FF low-light advantage vs. crop?

kcbrown wrote in post #9470687
There is a physical difference on a per-frame basis due to the difference in sensor size.

Excellent post! Very clear and informative. Thank you for contributing to this thread.

xarqi wrote in post #9470210
Sensor size is totally irrelevant. Photosite (loosely termed "pixel") size matters.

That is incorrect. If it were true, then a digicam-sized sensor with 5D2 pixels would have the same noise as the 5D2 itself. This is easily proven false using an experiment like I did above.

blackhawk wrote in post #9475344
FF has nothing to do with low light performance and may even diminish it...

That is incorrect. A larger sensor size with the same f-number and performance characteristics results in much improved low light performance.

blackhawk wrote in post #9475344
The nuts and bolts of it: http://www.clarkvision​.com/articles/index.ht​ml#part_4

I think you are referring to the cases where Roger Clark makes statements about performance at the Nyquist frequency (pixel level). Unfortunately, he does not consider the performance at a fixed spatial frequency, which is the only measurement that reflects real world photography (e.g. any given crop size or print size).

Daniel Browning wrote in post #9480249
That is incorrect. If it were true, then a digicam-sized sensor with 5D2 pixels would have the same noise as the 5D2 itself. This is easily proven false using an experiment like I did above.

Read my analysis above.

xarqi wrote in post #9477335
1) Larger photosites are inherently less noisy, therefore photosite size is a significant contributor to noise performance.

That is never true for tonal levels predominated by photon shot noise at any given spatial frequency. Pixel sizes from 3 microns up to 15 microns have the same QE. Even in read noise dominated levels it is only a loose correlation; there are often cases where read noise scales with pixel size to result in the same read noise level. The 5D2 and 7D is just one such case: when sensor size and spatial frequency is equalized between them, they are the same.

xarqi wrote in post #9477335
2) For the purposes of generating RAW data for one photosite, the total number of photosites in the sensor is unimportant. There is no cross-talk, either of signal or of noise.

I agree.

xarqi wrote in post #9477335
3) The only noise present in the raw data for the entire sensor is the noise present in the data for each photosite.

That's like saying "the only horsepower present in the engine is the horsepower present for each cylinder."

xarqi wrote in post #9477335
4) Sensor size does not affect the noise in the data.

Sensor size does not affect noise per pixel, but it does affect noise per spatial frequency. Just as engine size does not affect horsepower per cylinder, but it does affect the horsepower of the entire engine.

xarqi wrote in post #9477335
That is the logic that lead me to state that photosite size is important in determining noise, whereas sensor size is not.

Ignore spatial frequency with the wave of a hand invalidates the logic.

xarqi wrote in post #9477335
5) This noise is what I think is being called "per pixel noise", but there are claims of a different sort. If you accept the logic above, then this other type of noise can only derive ultimately from pixel noise as modified during post-processing.

6) There is noisy data, the magnitude of the noise depending on photosite size, but not sensor size. If, during processing, this data is interpolated, or averaged in any way then random noise will tend to cancel.

7) To reach the same print size, data from a larger sensor will need a greater degree of interpolation since the required magnification is less. Therefore, in the post-processing of a larger format image, there will necessarily be a greater degree of noise cancellation than for a smaller format. The noise remaining after processing is the "per frame" noise.

I wouldn't put it like that, but I understand you and I wont object.

xarqi wrote in post #9477335
Where does that leave us?
The ultimate source of noise is the photosite, and larger photosites are less noisy.

Post-processing of noisy data can cancel some noise, the degree of cancellation depending on sensor size.

Are larger sensors inherently less noisy?
No. However, they allow a greater degree of noise cancellation during post-processing.

The way you put it is very misleading, if technically correct. It's like this:

Where does that leave us?
The ultimate source of horsepower is the cylinder, and larger cylinders have more horsepower.

Using all the cylinders in an engine at one time can increase total horsepower, the degree of horsepower increase depending on engine size.

Do larger engines inherently have more horsepower?
No. However, they allow a greater degree of horsepower increase when using more than one cylinder at the same time.

Again, it's technically true, but misleading. First because you make it sound like there is some kind of extra post processing required. There's not. Just display the image at the same size and the larger sensor will appear less noisy. The only time it appears same or more noisy is when you display them at different sizes, or crop differently (e.g. compare headshot vs environmental portrait).

Modification of pixel noise during post processing is only required in the case where the crop or print is so large that the increased detail and noise power of the finer spatial frequency sampled by the smaller pixel is visible to the viewer's CoC, and furthermore that they find the additional detail and noise more objectionable than having it filtered through downsampling (or preferably something better) to the same level as the corser large pixel sampling rate.

In many common cases, the increased noise power at the higher spatial frequency will occur beyond the frequency of the viewer's CoC, or the increased detail that goes along with the noise power will be preferable. Either way, no difference in post processing would be required.

Furthermore, even if the burden of requiring additional post processing in those rare circumstances was seen as a negative for smaller pixels, one may simply use in-camera raw formats that do not allow the choice in that rare circumstance such as sRAW (though I do not recommend it).

The experiment I did above proves that noise in any given display or crop scales with sensor size, not pixel size. It shows that a large sensor with 6.4 micron pixels has far less noise than a small sensor with 6.4 micron pixels.

Daniel Browning wrote in post #9480484
That is never true for tonal levels predominated by photon shot noise at any given spatial frequency. Pixel sizes from 3 microns up to 15 microns have the same QE. Even in read noise dominated levels it is only a loose correlation; there are often cases where read noise scales with pixel size to result in the same read noise level. The 5D2 and 7D is just one such case: when sensor size and spatial frequency is equalized between them, they are the same.

I'd have to study some more before commenting on this.

Which it is.

