be a deciding factor. For others it may be something else entirely. For DxO who has made sorting/ ranking a big part of their site, that is important. For us as users I think the stats are separate from the images. We look at the stats to answer certain things. We look at the images to get a feel of the camera's output.

[Fine_Art]

Maybe the way to say it is sometimes the side by side relative comparison is important see DPR reviews. Sometimes the actual output is what you need to see as in Imaging resource full images. IIRC DPR shows small sections at native resolution. You can scroll around to see how each camera handles the same part of the scene. I find I can mentally compare even if one is bigger than the other.

[synn]

Other things being equal, a larger pixel will have a better signal to noise ratio since it collects more photoelectrons. SNR is proportional to the square root of the number of photons collected, so doubling the number of photons collected by doubling the pixel area will improve the SNR by a factor of 1.414. However, for a given sensor size, increasing the pixel size will decrease the resolution. In comparing sensors with different resolutions, it is necessary to normalize the SNR and DR values. DXO does this by resampling to 8 MP. This does not remove the fundamental advantage of a higher resolution sensor. One could alternatively upres the lower resolution sensor. This is the concept of the DXO print DR figures. If you are printing at a given print size and resolution, the file has to be resized to match the resolution of the printer. If one has downsized the image with a higher resolution, the SNR will be improved by pixel binning.

Bill

Talk is cheap.
I'm still awaiting the awesome pictures you've made with all this science.

[Fine_Art]

No need for that, I am sure I have seen good shots from Bill over at Photo.net

[ErikKaffehr]

Hi,

The images below show a shadow crop from a landscape shot  on a P45+ and a Sony Alpha SLT, both pushed 4 EV. No noise reduction is used but same sharpening applied. What we would see on screen or in print would be more like the top row than the bottom row.

Comparing sensor technology, it is more tricky. Using similar size crops and scaled to same dimension would be most proper in my view.

	P45+ (39 MP MFD)	Sony Alpha 99 (24 MP FF)
Dowscaled both to same size
Actual pixels

Best regards
Erik

Other things being equal, a larger pixel will have a better signal to noise ratio since it collects more photoelectrons. SNR is proportional to the square root of the number of photons collected, so doubling the number of photons collected by doubling the pixel area will improve the SNR by a factor of 1.414. However, for a given sensor size, increasing the pixel size will decrease the resolution. In comparing sensors with different resolutions, it is necessary to normalize the SNR and DR values. DXO does this by resampling to 8 MP. This does not remove the fundamental advantage of a higher resolution sensor. One could alternatively upres the lower resolution sensor. This is the concept of the DXO print DR figures. If you are printing at a given print size and resolution, the file has to be resized to match the resolution of the printer. If one has downsized the image with a higher resolution, the SNR will be improved by pixel binning.

Bill

[Fine_Art]

I still have to disagree. When I look at your P45 shot down-sampled, I see large swaths of magenta and the same on blue-green. It looks unrecoverable. No random noise filter is going to recognize it is not part of the image. When I look at the full size P45 shot I see dots of these colors that I think I could strip out. The impression of the file completely changes.

[ErikKaffehr]

Hi,

Good observation. Noise may be more or less managable.

Best regards
Erik

I still have to disagree. When I look at your P45 shot down-sampled, I see large swaths of magenta and the same on blue-green. It looks unrecoverable. No random noise filter is going to recognize it is not part of the image. When I look at the full size P45 shot I see dots of these colors that I think I could strip out. The impression of the file completely changes.

[Fine_Art]

The noise would look a lot more like noise in RT. Even on the full size image the splotching of magenta, cyan is clearly not part of the scene. That is a LR failure. If you (the software) can't remove color noise, you should leave it at fine pixel by pixel randomness for a capable noise program.

[ErikKaffehr]

Hi,

If you check my posting you would note that I had noise reduction disabled. I wanted to investigate the noise in the sensor and not how it affects noise reduction.

The reasons for that were threefold:

1 - normally try to keep noise recuction low
2 - if you go into noise reduction there will be a an infinite number of possible combinations of options
3 - it is my belief that noise should be kept to minimum, an image with no or little noise is preferable to an image that depends on noise reduction

Best regards
Erik

The noise would look a lot more like noise in RT. Even on the full size image the splotching of magenta, cyan is clearly not part of the scene. That is a LR failure. If you (the software) can't remove color noise, you should leave it at fine pixel by pixel randomness for a capable noise program.