which camera quiz

[arash_hazeghi]

Here is another brain storming quiz for gear lovers! :D
I taped this pretty beat up dollar bill on the wall, as a low contrast target. set my tripod 10m away and did two test shots with 500 f/4 + 1.4X TCII, aperture set to f/5.6 and remote release. One was done with 7D and one with 5D MKII. I dumped the RAW files in LR and processed them with default parameters. Obviously the bill is larger in 7D file, in order the equalize the outputs I used bi-cubic method in Photoshop to downsample 7D file to match 5DII side by side, the applied some mild sharpening to recover the details.

Here is the result, which one is 5DII and which one is 7D :D

[arash_hazeghi]

Here is how the FF looked like before downsampling 7D file





Let see who gets it right!!! :)

[Alfred Forns]

This is going to be interesting :)

[William Malacarne]

Arash

How much difference would it make if you placed the cameras at different distances to get the same field of view?

Bill

[arash_hazeghi]

Arash

How much difference would it make if you placed the cameras at different distances to get the same field of view?

Bill

If you move the MKII closer to equalize fov there is no contest, it will blow 7D out of water. But you can't always get closer to the birds ;)

[John Chardine]

You've done these before Arash. I got the wrong answer then but, undaunted, I will dive in. Nothing lost, nothing gained as they say.

The left hand image has slightly more detail than the one on the right, and it also appears to have a bit more chroma noise. Therefore, I am going to say 7D on the left, 5DII on right. So, megapixels trump pixel size in this particular test.

Of course this is all wrong- that should hedge my bets!

[arash_hazeghi]

You've done these before Arash. I got the wrong answer then but, undaunted, I will dive in. Nothing lost, nothing gained as they say.

The left hand image has slightly more detail than the one on the right, and it also appears to have a bit more chroma noise. Therefore, I am going to say 7D on the left, 5DII on right. So, megapixels trump pixel size in this particular test.

Of course this is all wrong- that should hedge my bets!

Let's wait for the answer, you might be right this time!

BTW, there is no chroma noise in either, there was some subtle ink residue on the bill as well some color moire so don't judge based on that.

Also did I do this before? can't remember!!!

[Alan Stankevitz]

Hi Arash,

I've done similar tests with both cameras, but cropping both images at 300 ppi with the 5DM2 being cropped to a higher degree to frame them equally. I found the 7D to be slightly better than the 5DM2 resolution-wise, with noise levels being very close. Just like your test, it was a focal-length-limited test.

In your test, I think the 7D looks like the photo on the left, but they are both close enough that you really can't tell them apart.

Alan

P.S. Say Alfred...Weren't you going to do a similar test between the 7D and Mark IV? I'd sure like to see how they compare when framed identically from a set distance.

[arash_hazeghi]

Alan,

Your findings might be different depending on the lens you used, whether you used a TC or not, whether your both bodies were micro-adjusted to perfection as well your post processing techniques. This particular test has a TC that is used wide open.

This was done at low ISO so there is no visible noise that would affect detail. also keep in mind that bicubic method is not very good for upsampling photos, so again the results will be different if you upsample MKII files.

Stay tuned for the answer.

Hi Arash,

I've done similar tests with both cameras, but cropping both images at 300 ppi with the 5DM2 being cropped to a higher degree to frame them equally. I found the 7D to be slightly better than the 5DM2 resolution-wise, with noise levels being very close. Just like your test, it was a focal-length-limited test.

In your test, I think the 7D looks like the photo on the left, but they are both close enough that you really can't tell them apart.

Alan

P.S. Say Alfred...Weren't you going to do a similar test between the 7D and Mark IV? I'd sure like to see how they compare when framed identically from a set distance.

[Alan Stankevitz]

I used a 600mm (no TC), adjusted using Live View with a PC, mirror lockup, etc...

[Alan Stankevitz]

Just curious Arash why you downsampled the 7D image? In a real-world scenario, would you do such a thing? For example, if a publisher wanted a cropped image of Washington's head at 300 ppi, wouldn't you crop the image as you did, but upsample instead of downsample?

In effect you are "dumbing down" the 7D's image resolution to match that of the 5DM2.

Alan

[arash_hazeghi]

Just curious Arash why you downsampled the 7D image? In a real-world scenario, would you do such a thing? For example, if a publisher wanted a cropped image of Washington's head at 300 ppi, wouldn't you crop the image as you did, but upsample instead of downsample?

In effect you are "dumbing down" the 7D's image resolution to match that of the 5DM2.

