M8 technical exploration, Luminous Landscape

[squareframe]

I didn't realize I had to consult Nyquist or Shannon to determine if an image is pleasing or not. reminds me of what I do almost daily ... I sail with the wind on my cheek. sure, I understand the physics, but I have little interest in that compared with the sensual feel of sailing a boat upon the wind. photography is much the same for me ... else I wouldn't photograph.

[Brad_]

The problem is, Bob, you don't have a technical background that enables you to

distinguish the difference between continuous-time analog systems, sampled by one's

vision, and discreet-time sampled data systems of a bandwidth-limited image aligned

(sampled) on a grid of pixels. That's evident from the article you referenced.

 

If you had such a background in sampled-data systems, you'd realize that perfect step-

function (brickwall) anti-alias filters do not exist. Rather, their attenuation has a slope,

with a corner frequency less than Nyquist (BTW, do you know who Nyquist is?). As a result,

specifying a resolution at Nyquist makes little sense. You can't have it both ways - either

you accept much less sensitivity, with a reasonable AA filter, or accept alias components

on the upper side of Nyquist manifesting as Moire with a filter starting to roll-off at

Nyquist (but still letting significant energy in on the other side of Nyquist). That is why in

digital systems specing resolution at Nyquist (half the sampling frequency) is faulty.

Specing at 1/3 of the sampling frequency (or 2/3 of Nyquist) makes much more sense.

Unless of course you like aliasing...

 

And I agree with Dan Taylor... Just hate seeing people without technical backgrounds

parroting information found on the net without regard to a real understanding. Would be

like me spouting all sorts of stuff on neurosurgey found on google and Gray's anatomy.

[steve_harris9]

The proof is in the print. Get a file of a photograph from the camera, you are considering buying, with the type of lens that you would use. Then print it (1) in the size that you normally print, on the printer that you will use. Then (2) make a print of the largest size that you would ever print -- something like a 36" x 24" on the latest epson. Do the same with a 2-1/4 x 2-1/4 scan of a negative or slide that (of yours) you like from an Imacon or better scanner. Then make the prints of that photograph. Better yet, spent the money and have a pro-printer make the prints. Then make your decision.

[blakley]

<i>The problem is, Bob, you don't have a technical background that enables you to distinguish the difference between continuous-time analog systems, sampled by one's vision, and discreet-time sampled data systems of a bandwidth-limited image aligned (sampled) on a grid of pixels. That's evident from the article you referenced.

<p>

If you had such a background in sampled-data systems, you'd realize that perfect step- function (brickwall) anti-alias filters do not exist.</i>

<p>

If that were indeed the problem, Brad, it would be very interesting, given that we're discussing a camera which does not have any anti-alias filter at all.

<p>

<i>Rather, their attenuation has a slope, with a corner frequency less than Nyquist (BTW, do you know who Nyquist is?). As a result, specifying a resolution at Nyquist makes little sense. You can't have it both ways - either you accept much less sensitivity, with a reasonable AA filter, or accept alias components on the upper side of Nyquist manifesting as Moire with a filter starting to roll-off at Nyquist (but still letting significant energy in on the other side of Nyquist).

<p>

That is why in digital systems specing resolution at Nyquist (half the sampling frequency) is faulty. Specing at 1/3 of the sampling frequency (or 2/3 of Nyquist) makes much more sense.</i>

<p>

And here's the real problem, Brad. You & Stan aren't really arguing with me. You're arguing with Erwin, and he knows this and gets it right (for example, he says <i>"Experience tells us that good definition should be found at values that are 0.7 or 0.5 of the theoretical Nyquist value: in this case around 40 to 50 lp/mm."</i>). So exactly what mistake do you think he's made here, Brad? The mistake of agreeing with you about something? Or the mistake of mentioning the Nyquist limit?

<p>

<i>And I agree with Dan Taylor... Just hate seeing people without technical backgrounds parroting information found on the net without regard to a real understanding. Would be like me spouting all sorts of stuff on neurosurgey found on google and Gray's anatomy.</i>

<p>

I also agree with Dan that looking at pictures is a much better way to evaluate camera performance than looking at graphs. That's why I've posted a variety of M8 photos here, in none of which is there a problem of any kind with resolution, as expected; <a href="http://www.photo.net/bboard/q-and-a-fetch-msg?msg_id=00J5En&tag=">here's a thread</a> with some of the latest. But I just hate seeing people with or without technical backgrounds falsely accusing real experts like Erwin of elementary mistakes which they are not making.

[Brad_]

Bob, you're in over your head and don't even know it.

