Very interesting on whether camera resolution matters with a lousy lens

[Dan Wells]

I teach photography, and get a ton of gear questions. One that's always bugged me is "does a high-resolution body matter with any but the best lenses". i just ran across a fascinating test on Bob Atkins' website, which shows that body resolution DOES matter even with the worst lenses (I honestly had no idea how this one would come out until I read the article). The test is within his review of the Canon 5Ds, which is here:
http://www.bobatkins.com/photography/digital/canon_eos_5Ds_review.html
 
He first does what we're used to seeing - uses a world-class lens (Canon's 85 f1.2 L) at its best aperture (f4) and shows us the high-resolution 5Ds image as a 100% crop and the lower resolution 6D image, upscaled to the same size. NO Contest (nor would you expect one - if the 85 f1.2 at f4 couldn't take advantage of the resolution, what could)?

His second test is the one that I hadn't seen before - he went hunting for the WORST lens in his collection, and he did pretty well - how many of us own a piece of "glass" (a lot of this beauty may be plastic) worse than a Canon 22-55 kit lens from their APS film IX Lite camera? It seems to have been about a $100 lens even sold separately, probably $50 with a body, and it's a 15 year old plastic-mount special designed for a format smaller than the FF cameras he was using it on... Pretty satisfactory definition of a lousy lens? I honestly can't think of how to do worse without resorting to a DELIBERATELY distorting lens (a lensbaby, or a Holga lens somehow attached to a DSLR, or a pinhole)?

The two images show nearly as much difference as the two captured with the great lens. No, the sharp image is not as sharp, but the degree of improvement is NOT that different. I'm assuming his test conditions were decent, although he doesn't go into any detail, he's been around the Canon world a long time, and is well respected.

I've attached Bob's images below - I hope he doesn't mind the publicity for his interesting test. I can actually see the differences in the thumbnails on here easily enough.

No, this doesn't mean "all lenses outresolve all cameras", but, rather that low resolution is additive - even a less sharp lens will be improved by a sharper sensor.

Within the realm of "reasonable" combinations (nobody is actually going to use that 22-55 on a 5Ds, or a 1980s Nikon FM10 kit lens on a D810, except as a test), it seems like both body and lens matter. The high-pixel bodies will perform their best with fancy lenses, but they'll give any lens a boost.

I suspect that, unlike cellphone dynamic range showing up in tiny images (I can tell a phone picture from a great distance, due to blown highlights and blocked shadows), the resolution effects will still require decent scale. I wonder what that scale is - can our eyes perceive detail we aren't quite conscious of? Some photo connoisseurs claim to be able to tell the difference between an 8x10 contact print and an 8x10 enlarged from a 4x5 original. That's supposedly beyond the range of human vision (even enlarged 2x, film grain should still be comfortably microscopic). Are they seeing things, or is it the "pipe organ phenomenon"? A large pipe organ can play notes well below the range of human hearing, but you can tell that those notes are sounding. Similarly, sounds ABOVE the range of hearing have been shown to subtly affect our perception of music. Is detail in a photograph similar?

[b_rubenstein]

Hopefully, someone here will recall the formula for calculating system resolution from the resolution of the individual components. I do know it shows exactly what Bob describes.

(FWIW, Bob doesn't seem to have anything about himself on his site, but I we once worked at the same company so I know he was a Member of the Technical Staff with a PhD. He's not some guy on the internet.

[rdonson]

Aren't there a LOT of ways to "scale" an image?  Aren't some methods better than others?  Scaling to compare images always seems like a weak point in these kinds of analysis to me. 

[Hywel]

I agree, I think it depends on your usage as to what scaling method and what comparison is appropriate. Perceptual sharpness is a slippery mix of actual sharpness, micro contrast, purity of blacks and whites (a subtle flare can manifest as lack of sharpness rather than lack of contrast if you're not looking for it), etc. etc..

If you are producing huge prints, the method in the article is surely the right comparison to make.

If you're downsampling everything for the web, a more appropriate test would be to down-sample both cameras to some common smaller size using a sensible rescaling algorithm (which is what DxO do in some of their comparisons I believe).

