D4 x 5D3 ISO series in ACR, IR raws (large file)

DSPographer said:

Joofa said:

So with real cameras out there who is right: DSPographer or Bobn2?

Click to expand...

5 years ago state-of-the-art was about 3e- read noise independant of pixel size. Today it is about 2 e-, again independant of pixel size. Which sensors with 1.4 micron pixels have better read noise than a D3s? The D3s has about 2.8 e- read noise so which are better than 2.8*(1.4/8.4) less than 1/2 e- in read noise?

Click to expand...

DSPographer said:

Joofa said:

rhlpetrus said:

Joofa said:

So who is right: DSPographer or Bobn2?

Click to expand...

If you read posts by both you'll see that there a difference in perspective, explained to me by DSP:

He is talking about pure rescaling of sensor sites, while Bob assumes that rescaling involves updating the whole sensor as expected from a technical POV. In the former, IQ may decay in some aspects, in Bob's perspective, that would not happen.

Click to expand...

Thanks for the explanation. But that is why I asked about the "pencil and paper design vs. reality" in my message above, i.e., not what could be done in theory, but what is really happening out there? So lets come out of the "perspectives", and see what the real cameras are doing out there. Right?

So with real cameras out there who is right: DSPographer or Bobn2?

Click to expand...

5 years ago state-of-the-art was about 3e- read noise independant of pixel size. Today it is about 2 e-, again independant of pixel size. Which sensors with 1.4 micron pixels have better read noise than a D3s? The D3s has about 2.8 e- read noise so which are better than 2.8*(1.4/8.4) less than 1/2 e- in read noise?

Click to expand...

DSPographer said:

Joofa said:

rhlpetrus said:

Joofa said:

So who is right: DSPographer or Bobn2?

Click to expand...

If you read posts by both you'll see that there a difference in perspective, explained to me by DSP:

He is talking about pure rescaling of sensor sites, while Bob assumes that rescaling involves updating the whole sensor as expected from a technical POV. In the former, IQ may decay in some aspects, in Bob's perspective, that would not happen.

Click to expand...

Thanks for the explanation. But that is why I asked about the "pencil and paper design vs. reality" in my message above, i.e., not what could be done in theory, but what is really happening out there? So lets come out of the "perspectives", and see what the real cameras are doing out there. Right?

So with real cameras out there who is right: DSPographer or Bobn2?

Click to expand...

5 years ago state-of-the-art was about 3e- read noise independant of pixel size. Today it is about 2 e-, again independant of pixel size. Which sensors with 1.4 micron pixels have better read noise than a D3s? The D3s has about 2.8 e- read noise so which are better than 2.8*(1.4/8.4) less than 1/2 e- in read noise?

Click to expand...

DSPographer said:

5 years ago state-of-the-art was about 3e- read noise independant of pixel size. Today it is about 2 e-, again independant of pixel size. Which sensors with 1.4 micron pixels have better read noise than a D3s? The D3s has about 2.8 e- read noise so which are better than 2.8*(1.4/8.4) less than 1/2 e- in read noise?

Click to expand...

The_Suede said:

DSPographer said:

5 years ago state-of-the-art was about 3e- read noise independant of pixel size. Today it is about 2 e-, again independant of pixel size. Which sensors with 1.4 micron pixels have better read noise than a D3s? The D3s has about 2.8 e- read noise so which are better than 2.8*(1.4/8.4) less than 1/2 e- in read noise?

Click to expand...

At these levels, does it really matter?

Depending on the method used to Bayer-interpolate (mainly the support area size chosen) you get very different results from the two options.

One - large pixels - will have better SNR per pixel, as Qe * area increases without any noticeable increase in total RN.

The other - small pixels - will have a much better interpolation support, since you can increase the support area (counted in pixel widths, not µm of course!) without really affecting the available resolution in the finished image.

The difference this makes to the raw converter and the rest of the PP chain is not trivial. The larger pixels will GENERALLY have a better Bayer interpolation accuracy per pixel (the two missing colors will be more accurately estimated), but also a MUCH higher susceptibility to impulse noise, or just the unfortunate pixels where the photon statistics go totally off the chart in predictability (it has to happen in quite a few tens of thousands of pixels in a 20MP+ raw file...)

