Take the highest MP count you can get....or

[Slingers]

I would think the opposite since the active photosite area is larger with no vignetting from the circuitary above such a one in a regular CMOS.

My understanding is that BSI increases pixel crosstalk which then also increases diffraction due to less light reaching the pixels with red, green and blue each getting different amounts of light.

It's possible I am wrong though as I just searched for reputable source to link but couldn't find it.

EDIT: after reading some more I was wrong as it seems it was the older bsi sensors that suffered from crosstalk. New technology like samsung's isocell or olympus' stacked bsi doesn't have the same problems.

 

[daf]

Should be pretty easy to make a real world test.
I'm sure some of You own both a7ii and a7rii:
Same lens, same focus point, same apertures 5,6/8/11/16.
Let's ses what come through..

 

[Vivek]

Craig, Yes, technology has improved enormously. There was even a significant jump from the NEX-5 to NEX-5N sensor (smaller pixels but much better results).

Of course, it OK to be talking about old/archaic stuff since the L brand is invoked.

 

[JoelM]

Well, how about cropping?

Yep, that was my thought exactly. Having a higher pixel density sensor will allow you to have a higher res crop. To me, that is a major reason for wanting more pixels.

Joel

 

[dmward]

The point the speaker made in the video is that its the um size of the photo site related to the resolving power of a lens.
His point, to quote; "its the physics"

He said that optimum photo site size is about 5 um and that large MP count 24x35mm sensors are getting well below that size as the MP count goes up.

Presuming that he knows what he's talking about its simple: to achieve maximum resolving power from a lens use a sensor with approximately 5 um photo site.

The more important test, in my view, is to shot something on a 5 um sensor with a lens, then shot the same scene on a 2 um sensor with the same lens, then print both of them, using the same printer to something like 40x60 inches and see which one is more impressive.

I haven't bothered trying to see where defraction begins to degrade images. I accept the common statement that its about F11. So, I tend to use that as minimum aperture.

I agree with Guy's inclination that high pixel count will mean I can print bigger if I want to and that's where its important. For most other applications we are throwing away most of the pixels to get an image down to a useful size.

 

[Bob]

from a physics point of view, it all depends on the resolving power of the lens.
Lets pick, for example, the Leica APO-Summicron-M 50mm f/2 as an example of a reasonably sharp lens with good resolving power.
Looking at its MTF graph you see at f/5.6 you see 80% contrast at 40 line pairs per mm.
To resolve a line pair you must sample it at least twice the periodicity of the signal according to Shannon sampling criteria. If it were to be sampled less frequently, then the pair would merge and would not be resolved.
So that means to resolve what that lens produces at 80% contrast one must have at least 80 sensor sites per mm.
1/80 yields a pitch of 12.5 micron.
Most real lenses don't have a sharp cut-off in resolving power, it is just that the contrast (or signal to noise ration) simply declines as the frequency is increased.
So I think, although it is not on Leica's chart, that that lens will provide at least 50% contrast at 80LP per mm. The lens will indeed pass higher frequencies, it is just that the contrast will drop as the frequency goes up. So what is your taste for the minimum contrast ratio that you would consider to be acceptable from a resolving power point of view?
The idustry has used 50% for a long time as an engineering benchmark.
So lets say we do and assume that this lens will produce 50% or better contrast at 80LP/mm
that means it needs a sensor pitch of 6.25 micron at the largest to give us enough information to resolve the signal.
it also means that anything larger will simply not give you what the lens is capable of producing.
Anything smaller will then be in the territory of over-sampling the signal which unfortunately may lead to other phenomenon none of which is particularly pretty.
So in other words, I tend to agree that unless lenses get a whole lot better, a pixel pitch smaller than say 5 microns or so will not give you more information. More pixels, yes, but no more information.
Thus assuming one wants to print large, even having more pixels than that will not actually give you more information on the print.
The simple reason is that is just not there unless you are satisfied with very low contrast in that information.
Of course limiting the sample frequency does produce the effect of a resolution cut-off filter, but unless the lens produces more information, more frequent sampling just won't deliver the goods.
So for a full frame 35mm sensor, anything more than 24Mpx is not theoretically beneficial.
If you really want more information, and assuming the lenses are up to it, what you need is a larger sensor, not more pixels per inch.
-bob

 

[k-hawinkler]

Well, considering a better lens one might come to a different conclusion, no?