My own highly unscientific experience bears out the conclusions of the article, even though a lot of my work ends up viewed at web size. I've been using the same physical trio of cheap-ish Canon prime 35, 50 and 85 mm lenses on my Canons since my first D30 and the oversampling clearly had huge benefits going from that through 5D Mark I and 5D Mark II. So I'm not surprised to see the benefits carry on to even higher oversampling, even with crappy lenses. 

The A7R II photos with these same lenses are better again... but not as good as with brand new Sony 55mm and 28mm lenses. So lens and sensor resolution both play a part. Presumably it is a game of diminishing returns eventually.

Cheers, Hywel

[Dan Wells]

Completely agreed that it depens on what you're using it for - some amount of oversampling matters relative to final output pixels, but Bob Atkins' test doesn't resolve how much. His question was whether, if 50MP was the ultimate desired resolution , there was any difference between a real 50 MP camera and an upsampled ~20 mp camera, both with good lenses and with lousy lenses? This seems pretty conclusive that you DO get a better 50 MP file by photographing one in the first place, rather than upsampling, regardless of lens. If your goal is, say, a 2 MP file for web use, oversampling with an 8 MP camera is clearly better than using a 2 MP camera (the oversampling reduces noise), but is massively oversampling with a 50 mp camera better than oversampling with a 20 mp camera? I don't know, and this test doesn't answer that question. My working hypothesis would be that, all else being equal (and a phone camera is NOT equal - I can spot their restricted dynamic range across a room, even in a tiny image), there is some fairly low limit to how much oversampling makes sense - nothing over 16 MP probably matters for even the highest resolution non-printed display, if that.

For printing, however, 50 MP is 240 ppi at 24x36", so it DOES make sense for large prints - just for kicks, modern Canon and Epson printers work internally with ~600 dpi, so an astonishing 310 mp is where you "run out of printer"for a 24x36" print. Could any human tell a true 310 MP 24x36" print from a 50 MP print (which is at the resolution of a Retina display)? If yes, what about a 100 MP print that is over 300 DPI?  Ansel Adams and some friends once tried to work out what our visual acuity actually supported based on viewing distance, and it doesn't suggest huge resolutions (but they continued right on making 8x10" contact prints, which probably work out to something alarming well over 1200 ppi).

[Telecaster]

I don't currently own any lousy lenses, though the sharpness über alles folks might argue with my definition of "lousy". I do notice that the Sony A7r2 records the character of various lenses just as all my other electronic cameras do (and in the past have done). The higher res is just higher res, not some magical phase transformation to a new realm of image making. 

-Dave-

[ErikKaffehr]

Hi,

I guess that the best way to describe it is to see thing like deterioration of information. Image is transferred over a chain of information carriers. Lens, sensor, postprocessing, display and human vision.

At each stage some information is lost and some disinformation is added. So each stage transfers a percentage of information, that is what we call MTF, and adds some disinformation what we call artefacts, aberrations or just noise.

The MTF part multiplies. So, lens perhaps transfers 60% and sensor transfers 60%, so total transmission would be 0.6 *0.6 = 0.36 that is 36 percent.

So improving sensor will improve information transfer. A better sensor may transfer 80% of the information, so with the very same lens we may transfer 0.6 * 0.8 = 0.48.

This essentially means that a significant improvement to any part of the imaging chain will improve the fidelity of the whole imaging chain. So the imaging chain is not dominated by it's weakest link, but is best described as a product of the quality of it's components.

Best regards
Erik

[gwhitf]

Sometimes I stitch two or three frames together with the canon 24mm tilt shift lens, thinking I'll gain resolution by the stitch. But the quality of it is so bad that I'd probably be better off just shooting it straight with a sigma Art lens, in one frame. Maybe it's just my copy of the 24. I don't know. But when at 100% in Photoshop, a bad lens is instantly obvious. Save all the time and trouble of stitching, by just shooting it with a good zeiss or good sigma.

[Slobodan Blagojevic]

... But the quality of it is so bad...

I'd love to see that.

[gwhitf]

I think part of the problem is that TS lenses are manual focus.  I think with digital, it requires Autofocus, or Live View approach. There is very little tolerance for mistake in manual focus. Maybe architectural guys would use it with Live View, but for me, way too slow.