Alan

Alan, not sure if I agree. PPI is an arbitrary figure, I never upsample my photos, the up-sampled photos from a Bayer sensor will lack definition and do not meet my quality standards. If I am to print a photo for a book (which is actually what I am doing now) I make sure the bird is at least 2000 pixels wing to wing and the photo is at least 3000 pixels wide so I get a nice 8X10 with no up-sampling. Also for web representation such as the photos posted here you always downsample your photos.

Besides having more pixels on a target is only meaningful if the pixels resolve additional detail, if they don't it is just a waste of electrons.

[Robert Hardy]

i will play A is the 7d :)
Rob.

[arash_hazeghi]

OK, to make it more interesting, there is actually a subtle giveaway in the examples :D

[Alfred Forns]

... maybe not subtle enough :)

[Alan Stankevitz]

Alan, not sure if I agree. PPI is an arbitrary figure, I never upsample my photos, the up-sampled photos from a Bayer sensor will lack definition and do not meet my quality standards. If I am to print a photo for a book (which is actually what I am doing now) I make sure the bird is at least 2000 pixels wing to wing and the photo is at least 3000 pixels wide so I get a nice 8X10 with no up-sampling. Also for web representation such as the photos posted here you always downsample your photos.

Besides having more pixels on a target is only meaningful if the pixels resolve additional detail, if they don't it is just a waste of electrons.

Well, I guess it all depends on the publisher then. I am quite frequently asked for an image of a bird, x-inches by y-inches at 300ppi. They usually want some space surrounding the bird and I crop accordingly. If I upsize or downsize, it is dependent upon how much cropping the image requires. In most cases, the images are upsized except with the 7D.

For curiosity sake, I did a comparison a few months ago of the 5DM2, 50D and 7D and got different results than your comparison:

http://iwishicouldfly.com/iwishicoul...comparison.jpg

Alan

[arash_hazeghi]

Well, I guess it all depends on the publisher then. I am quite frequently asked for an image of a bird, x-inches by y-inches at 300ppi. They usually want some space surrounding the bird and I crop accordingly. If I upsize or downsize, it is dependent upon how much cropping the image requires. In most cases, the images are upsized except with the 7D.

For curiosity sake, I did a comparison a few months ago of the 5DM2, 50D and 7D and got different results than your comparison:

http://iwishicouldfly.com/iwishicoul...comparison.jpg

Alan

Alan you're saying the same thing again, not sure what's the point :eek:.

As I mentioned it will be different if you up-sample the photo and use a naked prime lens as you have done here. That's not what I am doing here. It is very simple math, when you upsample a photo you are creating data that did not existed by extrapolation to fill in the gaps, especially with bicubic method in Photoshop it will result in murky photos.

Also in above there is a very big difference between 50D and 7D which differ in pixel density by only 10%, this means 50D was not correctly focused.

good luck with your work.

[arash_hazeghi]

... maybe not subtle enough :)

Lol, I actually mentioned it once ;)

[arash_hazeghi]

So the answer is:
ADMKII B: 7D :D

The giveaway was the color moiré artifacts in the left one which John saw, this hints that A is a 100% crop and B was downsampled.

The key point here was the use of TC which is often overlooked. Smaller pixels will sample a higher spatial frequency but only if those frequencies are available at first place. An optical system has a limited bandwidth of the spatial frequencies it can transmit. The Nyquist limits for 7D and 5DMKII are 116 lpm and 78lpm respectively; of course the absolute resolution is less due to the optical low-pass filter that is on top of the sensor. A naked Canon 500 f/4 when focused correctly will provide about 150 lpm resolution at the center, which is pretty good, and will out resolve both sensors. However when you add the TC you lose a stop of light, reducing resolution to about 75 lpm by MTF50 conventions. This will hit the cut off frequency of the AA filter on the 5DMKII creating some moiré artifacts. 7D pixels will still sample at close to 116 lpm but since these tones are greatly attenuated in the projected image it will not resolve much extra meaningful detail. It will be stuck at 75 lpm just like the MKII. Throw in the MKIIs higher SNR pixels with better MLA and it will deliver a crisper image. By down sampling 7D file you can recover SNR a bit (in a lousy way) and by sharpening you can further improve it to barely match the 1:1 output of MKII which is what the above is.

The point is 7D’s tight pixels have pushed the envelope of Canon telephoto optics (which are btw at least 7 years old now and do not feature the newly engineered sub-wavelength coatings that are offered in more recent lenses) to the point that it has reached the limits of diminishing returns when adding a TC. I like to think about 7D as having a built-in TC and use my lenses naked.