Again, with the introduction of "noise per spatial frequency" concept, I can't comment until I know more; it may be the same thing as, or analogous to, the "per frame noise" issue raised.

I'm not sure if it does, or not. I suspect that once in the realm of "spatial frequency", we are in the realm of processed images, where averaging of pixel noise may have occurred.

How does it mislead? What false conclusions does it invite?

Other things like compression ratio, stroke, rpm, yadda, yadda being equal, OK, I'll accept that analogy, especially since you've cast doubt on the "larger photosites mean lower pixel noise" axiom.

Seems fair.

By larger, I take you to mean more cylinders.

Correct - whether the engine generates greater horsepower depends on the power generated by each cylinder, and the number.

How?

But there is! Without processing of the data, with the noise it contains, there is no image. With that processing, some noise averaging may occur, the amount depending on the degree of interpolation needed, with that depending on the sensor size and the print size.

Without processing, which potentially involves noise cancellation through interpolation, there is no image.

I'll have to chew on all that for a time.

To be honest, it wasn't very clear - I couldn't see any noise at all in either, so I had nothing to compare. I'll grant the truth of this assertion though, on the basis that I outlined. To scale the image from the sensor size to the display size required less magnification in the case of the FF image, thus a greater degree of noise cancellation due to interpolation.

Thanks for the great response.

xarqi wrote in post #9480562
Again, with the introduction of "noise per spatial frequency" concept, I can't comment until I know more; it may be the same thing as, or analogous to, the "per frame noise" issue raised.

Indeed it is. I should have mentioned that explicitly; you are perceptive. (One difference might be that "per spatial frequency" works equally well for any given crop out of the total image, whereas "per frame" carries connotations of using the whole image instead of crops).

xarqi wrote in post #9480562
I'm not sure if it does, or not. I suspect that once in the realm of "spatial frequency", we are in the realm of processed images, where averaging of pixel noise may have occurred.

Your suspicions are correct. I was wrong in how I interpreted your post, I'm sorry. Now I see now that you weren't referring to noise in an actual displayed image, just in the raw data file.

xarqi wrote in post #9480562
How does it mislead? What false conclusions does it invite?

Chiefly, the false conclusion that pixel size is important as it pertains to noise. (It is somewhat important, but only to the extent I hinted at in the first paragraph of my last post.) You may not agree, but here's why I think that's false: for a given sensor size, performance per area, and display size, there are a wide variety of pixel sizes that will result in the same noise level in the final displayed image. To me, that means pixel size is not important as it pertains to noise. On the other side, for a given pixel size and display size, larger sensors have less noise in the final displayed image. To me that means sensor size is important as it pertains to noise.

Second, to build up to the importance of pixel size, it invites the conclusion that noise per pixel in the raw data file is more important than the noise per spatial frequency in the final displayed image (in any form, such as a crop or print). At least, that is the impression I got from how things were phrased. In photography, the final displayed image is definitely far more important.

xarqi wrote in post #9480562
By larger, I take you to mean more cylinders.

Yes, that's what I meant.

xarqi wrote in post #9480562
How?

I think it is misleading by placing importance on things that don't matter. All that really matters is the total horsepower of the engine. Whether it's 16 small cylinders or 4 large ones isn't important except as a novelty -- what matters is what you can do with it. To emphasize cylinder size may cause a novice to think a small engine with large cylinders would have more horsepower than a large engine with small cylinders. The total engine horsepower is really what matters, and that is correlated with engine size, not cylinder size. Same with sensors.

xarqi wrote in post #9480562
But there is! Without processing of the data, with the noise it contains, there is no image. With that processing, some noise averaging may occur, the amount depending on the degree of interpolation needed, with that depending on the sensor size and the print size.

But as far as the photographer need be concerned, the amount of post processing is the same either way. They push the "print" button on their camera, it goes direct to the printer (that interpolates it) and the print comes out with all the benefits of the noise averaging. The software will certainly notice the difference between interpolating the 21 MP jpeg vs a 13 MP jpeg, but it's unimportant.

xarqi wrote in post #9480562
To be honest, it wasn't very clear - I couldn't see any noise at all in either, so I had nothing to compare.

If you download the raw files, you'll see that the ISO 1600 FF has more noise per pixel than the APS-C ISO 640, but in the displayed images the noise is the same. Something you already knew would be true -- I just didn't realize you knew it. Also, thanks for the feedback. It's a pretty fast and easy experiment, so I'll repeat it when I get a chance and make things clearer.

xarqi wrote in post #9480562
I'll grant the truth of this assertion though, on the basis that I outlined. To scale the image from the sensor size to the display size required less magnification in the case of the FF image, thus a greater degree of noise cancellation due to interpolation.

I see that we agree on the fundamentals -- it's only emphasis where we differ.

Phew!

xarqi wrote in post #9480710
Phew!

Uh, yeah. +1

xarqi wrote in post #9480710
Phew!

After reading some the MK-4's white papers that's what I said.
Canon, on their who's got the mostest MP quest, decreased the 4's pixel pitch. Warning, Will Robinson, watch that f/stop setting...

They did a real slick trick with the main LCD, but don't damage that glass cover screen as it will be right into that pretty looking eye candy LCD.
63 custom functions case you have any spare time, and more assisted focus points or whatever then are then you can shake a Nikon at.
Less battery life, but it can take images in starlight and ICE loves them because they don't need those expensive night vision devices any more. Does HD vids too.
It slices and dices... and cost a whole lot more.

My first reaction is, it's really complicated with ALL the bells, whistles, horns and ultrasonic sheep herding devices you could ever want for the next 1.2 years...
and gawd help you if you drop it (that never happens).

I want more coffee...


This is the twentieth century
But too much aggravation
This is the edge of insanity
I'm a twentieth century man but I don't wanna be here.

great technical info guys