[Petrus]

I teach photography, and get a ton of gear questions. One that's always bugged me is "does a high-resolution body matter with any but the best lenses". i just ran across a fascinating test on Bob Atkins' website, which shows that body resolution DOES matter even with the worst lenses

Of course it does. The throughput quality factor (?) is not the lesser quality component bottleneck, but the product of the quality factors or each component. Quality factor is 1 for a perfect component and 0 (zero) for a totally nonworking one.

Mediocre camera with bad lens: 0.6 x 0.2 = 0.12
Great camera with the same bad lens: 0.9 x 0.2 = 0.18

Simple as that.

With this method it is possible to estimate where to put the money when improving the throughput quality of a system: to the component where the component quality improvement is the biggest per dollar invested. For hi-fi systems this is quite clear; electronic components are practically perfect compared to even the best speakers. For cameras more sensor resolution is often better investment than marginally better lens which costs 4 times as much as a good one.

[bjanes]

Of course it does. The throughput quality factor (?) is not the lesser quality component bottleneck, but the product of the quality factors or each component. Quality factor is 1 for a perfect component and 0 (zero) for a totally nonworking one.

Mediocre camera with bad lens: 0.6 x 0.2 = 0.12
Great camera with the same bad lens: 0.9 x 0.2 = 0.18

Simple as that.

It is not that simple in practice. MTFs do multiply, but one must convert from the spatial domain to a frequency domain by first performing a Fourier transform, multiply the MTFs, and then convert back to the spatial domain as explained by Norman Koren here (scroll down to the Imaging Systems section). Not being an expert in these matters, I don't how one would do this in practice.

Bill

[Slobodan Blagojevic]

Of course resizing something from 20 Mpx to 50 Mpx is generally going to look like crap. Downsizing 50 Mpx to 20 Mpx, however, might seem like a better idea. After all, bigger sensors show their advantage when printed big, no doubt.

[Petrus]

It is not that simple in practice.

My example was not exact science, just the general principle.

[Hywel]

And you have to remember that noise behaves differently, too. You can never truly remove noise from the data once it is added; later stages of amplification and processing can at best amplify the noise already there along with the signal, and probably add some noise of their own.

From this point of view it is necessary to do the best job you can on the earliest stages of processing, because noise added there is there forever. That's why low noise preamps are so critical for sound recording (and data recording for scientific data in general).

There is a sense in which this is relevant to lenses and cameras, too. We don't normally think of a lens introducing noise, but it does if you consider that the desired "signal" is the photons at the sensor which come from the "correct" light path. MTF might go multiplicatively through the chain, but "noise" in this sense of light contamination from flare and internal reflections (and even in some sense physical effects like diffraction and aberrations) screws your signal up at source. Once it is there you can't get rid of it, your signal is degraded forever, and other parts of the processing chain are stuck with this noise.

Some stuff you might be able to deconvolve with sophisticated image processing (see Hubble telescope pre-fix, or the way that modern cameras correct for some aberration and distortion, even sometimes in the raw files) So it's not 100% accurate to characterise this in the same way as truly random sources of noise which are truly not separable from the signal.  In any case it'll never be as good as getting a clean signal in the first place- see Hubble telescope post-fix.

But it is useful to realise that not everything in the imaging chain is as simple as it might look and that the naive "these sensors are out-resolving the lenses so higher resolution doesn't help" calculations don't hold up to actual measurements. Tests are the only way to really know for your own use case.

Cheers, Hywel




 

[ErikKaffehr]

Hi,

Our computer screens used to be around 2MP. So 12, 24 or 50 MP is not visible on computer screen without pixel peeping.

Making prints, it is a different thing.

Epson printers have normally 360 PPI internal resolution which is a good match for what is considered normal human vision at 25 cm / 12" viewing distance.

So, having a 24 MP camera we could print something like  28x42 cm (11x16")  using the 360 PPI resolution. Anything larger, and it will be downsized anything smaller and it will be upsized.

With 42 MP, you could print 37x55 cm (14x22") at 360 PPI, that is a bit smaller than European A2.