[Alan Stankevitz]

Also in above there is a very big difference between 50D and 7D which differ in pixel density by only 10%, this means 50D was not correctly focused.

good luck with your work.

Disagree. All cameras were focused accurately using mirror lock, manual focus and tethered to a PC for visual accuracy. You cannot base a camera's resolving abilities strictly on the sensors pixel density. AA filters also play a role in total image quality/resolution.

Alan

[Alan Stankevitz]

I still don't understand what this comparison is suppose to prove when your test is dumbing-down the 7D to match the resolution of the 5D when equally cropped. The test takes away extra pixels that the 7D has available which decreases the resolution of the 7D's image.

[arash_hazeghi]

Disagree. All cameras were focused accurately using mirror lock, manual focus and tethered to a PC for visual accuracy. You cannot base a camera's resolving abilities strictly on the sensors pixel density. AA filters also play a role in total image quality/resolution.

Alan

I had a 50D too before I got the 7, its AA filter is a bit stronger than 7D but that is barely a 5% visual difference. in your example the difference is more like a whopping factor of 2, with the 50 being almost like MKII, numbers just don't add up...

[arash_hazeghi]

I still don't understand what this comparison is suppose to prove when your test is dumbing-down the 7D to match the resolution of the 5D when equally cropped. The test takes away extra pixels that the 7D has available which decreases the resolution of the 7D's image.

Alan you are just repeating your argument, not sure what the word "dumbing-down" means, I already explained that 7D file does not contain higher spatial frequencies, no matter how many extra pixels it has it would be the same even if it had one extra GIGA pixels, it is simple physics and math, can't explain any better-sorry :)

[Alan Stankevitz]

Arash,

Check out this example. In this example, the 5DM2's image has been upsized to equal that of the 7D. Yes, I know this adds extra pixels, but in this case you are not taking away pixels from the 7D's image:



Now take a look at 100% view of each:

First the 5DM2:



Now the 7D:


You can clearly see that the 7D has better resolution than the 5DM2. The 7D has a higher pixel density than the 5DM2 and this is what one would expect to see.

Alan

[arash_hazeghi]

Alan,

I am really surprised :eek:

First these two samples both look soft to me and lack of definition, because you have up-sampled both of them, I would reject both of them, they are both deletes in my book so it doesn't matter which one looks better at first place.

This is my idea of a sharp 100% crop, no processing or sharpening, straight out of camera:



plus

1) how many times did I mention that I used a TC wide open? that was the whole point you missed. Did you use a TC? what was your aperture with the TC? I did mention that with a naked prime 7D DOES have resolution advantage.

2) How many times did I mention that if I up sample results would be different?

The reason for this is you are filling the gabs between MKII samples from any given tone in the spectrum of the image with poorly interpolated data in Photoshop.

If you want to upsample try Spline method not bicubic.


Pixel density does NOT mean higher resolution when the system is limited by optics, when you have 50 lpm optical resolution it doesn't matter if you sample it at 200 lpm or 400 lpm ,simple math you can't beat, at least in this world.

If pixel density equates resolution all of us should just go buy a compact 15 mpixel camera with super high pixel density and a 600mm mega zoom lens because that will produce better poorly up-sampled results. If your idea of resolution is murky upsampled photos using bicubic method I believe that is the best choice, some of them even have "digital zoom" which will upsample and crop the photo inside the camera for you, so you can get your tightly cropped 300dpi billboards right away :D ;)

Good luck with your work, I agree 7D is a very good camera if used within its limits and properly, I look forward to seeing your photos in the avian forum :)

[Alfred Forns]

.... I picked the wrong one !!!! :)

[arash_hazeghi]

.... I picked the wrong one !!!! :)

haha you shouldn't Al cause you guys have both 7D and 5DMKII! :)

BTW, was shooting in the rain the other day with MKII dropped in the mud and scratched the base, still rocks!!! I am using it a LOT more since you and Fabs mentioned AF for flight was good (100% true)! just discovered its potentials!!! Got my best harrier dive shot the first day I tried it with 500! wish it was faster :(

Thanks!!!

[Alan Stankevitz]

Arash,

If I find any spare time tomorrow, I will take a few images with my 600mm and a 1.4TC with both a 7D and 5D2. I seriously doubt that the images will look any different. I have yet to find any camera that out-resolves my lenses including the use of a 1.4TC.

Here are the un-upsized or down-sized images that I took this evening:

5DM2:


7D:


As you can see, without any upsizing or downsizing, these 100% images do show the 7D has the ability to resolve to a higher level than a 5DM2.