 

[fotografz]

from a physics point of view, it all depends on the resolving power of the lens.
Lets pick, for example, the Leica APO-Summicron-M 50mm f/2 as an example of a reasonably sharp lens with good resolving power.
Looking at its MTF graph you see at f/5.6 you see 80% contrast at 40 line pairs per mm.
To resolve a line pair you must sample it at least twice the periodicity of the signal according to Shannon sampling criteria. If it were to be sampled less frequently, then the pair would merge and would not be resolved.
So that means to resolve what that lens produces at 80% contrast one must have at least 80 sensor sites per mm.
1/80 yields a pitch of 12.5 micron.
Most real lenses don't have a sharp cut-off in resolving power, it is just that the contrast (or signal to noise ration) simply declines as the frequency is increased.
So I think, although it is not on Leica's chart, that that lens will provide at least 50% contrast at 80LP per mm. The lens will indeed pass higher frequencies, it is just that the contrast will drop as the frequency goes up. So what is your taste for the minimum contrast ratio that you would consider to be acceptable from a resolving power point of view?
The idustry has used 50% for a long time as an engineering benchmark.
So lets say we do and assume that this lens will produce 50% or better contrast at 80LP/mm
that means it needs a sensor pitch of 6.25 micron at the largest to give us enough information to resolve the signal.
it also means that anything larger will simply not give you what the lens is capable of producing.
Anything smaller will then be in the territory of over-sampling the signal which unfortunately may lead to other phenomenon none of which is particularly pretty.
So in other words, I tend to agree that unless lenses get a whole lot better, a pixel pitch smaller than say 5 microns or so will not give you more information. More pixels, yes, but no more information.
Thus assuming one wants to print large, even having more pixels than that will not actually give you more information on the print.
The simple reason is that is just not there unless you are satisfied with very low contrast in that information.
Of course limiting the sample frequency does produce the effect of a resolution cut-off filter, but unless the lens produces more information, more frequent sampling just won't deliver the goods.
So for a full frame 35mm sensor, anything more than 24Mpx is not theoretically beneficial.
If you really want more information, and assuming the lenses are up to it, what you need is a larger sensor, not more pixels per inch.
-bob

Mystery solved.

In fact, a couple of mysteries solved.

 

[Bob]

Well, considering a better lens one might come to a different conclusion, no?

Sure,
Have an example?
-bob

 

[ErikKaffehr]

Hi,

Well, I have a great respect for Mr. Fossum but that presentation was quite a few years ago. Things develop quite a lot in five years. What happens is that design rules get smaller. A single junction (transistor) get smaller, that means that optimum resolution shifts. With 500 nm resolution, it may be that 8 MP is optimal, but with 180 nm technology used today 54 MP may be optimum. Modern sensors are also much simple designs, so sensor to wiring ratio has been much improved.

Now, getting back to diffraction. Diffraction is really benign to sharpening there is some empirical research indicating that say 36 MP stopped down to f/22 still beats 24 MP stopped down to f/8, once optimal sharpening is applied. Check here: https://www.onlandscape.co.uk/2012/08/the-diffraction-limit-how-small-is-too-small/

There is also empirical research showing that putting a mediocre zoom on a 36 MP DSLR beats the best of class zoom on a 21 MP camera.
LensRentals.com - Roger Buys a Camera System: A 24-70mm System Comparison

The final issue is that the resolution of the lens should not exceed the resolution of the sensor. Once this happens the sensor will show a lot of detail, all of which is fake. So it is better to have high resolution sensor that actually resolves the detail than a sensor that cannot resolve the detail and turns it in artefacts. OK, the correct image will look soft at actual pixels and the low resolution will show a lot of fake detail at reasonable contrast.