 

All cameras with the exception of the M8 and old Kodak dSLRs (that I know of - there are

probably more) have AA filters. Thus, articles that go into sampling theory wrt digital

cameras assume that. The M8 is a special case without an AA filter, and specing

resolution at Nyquist is even MORE BIZARRE for the reasons I mentioned above. Unless, as

I pointed out earlier, you like Moire. <P>

 

<I>You're arguing with Erwin, and he knows this and gets it right (for example, he says

"Experience tells us that good definition should be found at values that are 0.7 or 0.5 of

the theoretical Nyquist value: in this case around 40 to 50 lp/mm.").</I><P>

 

What? Jeez Bob, try and stay consistent. I took exception with your comment that

Edmunds says resolution is measured at Nyquist (<i>As far as "generally accepted" is

concerned, authorities other than Erwin, including Edmund Optics...</i>). Which you kept

on defending. You, now quoting Erwin at 0.7 * Nyquist is essentially what I've been saying

all along; ie 1/3 the sampling frequency or 2/3 Nyquist. But I bet you didn't even realize

that. Too funny! See what happens when you don't have the underlying background to

make judegements on your own and can only provide links to things you don't fully

understand?<P>

 

<I>That's why I've posted a variety of M8 photos here, in none of which is there a problem

of any kind with resolution, as expected;</I><P>

 

Huh? Warning Will Robinson, straw man alert! No one here has said the M8 does not have

excellent reolution. I'm just taking exception to the bad science you're propagating!

[Brad_]

Just to make it really clear: 1/3 Fs = 2/3 Fn. 2/3 Fn = 0.6666 Fn, which is esentially 0.7

Fn.

 

Heh, heh...

[blakley]

Brad, we've had this problem before. If you're going to quote me, why not quote what I actually wrote? You say

<p>

<i>I took exception with your comment that Edmunds says resolution is measured at Nyquist (As far as "generally accepted" is concerned, authorities other than Erwin, including Edmund Optics...).

</i>

<p>

I did <i>not</i> write that "Edmunds says resolution is measured at Nyquist". What I <i>did</i> write is in fact exactly the opposite - reproduced here via cut-and-paste, which we know from previous threads is a favorite of yours:

<p>

<b>Stan, let's be fair to Erwin. He didn't say that the Leica's sensor's resolution is 73.5mm, he said it's Nyquist limit (maximum sampling frequency in pixels/mm divided by two) is 73.5 lp/mm.

<p>

As far as "generally accepted" is concerned, authorities other than Erwin, including Edmund Optics and our own Bob Atkins, define the sensor's maximum resolution as its Nyquist limit.</b>

<p>

That little word "maximum" is important. I <i>know</i> that contrast decreases and artifacting increases as you approach the Nyquist limit, and Erwin knows this too, and he didn't make any mistakes explaining it.

<p>

<i>You, now quoting Erwin at 0.7 * Nyquist is essentially what I've been saying all along; ie 1/3 the sampling frequency or 2/3 Nyquist. But I bet you didn't even realize that.</i>

<p>

You lose, and I can't possibly imagine why you're confused about this after I explicitly called attention to the fact that you were agreeing with Erwin (again via cut-and-paste): <b>So exactly what mistake do you think he's made here, Brad? The mistake of agreeing with you about something?</b>

<p>

<i>Too funny! See what happens when you don't have the underlying background to make judegements on your own and can only provide links to things you don't fully understand?</i>

<p>

Nice try, Brad. Wanna tell the crowd how fully you understand my technical background?

[blakley]

<i>Just to make it really clear: 1/3 Fs = 2/3 Fn. 2/3 Fn = 0.6666 Fn, which is esentially 0.7 Fn.</i>

<p>

Nice job. You multiply much more accurately than you read.

[Brad_]

<I>You lose, and I can't possibly imagine why you're confused about this after I explicitly

called attention to the fact that you were agreeing with Erwin (again via cut-and-paste): So

exactly what mistake do you think he's made here, Brad? The mistake of agreeing with you

about something?</i><P>

 

I lose?? You're talking crazy, Bob. I never said Erwin made a mistake. I'm *agreeing* with

him. But apparently you didn't even realize that.<P>

 

Too funy...

 

Again, back to your straw man, where was an issue made about the M8's resolution?

[Brad_]

<I>Nice job. You multiply much more accurately than you read.</I><P>

 

Just trying to help you out in an area you don't really understand.

[john_morris4]

<p><i>"Just trying to help you out in an area you don't really understand."</i></p>

<p>

The mathmatics that relate values in continuous time (or space, since we're talking about imagery) to sampled time (or space) are not particularly complicated, subtle, or hard to grasp. Many people still get these things wrong, and some also write wrongly about them. But Bob does seem to have taken the time to understand how sampling applies to continuous-space images.