But of course, my test is invalid because I did not use a 1.4x TC.

Alan

[arash_hazeghi]

Big improvement over previous post Alan!!!

now try with the TC and wide open aperture. When you add the TC the lpm resolution of the combo will drop by a factor of 2, it is the laws of optics not something you and I decide on. I do not have a 600mm f/4 lens but I find it hard to believe it has a resolution of 250 lpm in order for it to out-resolve the 7d with the TC.

Besides, looking at your own samples, I can only see one area of the image where 7D is delivering just slightly more visible detail and that is the text, the difference is barely 15% to my eyes. How in the world you are calling this a big difference? :eek: The fine lines are still separated in both samples. Now throw in some noise and low contrast detail to the equation and this little difference you see will disappear in a heartbeat. In real field conditions such a small difference becomes irrelevant due to vibrations, tolerance in AF system, atmospheric conditions etc.

7D has 2.2X the pixel density of the 5DMKII, the 5DMKII also has 2X the density of 5DMKI but the pixels in the latter two are much larger, if you have the 5D do the same test side by side with the 5D MKII and see how dramatic the difference is. Resolution will nicely scale in that case because there is no limitation from the optics. that is what I call a meaningful difference for field purposes. The sensor is only 50% of the equation the other 50% is the optics.

Arash,

If I find any spare time tomorrow, I will take a few images with my 600mm and a 1.4TC with both a 7D and 5D2. I seriously doubt that the images will look any different. I have yet to find any camera that out-resolves my lenses including the use of a 1.4TC.

Here are the un-upsized or down-sized images that I took this evening:



As you can see, without any upsizing or downsizing, these 100% images do show the 7D has the ability to resolve to a higher level than a 5DM2.

But of course, my test is invalid because I did not use a 1.4x TC.

Alan

[Alan Stankevitz]

Arash,

I never used the words "big difference"... re-read the entire thread. I believe those are your words? I merely stated the 7D has better resolution than the 5DM2.

Alan

[Roger Clark]

Smaller pixels will sample a higher spatial frequency but only if those frequencies are available at first place. An optical system has a limited bandwidth of the spatial frequencies it can transmit. The Nyquist limits for 7D and 5DMKII are 116 lpm and 78lpm respectively; of course the absolute resolution is less due to the optical low-pass filter that is on top of the sensor. A naked Canon 500 f/4 when focused correctly will provide about 150 lpm resolution at the center, which is pretty good, and will out resolve both sensors. However when you add the TC you lose a stop of light, reducing resolution to about 75 lpm by MTF50 conventions. This will hit the cut off frequency of the AA filter on the 5DMKII creating some moiré artifacts. 7D pixels will still sample at close to 116 lpm but since these tones are greatly attenuated in the projected image it will not resolve much extra meaningful detail. It will be stuck at 75 lpm just like the MKII. Throw in the MKIIs higher SNR pixels with better MLA and it will deliver a crisper image. By down sampling 7D file you can recover SNR a bit (in a lousy way) and by sharpening you can further improve it to barely match the 1:1 output of MKII which is what the above is.

The point is 7D’s tight pixels have pushed the envelope of Canon telephoto optics (which are btw at least 7 years old now and do not feature the newly engineered sub-wavelength coatings that are offered in more recent lenses) to the point that it has reached the limits of diminishing returns when adding a TC. I like to think about 7D as having a built-in TC and use my lenses naked.

A few things that might clarify this thread for people. Things are not always so black and white. First there isn't a simple sharp cutoff of line pairs per mm (lpm), and there is still perceived detail beyond MTF50. There is detail that people do perceive in images close to MTF0 (that's 0% contrast for those not familiar with MTF). Next, while putting on a 1.4x TC does reduce light by a stop, the magnification is only 1.4x so the resolution loss due to increased diffraction is only 1.4x. Depending on the quality of the TC, the loss could be a little greater, but beyond that, it will not significantly change the angular resolution of the main lens, which is what we are concerned with. In general if the use of a 1.4x TC meant a 2x loss in resolution, most here would not use TCs. Clearly this is not the case.

So now let's look at the 7D case. At the 4.3 micron pixel spacing, MTF (contrast) is down to about 20% (about 35% at f/4) pixel to pixel with perfect optics (diffraction limited). The blur filter will reduce that a a little more. But as has been discussed on BPN with macro photography working at f/16 and slower with some cameras, the mere fact that there is a hint of contrast means that the contrast can be boosted. and some of the image quality restored using tools like unsharp mask. If one uses more sophisticated tools, like Richardson-Lucy image deconvolution, more can be recovered. The cost is generally more noise, and of course if pushed too far, one gets algorithm artifacts.