So, to properly pixel peep, for A3 size print the images should be downsized to around 24 MP. While for A2 the images would be needed to upsized to something like 54 MP.

Now, an upscaled image will look ugly on screen as actual pixels on screen corresponds to a large magnification. But, that is exactly what is gong to sent to the printer.

In practical experience, it seems that 180PPI is quite enough for a good print. The reason is that our vision is not dominated by resolution but by contrast at lower frequencies. So good post processing makes it possible to achieve very good print quality from smaller formats, although they may lack ultimate detail.

An example of that was that I made some A2 (16 x 23") prints for an exhibition, one of those was from a 12 MP camera, 12 MP with good light beats 24 MP with boring light any day. That print looked really good, except it had a lot of dust spots. Looking at the image, I realised that those dust specks were birds. With 50 MP the print perhaps would show enough details to see that those were birds.

Structures, like mountain sides, still looked good in that A2 size print. As a general comment, when I have gone from 12 MP APS-C to 24 MP full frame I have shot some comparison images and printed at A2, at least in one case I couldn't tell the images apart, but in general the 24 MP images were a bit better. But the difference was far less than on screen. Going from 24 MP full frame to 39 MP on MFD I really found that I could not observe difference between 24MP and 39MP with the naked eye in A2-size prints. The difference was obvious when viewing the image with a loupe.

Best regards
Erik

Of course resizing something from 20 Mpx to 50 Mpx is generally going to look like crap. Downsizing 50 Mpx to 20 Mpx, however, might seem like a better idea. After all, bigger sensors show their advantage when printed big, no doubt.

[dwswager]

No, this doesn't mean "all lenses outresolve all cameras", but, rather that low resolution is additive - even a less sharp lens will be improved by a sharper sensor.

It is generally all additive from the lens out front to the output medium and including technique and lighting conditions. 

Looking at it from the pessimistic view point, there is a perfect image out there and all our technology and technique does is degrade it.  Of course, what we are striving for is the minimum amount of degradation throughout the chain!

[Bart_van_der_Wolf]

Making prints, it is a different thing.

Epson printers have normally 360 PPI internal resolution which is a good match for what is considered normal human vision at 25 cm / 12" viewing distance.

So, having a 24 MP camera we could print something like  28x42 cm (11x16")  using the 360 PPI resolution. Anything larger, and it will be downsized anything smaller and it will be upsized.

Hi Erik,

Yes, 300 - 360 PPI is a good match for human visual acuity at normal reading distances, at good illumination levels. However, the printer is actually printing at 600 PPI or 720 PPI, unless we reduce that to 300 PPI or 360 PPI by not selecting the proper settings. This may seem overkill, if we cant see it why would we do it, one could ask? Well, there can be subject matter that is resolvable by eye and that exceeds the average 20/20 vision acuity, calling on the Vernier resolution capability of our vision, AND with more (smaller) pixels to print we can use better output sharpening.

While I agree that, also due to larger viewing distance, we can print with quite a nice image quality at the lower PPI choices, there is a distinct benefit to using the higher resolution settings, even if that means we need to upsample the output before sending it to the printer. When one uses high quality upsampling software, e.g. PhotoZoom Pro, one can even increase resolution of edge detail beyond what the camera sensor has to offer.

There is also a chance that, on certain subject matter, what looks like aliasing is just bordering on aliasing but due to interference the detail looks crude/rough/pixelated, but once upsampled it looks correct because it was not really aliased.

In practical experience, it seems that 180PPI is quite enough for a good print. The reason is that our vision is not dominated by resolution but by contrast at lower frequencies. So good post processing makes it possible to achieve very good print quality from smaller formats, although they may lack ultimate detail.

In principle yes but, given the choice, who could object to better image quality also when viewed from shorter distances. The viewing experience becomes a lot more immersive, and material structure is much better depicted with more than barely adequate resolution.

And as to the OP's question, improving sensor resolution will always improve the captured lens quality, because denser sampling of the projected lens image will boost the overal MTF and increase limiting resolution. As others have said, component MTFs multiply to achieve an overall system MTF, and it's the worst contributing component that sets the achievable ceiling. Sensors are usually worse at resolving detail than decent lenses, so usually most gain is to be had by increasing sensor resolution (which also opens opportunities for signal restoration with dedicated software).