The simple fact is that high resolution is always good. A high resolution sensor will always give better detail with smoother contours and more malleable images, until pixels get that small that resolution is lost. There is a limit to resolution, but it is not at 12, 16 or 24 MP on a full frame sensor. Correct rendition with the best lenses today may require as much as 250-300 MP.

Best regards
Erik

Guy mentioned if two cameras were on a table he would take the one with the highest MP. He is not alone, for sure.

But's it's very interesting to hear what the inventor of CMOS sensors has to say on exactly this issue:

https://youtu.be/JkBh71zZKrM

This is a long lecture at Yale I think, but you can just forward to 38 minutes where he talks about the problems with small pixels and how marketing is trumping engineering and physics with high mp cameras like the A7r2. It's really worth a listen as he basically says: high MP counts "sell", so we make them, despite the fact they are not better because of the diffraction limit.

I did not get to this video as a way to "bash" the A7r2, which is obviously a very nice camera. I'm not mentioning the "L" word here at all. What got me here was another discussion about whether lens performance in general can vary at an aperture like f/11 or F/16.

In trying to understand this issue I came across articles which claimed the diffraction limit was f/13 on a 12mp FF but f/11 on a 24mp FF camera. Several of these stated outright, that unless you planned on shooting fast all the time, a larger pixel was more desirable because diffraction did not set in so soon and thus you could use f/11 and f/16 to full effect when you wanted DOF.

Now some think when you downsize a 42mp to 24mp the diffraction effects, which are greater at f/11, will magically disappear and the two images will be the same.

I don't really know, but I would like to find out, because I do shoot alot of landscapes and I do use f/11 often

I can't say I don't find it interesting the inventor of CMOS basically says outright: the MP race is a marketing scam. I'm not qualified to say to what degree it's true or not, but again I'd like to understand the issues better.

 

[Vivek]

So for a full frame 35mm sensor, anything more than 24Mpx is not theoretically beneficial.

If one takes that as true, the whole of m43rds, the top tier Canon, Nikon and Sony cameras can be summarily dismissed as ineffective.

 

[ErikKaffehr]

Hi,

I see a couple of problems with your reasoning.

1) It is quite true that 50% (or 35%) MTF is an often used criterium for perceived sharpness, but that reasoning excludes sharpening which is a part of the digital work flow. I guess no one with argue that it is good to achieve say 50% MTF at 180 PPI in print without sharpening. But it may matter little if it is native MTF from the lens or achieved by sharpening.

2) Now, consider two cases, say you make an A2 size print from a camera that delivers 50% MTF at 180 PPI and is limited to 180 PPI which would correspond to 16 MP. Now let's put the same lens on a camera with 64 MP. That camera would still produce at least 50% MTF at 180 MPI. Actually, it would produce quite a bit more as system MTF is MTF (lens) * MTF (sensor) and MTF of the 64MP sensor will be higher at 180 PPI than that of the 16 MP sensor.

3) The higher resolution sensor will still give a lot of detail once you go beyond A2 size.

I am pretty sure I have actually seen this in prints. I have shot 16 MP pictures which I felt were more usable than 24 MP pixels . Technically this was a newer generation APS-C sensor compared with older generation FF. This was under windy condition with both shutter speed, accurate focusing and shutter speed being an issue. The APS-C camera had live view, allowing more accurate focusing. With the APS-C camera I needed a shorter focal length so I could use a slightly better lens. The APS-C camera also had a cleaner high ISO, so I could have a shorter shutter time. Both images printed fine in A2 size.

On the other hand I have seen very little advantage from 39 MP MF over my APS-C sensor at A2 size. Going to A1 which is 41% larger (in linear dimension) I felt the 39 MP MF came to advantage.

So, what I see is that more megapixels actually yield a better image. But, you have to work for it. Also, it may be that the advantage of more pixels may be visible only in relatively large prints.

These are of course just reflections of what I have seen in my own work. My standard print size is A2, around 16"x23", with a few 30"x40".