They each give very different "qualities" to the finished image, and I must say that i prefer the downsampled high-res image by quite a large margin. The noise grain, especially the Chroma noise, is a lot "tighter", which has two very tangible effects in post: You can apply more NR without affecting image detail, and you don't NEED to apply as much NR - as the tighter noise grain is a lot less "digital", and hence also causes less objectionable responses in the human visual recognition system.

Finer pitched noise grain can be stronger in total P-P oscillation than coarse grain noise (which means that it will have a much higher total energy content), and you will still accept it as a part of the image rather than a digital artefact.

So it's not just about the numbers, there's an entire path of interpolations and convolutions going on before the image ends up on your screen or in your print - at the intended viewing size. The lowered SNR per pixel may actually end up as an increased perceived SNR / detail in the finished image.

Click to expand...

Mikael Risedal said:

d7000 ?

Click to expand...

Joofa said:

DSPographer said:

Joofa said:

So with real cameras out there who is right: DSPographer or Bobn2?

Click to expand...

5 years ago state-of-the-art was about 3e- read noise independant of pixel size. Today it is about 2 e-, again independant of pixel size. Which sensors with 1.4 micron pixels have better read noise than a D3s? The D3s has about 2.8 e- read noise so which are better than 2.8*(1.4/8.4) less than 1/2 e- in read noise?

Click to expand...

Thanks for the message. I guess you win then. .

--
Dj Joofa

Click to expand...

DSPographer said:

Mikael Risedal said:

d7000 ?

Click to expand...

No.

I was quoting the minimum read noise from sensorgen.info for the D3s, but some care is needed with that site. If we check the table for all ISO values we see that the read noise of the D3s continues to drop for extremely high ISO value: but these may not be correct. The sensorgen table comes from DXOmark data that derives from a fit for read noise, shot noise and PRNU. At high ISO levels sometimes the DXOmark fit goes haywire with a negative value for PRNU (which is physically impossible). So if we avoid the extreme ISO settings we get 3.5 e- minimum read noise for the D3s at ISO 6400 and 2.9 e- minimum read noise at ISO 1600 for the D7000. But the D7000 has 4.73 micron pixels, so to have equal read noise per area the D7000 value would need to be: 3.5 * (4.73 / 8.4) = 2 e-. So the per area read noise of the D7000 is greater than for the D3s.

Click to expand...

The_Suede said:

DSPographer said:

5 years ago state-of-the-art was about 3e- read noise independant of pixel size. Today it is about 2 e-, again independant of pixel size. Which sensors with 1.4 micron pixels have better read noise than a D3s? The D3s has about 2.8 e- read noise so which are better than 2.8*(1.4/8.4) less than 1/2 e- in read noise?

Click to expand...

At these levels, does it really matter?

Depending on the method used to Bayer-interpolate (mainly the support area size chosen) you get very different results from the two options.

One - large pixels - will have better SNR per pixel, as Qe * area increases without any noticeable increase in total RN.

The other - small pixels - will have a much better interpolation support, since you can increase the support area (counted in pixel widths, not µm of course!) without really affecting the available resolution in the finished image.

The difference this makes to the raw converter and the rest of the PP chain is not trivial. The larger pixels will GENERALLY have a better Bayer interpolation accuracy per pixel (the two missing colors will be more accurately estimated), but also a MUCH higher susceptibility to impulse noise, or just the unfortunate pixels where the photon statistics go totally off the chart in predictability (it has to happen in quite a few tens of thousands of pixels in a 20MP+ raw file...)

They each give very different "qualities" to the finished image, and I must say that i prefer the downsampled high-res image by quite a large margin. The noise grain, especially the Chroma noise, is a lot "tighter", which has two very tangible effects in post: You can apply more NR without affecting image detail, and you don't NEED to apply as much NR - as the tighter noise grain is a lot less "digital", and hence also causes less objectionable responses in the human visual recognition system.

Finer pitched noise grain can be stronger in total P-P oscillation than coarse grain noise (which means that it will have a much higher total energy content), and you will still accept it as a part of the image rather than a digital artefact.

So it's not just about the numbers, there's an entire path of interpolations and convolutions going on before the image ends up on your screen or in your print - at the intended viewing size. The lowered SNR per pixel may actually end up as an increased perceived SNR / detail in the finished image.

Click to expand...