Cheers,
Bart

[NancyP]

Just what I need - more G.A.S. !
Atkins' result is really not too surprising. (Neither is the G.A.S.    )
I am still wondering about the TS-E 24 v. II, a long-time subject of my personal G.A.S, not yet tried by me. Wouldn't most people use Live View with such a lens? And even with Live view, they have problems with sharpness? Is it just a matter of not quite getting the tilt right? Or is it expecting too much out of the tilt, since anything out of the exact plane of critical sharpness is going to be "less sharp"?

[dwswager]

Trying to make a direct comparison of a dithering printer to a continuous tone device is a fools errand.  There is a 1-1 correspondence between image pixels and printed output with each pixel being any one of the total number of colors the continuous tone output medium has to offer. 

Epson (all inkjets) have a limited color set.  The hardware has the ability to place a certain number of drops per linear inch in each dimension of a specific size and specific color.  This is where the 360, 720, 1440 and 2880 numbers come from.  But there is no direct correspondence between a droplet and a pixel.  The apparent resolution of the image is significantly impacted by the substrate on which the ink is placed (how much ink can be laid down and how much spread of the drop occurs).

I have owned just about every Seiko Epson Stylus Photo desktop printer ever made.  I also worked on a government program with Epson.

Some Points directly:

1. Epson printers do not resample the image prior to passing through the dithering algorithm

2. The minimum resolution that should be sent to the printer reasonable quality output is 180ppi

3. The algorithm performance tends to top out at around 300-360ppi, except for images with specific characteristics (Epson declined to identify these characteristics) when higher pixel densities might yield marginal increases in apparent quality.

4. Only up sample an image if the up sampling provides a benefit to the image itself.  It is not beneficial to up sample an image merely to provide the dithering algorithm with more data.

5. Long time Epson users will verify that sending multiples of 60ppi is best and that, in general, 360ppi is a good standard.  Some will say that 720ppi is also reasonable.  What I will say is stay away from resolutions not quite a multiple of 60ppi.  For example, don't use 359ppi!  Better to be 330ppi than 355ppi for example.  This is because of how the algorithm itself handles the data prior to the dithering.

Hi,

Our computer screens used to be around 2MP. So 12, 24 or 50 MP is not visible on computer screen without pixel peeping.

Making prints, it is a different thing.

Epson printers have normally 360 PPI internal resolution which is a good match for what is considered normal human vision at 25 cm / 12" viewing distance.

So, having a 24 MP camera we could print something like  28x42 cm (11x16")  using the 360 PPI resolution. Anything larger, and it will be downsized anything smaller and it will be upsized.

With 42 MP, you could print 37x55 cm (14x22") at 360 PPI, that is a bit smaller than European A2.

So, to properly pixel peep, for A3 size print the images should be downsized to around 24 MP. While for A2 the images would be needed to upsized to something like 54 MP.

Now, an upscaled image will look ugly on screen as actual pixels on screen corresponds to a large magnification. But, that is exactly what is gong to sent to the printer.

In practical experience, it seems that 180PPI is quite enough for a good print. The reason is that our vision is not dominated by resolution but by contrast at lower frequencies. So good post processing makes it possible to achieve very good print quality from smaller formats, although they may lack ultimate detail.

An example of that was that I made some A2 (16 x 23") prints for an exhibition, one of those was from a 12 MP camera, 12 MP with good light beats 24 MP with boring light any day. That print looked really good, except it had a lot of dust spots. Looking at the image, I realised that those dust specks were birds. With 50 MP the print perhaps would show enough details to see that those were birds.

Structures, like mountain sides, still looked good in that A2 size print. As a general comment, when I have gone from 12 MP APS-C to 24 MP full frame I have shot some comparison images and printed at A2, at least in one case I couldn't tell the images apart, but in general the 24 MP images were a bit better. But the difference was far less than on screen. Going from 24 MP full frame to 39 MP on MFD I really found that I could not observe difference between 24MP and 39MP with the naked eye in A2-size prints. The difference was obvious when viewing the image with a loupe.

Best regards
Erik