Best regards
Erik

from a physics point of view, it all depends on the resolving power of the lens.
Lets pick, for example, the Leica APO-Summicron-M 50mm f/2 as an example of a reasonably sharp lens with good resolving power.
Looking at its MTF graph you see at f/5.6 you see 80% contrast at 40 line pairs per mm.
To resolve a line pair you must sample it at least twice the periodicity of the signal according to Shannon sampling criteria. If it were to be sampled less frequently, then the pair would merge and would not be resolved.
So that means to resolve what that lens produces at 80% contrast one must have at least 80 sensor sites per mm.
1/80 yields a pitch of 12.5 micron.
Most real lenses don't have a sharp cut-off in resolving power, it is just that the contrast (or signal to noise ration) simply declines as the frequency is increased.
So I think, although it is not on Leica's chart, that that lens will provide at least 50% contrast at 80LP per mm. The lens will indeed pass higher frequencies, it is just that the contrast will drop as the frequency goes up. So what is your taste for the minimum contrast ratio that you would consider to be acceptable from a resolving power point of view?
The idustry has used 50% for a long time as an engineering benchmark.
So lets say we do and assume that this lens will produce 50% or better contrast at 80LP/mm
that means it needs a sensor pitch of 6.25 micron at the largest to give us enough information to resolve the signal.
it also means that anything larger will simply not give you what the lens is capable of producing.
Anything smaller will then be in the territory of over-sampling the signal which unfortunately may lead to other phenomenon none of which is particularly pretty.
So in other words, I tend to agree that unless lenses get a whole lot better, a pixel pitch smaller than say 5 microns or so will not give you more information. More pixels, yes, but no more information.
Thus assuming one wants to print large, even having more pixels than that will not actually give you more information on the print.
The simple reason is that is just not there unless you are satisfied with very low contrast in that information.
Of course limiting the sample frequency does produce the effect of a resolution cut-off filter, but unless the lens produces more information, more frequent sampling just won't deliver the goods.
So for a full frame 35mm sensor, anything more than 24Mpx is not theoretically beneficial.
If you really want more information, and assuming the lenses are up to it, what you need is a larger sensor, not more pixels per inch.
-bob

 

[jlm]

i think it was Smokey Yunick of Nascar fame whose relevant quote was "there's no substitute for cubic inches" you can fit the metaphor to your own liking in this context...

 

[Jack]

Anything smaller will then be in the territory of over-sampling the signal which unfortunately may lead to other phenomenon none of which is particularly pretty.

Isn't one of the things it can do however, is render smoother tonality? When oversampling by a factor of N, the dynamic range increases by log2(N) bits, because there are N times as many possible values for the sum. (Hence one reason we typically get more DR as sensor resolution goes up past the 24MP point on 35mm sensors.) Of course noise goes up too, but raw algorithms have gotten very good at dealing with it (averaging). So while the data may be an interpolation and not "real," it gets implemented in a way most photographers (or audiophiles) pleasing -- the smoothing out of harsher point-to-point transitions between rapidly changing values. Analog film did this as well through halation (and analog tube-amp audio did it through latency).

 

[ErikKaffehr]

Hi,

Oversampling is always good. It is undersampling that is bad. Oversampling yields smooth countours, undersampling gives jagged contours just as an example.

Regarding noise it is mostly dependent on the number of photons collected. It doesn't matter if those photons are collected on an 8MP, 16 MP, 36MP or 42MP as long as similar technology is used. A larger sensor will collect more photons, there a full frame 24 MP sensor will have less noise than a 24 MP APS-C sensor. (*)

Regarding dynamic range it may decrease a bit with pixels size, according to the math at least, but that decrease has been met with better readout noise from modern sensors.

So smaller pixels are mostly advantageous. The exception is really high ISO capability where large pixels are at some advantage. (**)

Best regards
Erik

(*) This is not entirely true. With smaller pixels there will be more wiring so photodetector area to wiring area ratio will worsen. But design rules also shrink, so wiring area also gets thinner as sensor making technology develops. For each generation of technology there will be an optimum megapixel count.