</p>

[blakley]

<i>

Again, back to your straw man, where was an issue made about the M8's resolution?</i>

<p>

Right here, Brad (cut and paste again):

<p>

<i>Why would the resolution of the Leica's sensor be 73.5lp/mm?! It is only possible if one assumes that 1lp=2 pixels. It is generaly accepted that 1lp=3pixels. That bring resolution of the sensor to of less that 50/lp/mm. Consequently it will bring the resolution of the lens/sensor system to of about 50lp/mm. Reala ($2.5 a roll) resolves much better.</i>

<p>

Perhaps reading glasses would resolve this little problem you're having?

[Brad_]

Why the personal insults, Bob?

[blakley]

Which personal insult are you referring to, Brad? This one?

<p>

<i>

Just hate seeing people without technical backgrounds parroting information found on the net without regard to a real understanding</i>

<p>

Or maybe this one?

<p>

<i>

you're in over your head and don't even know it.</i>

<p>

Or was it this one?

<p>

<i>

See what happens when you don't have the underlying background to make judegements on your own and can only provide links to things you don't fully understand?</i>

<p>

Oh, wait, I know... it was this one:

<p>

<i>

You're talking crazy</i>

<p>

I'm not sure why, Brad.

[bobatkins]

The Nyquist limit is sampling at half the maximum spatial frequency for a periodic waveform. All this really means is that if you sample at half the maximum spatial frequency you can mathematically reconstruct the original waveform. If you sample less frequently you have abiguity (aliasing) in the reconstructed waveform.

 

Thus the Nyquist limit for the M8 is 73.5 lp/mm.

 

The Nyquist limit and the observed visual resolution will probably be different. The observed visual limit will be lower, depending on just how you define "visually resolved". In practice you'll probably resolve high contrast test patterns at around 65 lp/mm. Things start to look a bit nasty as you get very close to the Nyquist limit.

 

Interestingly enough this applies to all cameras, even Leicas. The presence of absence of an AA filter would make no difference if it were perfect. In practice an AA filter may lower the visual resolution limit by a couple of lp/mm, but will also significantly lower aliasing effects. You can't remove aliasing by software. That's what Nyquist tells you. You can "fudge" the signal to get rid of the nasty aliasing effects a bit, but mathematically you can't really differentiate between the true and aliased signals. The AA filter does remove the aliased signal (but not perfectly since there's no such thing as a perfect AA filter).

 

BTW those exchanging personal insults here will soon find themselves on a Xmas vacation from the forum. You have been warned.

[blakley]

Point taken, Bob. I'd welcome your edits.

[stan_belyaev]

1. Number pix per mm of that Kodak sensor is 1000/6.8=147pix/mm, so sampling frequency is 147pix/mm. N frequency is 50% of sampling frequency , so it is 73.5pix/mm (it MUST be expressed in the same units ? pix/mm). Resolution is expressed in line-pairs/mm. Let?s assume that 1lp = 2pixels, so we have to divide N frequency by 2. It gives us 36.75lp/mm. Actually it is quite good, even though the number looks small.

 

2. In order to utilize this resolution one should use a lens with a very high MTF at the same frequency (about 40lp/mm). Excellent MFT at lower frequencies or higher will not work. MFT should be as high as possible at about 40lp/mm. In that test two lenses were used Apo Summicron 75 and Canon zoom lens. Let?s look at MFTs of both of them. MFT of Summicron is about 75% and the line is almost flat. (f5.6, 40lp/mm). Canon?s MFT is about 80% at 30lp/mm (there is a steep slope) (30lp/mm is N frequency of the Canon?s sensor). But both lenses were used at the same focal length ? 75mm. So Canon?s 75mm is real (FF sensor), but Leica?s 75mm became 100mm (1.33crop). It is unfair. Moreover, Apo-Summicron is a very highly corrected lens (APO), so visible light of different wavelength is precisely focused (fall on the same pixel). Canon?s zoom, probably, is not as well corrected as Apo-Summicron. It is possible that B, R, and G lights may fall not on the same pixel but on neighbor pixels. It will affect the resolution of Canon?s camera.

 

3. Two vs 3 pix/mm. In a perfect world it should be 2, but in a real world a white or a black line can partially cover two neighbor pixel, so instead of one black and one white pixels we get 2 gray pixels. In this situation an extra pixel is used in order to calculate resolution (it is more likely that one of the three pixels will have a pure white or black color.