(**) DR is normally defined as Full Well Capacity / Readout Noise. If we merge two pixels we double FWC but readout noise is the same. Now if we do it in software FWC will be the same as with the bigger pixel, but readout noise will be 41% (square root of 2) higher.

In the figures below you can see that DR increases a bit when Sensor+ kicks in on the IQ-180 but the noise curve is not affected. On screen there would be a dramatic change if we viewed the image at actual pixels, but activating sensor plus is like going from 100% view to 50% view. For those who don't know Sensor+ it is a Phase One technology doing pixel binning in hardware. So when Sensor+ is activated 4 pixels will be interconnected and form one pixel. Megapixels go down to 25% while linear resolution is halved. That binning is a Phase One invention and it avoids colour sampling errors. Quite smart.

Isn't one of the things it can do however, is render smoother tonality? When oversampling by a factor of N, the dynamic range increases by log2(N) bits, because there are N times as many possible values for the sum. (Hence one reason we typically get more DR as sensor resolution goes up past the 24MP point on 35mm sensors.) Of course noise goes up too, but raw algorithms have gotten very good at dealing with it (averaging). So while the data may be an interpolation and not "real," it gets implemented in a way most photographers (or audiophiles) pleasing -- the smoothing out of harsher point-to-point transitions between rapidly changing values. Analog film did this as well through halation (and analog tube-amp audio did it through latency).

 

[ErikKaffehr]

Hi,

I would agree, with some small remarks.

1) Having a slightly larger sensor with a similar pixel count makes the pixels larger. The larger size of pixels makes less demands on the lens. So the lens on the Sony may need to perform say 25-35% better than a lens on the 50 MP CMOS back to yield the same fine detail contrast.

The other point is that as both sensors use similar generation technology, so noise levels per sensor area will be about the same. But, the 50MP CMOS sensor being larger it will have it will have a slightly better Signal to Noise Ratio (SNR). Not a dramatic difference just corresponding to 2/3 EV exposure change.

2) Now, in a real world we probably have plenty of SNR with any modern imaging system at low ISO, assuming the sensor is not too small. So the real advantage of that larger sensor will show at high ISO. High ISO has not been the forte of MFD, but that has been due to the relatively noisy readout of the CCD sensor combined with low quantum efficiency. With the CMOS based MFD sensors this turns around. The Pentax 645Z is probably the best performing camera of all at high ISO. The other CMOS backs are probably as good in a real world just a bit restrained by reality

The 645Z should have a 2/3 stop advantage over say a Canon 5DIII at high ISO and it probably has. So you now can put your 85/1.4 Otus on the Pentax and have 2/3 stop advantage over the Canon 5DIII.

3) Just to explain, the Canon sensor is a different story. Canon sensors have noisy readout, so they cannot deliver the dynamic range of the Sony sensors at base ISO. But the signal coming of the sensor i actually quite clean, the purity is lost in the AD-converter. At high ISO the image is underexposed so amplification (analogue gain) is applied to the signal before AD-conversion. The camera makes good use of the sensor and the camera shines at high ISO. But, high ISO will not help with shot noise :-(

4) Now, why are Sony sensors that good at low ISO? They use a large number of simple but accurate ADCs situated on the sensor chip. So connections are short. There is an ADC for each column of pixels so each converter handles just a few thousand pixels.

Sony is not exactly alone with this technology. Leica has it on their CMOSIS designed sensors and so have Toshiba and Samsung, I think.

Digital sensors are probably among the hardest challenges in analogue microcircuit design. So the design of the pixels also matters a lot.

Best regards
Erik

The real difference will come down to tonal range and smoother transition because of the bigger sensor. Better said it looks smoother but that gap is getting smaller. They really are Sony sensors . I tested the leaf cmos back , same sensor as the Pentax and it's a great sensor but it's not really night and day, it's close.

 

[turtle]

I know its different, but this reminds me a little of the discussions on film formats years ago, where people cited maths and physics to argue that a 35mm Leica with the right film and lens would outperform 5x4" and therefore there's no point in using a big heavy LF camera. Unfortunately it tended to fall flat on its face in practice (not to mention the other factors at play, which impacted the look of the printed image).

I agree that the 50MP FF cameras seem to be nudging the practical limit for FF sensors and lenses to really deliver a meaningfully better print in normal use (i.e. regularly used apertures). However, tests show that at commonly used apertures the 5DSR does indeed resolve a little more detail than the D810, even though both have very small pixels (and where the difference in resolution is smaller than 24:42MP). At f16, I would agree that the difference may be gone or hard to see. Certainly the differences between the 5DS and 5DSR are already gone.

In my experience, 24MP and 36MP can be clearly distinguished in common use (f5.6-f11) and although the difference is not massive or guaranteed. Its more obvious with a very sharp lens at f4 (like the 55mm FE) than with a less sharp lens at f11.

In answer to what I regard the underlying question to be, 'can you really tell much of a difference between 42MP and 24MP in terms of resolution', the answer is 'yes'. As resolution increases, there are greater demands on technique and also aperture related limitations, but you really have to push all of that hard to the right to force 42MP down to the level of 24MP - far harder than one routinely encounters in normal use.

 

[ohnri]

As to the broader question of whether I would always take more MP ...

Took my A7s2 out for Haloween. Left my A7r2 behind.

I popped on the insanely great 35/1.4 FE.

I took many shots at over 50,000 ISO and quite a number at over 100,000 ISO. In some of these I still had to lift the shadows.

My Slideshow of images to friends and family was a big hit.

Mostly all color images, I only converted to B&W when the lighting was so garishly mixed that it was impossible to get all the faces looking decent without making a long term project out of it.

Could the A7r2 do the same thing? No idea really. I wanted easy. I got easy.

Is 12MP better than 42MP? No idea there either. I do know that the A7s2 does some things with less work than the A7r2. And it makes lovely files.

So, would I always choose more MP?

Nope. Apparently not.

-Bill

 

[Vivek]

Yeah, choice is good.

I have 12, 24, 36 and 42 to choose from. :grin:

 

[ErikKaffehr]

Hi,

I don't agree that diffraction is kicking in earlier with small pixels. I have made aperture series tests on almost all my cameras and in every case I achieved best result at f/5.6. These cameras were 6MP APS-C, 12MP APS-C, 24MP full frame and P45+.

What is true is the you loose more MP with small pixels than with large pixels when stopping down, the more you have the more you loose.

But, there is another axis, and that is sharpening. With good sharpening MTF can be regained but resolution cannot be regained.

Tim Parkin, the publisher of OnLandscape has looked into this in some detail, and he found that he could get better fine detail from a Nikon D800 he tested at f/22 than from his Sony Alpha 900 at f/8, with adequate sharpening.

I must say, I have not been able to reproduce his results with the gear I had that time.

Best regards
Erik

I don't think that the 6+ MP is what sells the camera. The slight IQ difference with respect to the A7r is rather due to the new technology used :

1) The back side illuminated sensor (BSI).
2) Then the on sensor PDAF, which is very efficient working in tandem with CDAF.
3) Add IBIS

and this body ticks all you need for a good picture (except those needing tracking AF for fast action). Personally I see too many pixels as a draw back for many because it eats storage space like mad and suppose higher processing resources. People are very aware of that.

Also diffraction kicking in earlier with higher pixels density is nothing new. It has already been abundantly discussed with the Nikon D800 and the A7r (F8 being judged the sweet spot, smaller than that diffraction begun to kick in). It has even been discussed for MFT sensors where you try to keep your lens at F5.6 or wider to avoid it.

One can use focus stacking to avoid the lack of DOF. Olympus for instance has announced they would implement that feature directly in the body with a firmware update for both the E-M1 and E-M5II. Rumors say that the E-M1II will get 20MB, which could explain why they are working on that feature.