 

4. Object ? Probably, a constantly moving object (human) is not the best one for a resolution test

 

These are my thoughts why some of the articles are not very accurate. In other industries similarly conducted tests would be unacceptable. There are too many factors that may influence the outcome of digital camera tests, most of them should be equal (it is very difficult to achieve), otherwise the test should be considered inconclusive.

 

Please, correct me if my math is wrong.

 

I really wanted to buy m8, but after all these reports I decided to wait, instead I got MP-3 LHSA kit and Rollei FW 4.0. Absolutely beautiful cameras!

[blakley]

Stan, Nyquist frequency is already line pairs; you're dividing by two once too often for limiting resolution.

<p>

LHSA MP-3 is a great camera and a really nice choice; you'll get wonderful pictures. It was probably a good idea not to get the M8 until the fixes are implemented. But if you think resolution's a problem with the M8 you're just wrong. <a href="http://www.adamsoneditions.com/EDITIONS/EDITIONS_pubs.html">David Adamson</a> has been working with M8 files. He says <a href="http://www.leica-camera-user.com/digital-forum/9022-30-x-40-inch-m8-prints.html"></a> that the files are as good as scanned 4x5 film, and <a href="http://www.leica-camera-user.com/digital-forum/11267-m8-siberia-plus-new-acr-wb.html">here</a> that people who see his prints can't believe they're not looking at medium format.

[blakley]

(sorry, that should have read, "he says <a href="http://www.leica-camera-user.com/digital-forum/9022-30-x-40-inch-m8-prints.html">here</a> that the files are as good as scanned 4x5 film...")

[stan_belyaev]

N frequency should be expressed in the same units as a sampling frequency - l/mm or pixels/mm (not pixel-pair/mm or lp/mm). Pls, see the example from wikipedia.

In this example both sampling frequency and N frequency are exprssed in the SAME UNITS. (Hz)

http://en.wikipedia.org/wiki/Nyquist_frequency

 

 

40lp/mm is a lot. The size of the sensor is 27mm (horizontal). Human eye can not reslove more that 5-6lp/mm. One can magnify the image x6.5-8 and still get an excellent result! It means that one may get a nice print 19x13cm without noticeable loss of resolution/definition.

4x5 scan....- may be, it really depends upon ones taste. But it one is really interested in a high definition photo, then 19x13 is the max. size. There are very nice books on high definition black and white photography by Barry Thornton: "Elements" and "Edge of darkness". They were very nicely written, easy to read and they explain a lot! I strongly recommend these books.

 

I remember my first digital camera - Oly 2040 (2mpx). Back in 2001 it was a higher end camera. Many people stated that it was as good as 35mm. Well, I bought it, enjoyed sending pictures by email, but when it came to printing I couldn't print a decent photo of bigger than 4x6cm size (olympus p400 dye-sub printer $900 in 2001).

 

Unfortunately, somtimes we can not trick the numbers, so we have only two option: either to increase the numbers or to decrease our standards. Each of us makes his/her own choice.

[blakley]

There's something we can agree on, Stan: "Edge of Darkness" is indeed a fine book.

[stan_belyaev]

"Elements" is even better. It contains more picture, more explanations. Unfortunately it is out of print and very expensive ($500-$1500). I was fortunate to get an autographed copy from England for a reasonable price.

[stan_belyaev]

Actually that was a very interesting discussion. Each of the contributors and readers has learnd somthing new (i would guess).

Just to unwind, would you be interested to take a look at old soviet photos (published in "Soveiet Photo" magasine (I'm russian). Some of them. probably, weree taken with Leicas (popular among soviet reporters), so it should be withig the scope of this forum.

http://sov-photo.livejournal.com/

[blakley]

That's a really great link, Stan; I love <a href="http://sov-photo.livejournal.com/67827.html">this picture</a> especially. I've never even heard of "Elements" - if I run across it I'll certainly take a look.

[bobatkins]

Sampling theory is based on cycles/mm, not line pairs. However the usual convention is that one "line pair" (one black line and one white line) is equivalent to one cycle. A line pair isn't two black lines with a white line between them or two "black/white" sets.

 

It gets more complex because the cycles concerned really apply to sine waves and bar patterns are square waves (which is the sum of a sine wave and all its odd harmonics), so you can't just thow Nyquist at it and get an accurate result in "line pairs" when the lines are black and white. You can more correctly apply it to a target with sinusoidal modulation rather than square wave modulation, and such targets exist.

 

However all this applies only to a mathematical reconstruction of the original waveform from discrete samples.

 

For the purposes of photography you can, however, assume that one "line pair" equals one cycle, so you can say that the resolution of the M8 will be something less than 73.5 lp/mm. Around 65 lp/mm is probably a resonable practical estimate.