So, I find the premise of this test/comparison flawed and the specifics of the reasons the test came out this way did incorrect.

I feel your tone in your response earlier this evening to Alan was a little harsh. Please try and tone it down.

Some notes on that response:

If you want to upsample try Spline method not bicubic.

Note bicubic is a cubic spline in 2 dimensions. Every resampling algorithm in my experience does well on some subjects and not so well on others. Cubic spline generally does a good job.

Pixel density does NOT mean higher resolution when the system is limited by optics, when you have 50 lpm optical resolution it doesn't matter if you sample it at 200 lpm or 400 lpm ,simple math you can't beat, at least in this world.

And too simple math is not the answer to refute what others are trying to tell you. Lenses don't deliver, for example, 50 lpm and lower. It is a slow fall-off. You should quote MTF 0% if you want to talk limited by optics. And note it is also possible to detect isolated subjects that are much smaller than the diffraction disk; such subjects also impact image perception.

It is the basic fact that lenses roll off slowly in MTF why cameras like the 7D can squeeze out a little more resolution, and do it with a 1.4x TC, and why citing single MTF values is incomplete.

Roger

[Alan Stankevitz]

This is my final post on the subject.

The two images are 100% crops taken with the Canon 600mm., f/4 IS lens + 1.4x TC (800mm @ f/5.6) taken with the 5DM2 and 7D from the same distance and cropped the same without any interpolation. I think everyone will be able to judge if there is more resolution or not available from the 7D when compared with the 5DM2.

I apologize for how long and convoluted this thread has become, but when statements are made that I do not agree with, I must defend my own findings.

5DM2:


7D:


I would also like to apologize to Arash or anyone else if I came off being offensive. I was just trying to make a point regarding the comparison of resolution between these two cameras with good glass and with or without a teleconverter. I will not make any more comments about this.

In the words of Chevy Chase, Good Night and have a pleasant tomorrow.

Alan

[arash_hazeghi]

Alan.,

It's fine if you want to stop here, but these samples are blown up and washed out, the 5DMKII has completely lost its sharpness, in fact looks worse than w/o TC!!! not sure what happened here :eek: ;) above is your MKII sample w/o TC below is with the TC :eek:

But I agree let's leave it for the viewers to decide if there is a difference or which one is better.
Good luck :)

5DM2:


Alan

[Roger Clark]

Alan,
Well done. You've illustrated the point very well. The 7D does show more detail.

Roger

[arash_hazeghi]

This is my final post too, unless Alan wishes to continue. Given the past history I am not going to respond to some other comments.

Alan you were not offensive at all and I enjoyed the discussion, I sometimes have a critical tone hope you did not take anything personally, if you did I apologize :) Have fun with the cameras.

Thank you all for participating in the poll.

I am also providing 100% crops for both in case you want to examine them, I didn't use flash and WB was off. same parameters Canon DPP 3.8 no additional sharpening or NR.

[Dan Avelon]

Nice discussion here,

Alan,

I checked the EXIF for samples #28 and you used f/4 for 5D and f/10 for 7D, why is that? In those samples I don't really see any extera detail in the 7D shot but my eyes might be lousy.
Also I agree with Arash that TC samples are washed out, and in particular the 5DMKII sample looks worse than without TC, if you are saying that even the 7D resolves more detail with the TC, why should the MKII look worse then :confused:

When I got my 5DMKII I saw a big difference in detail compared to classic 5D but the advantge here seems negligeble ( to my eyes) , I also think MTF50 is a good way to quantitively compare the sharpness, which is why it is industry standard.

Does anyone have the charts for 500 and 600 lenses with TC? I searched the internet and didn't find anything.




Dan

[sergey kuplin]

Roger,
I very much enjoy reading your posts as they sometimes introduce new perspectives and topics for discussion. I understand this is not a signal processing forum but since you too seem to be interested in technical details, I figured I might throw in a couple of ideas.
Firstly, human perception is most closely modeled by a logarithmic scale, not linear. I have personally never seen a reporting of perceived detail even close to MTF0. That would mean –infinity on an MTF chart!!! (For those unfamiliar with the logarithmic scale, the distance between 1 and 10 is the same as that between 0.1 and 1 or 0.01 and 0.1 on a logarithmic scale, so it would take till infinity to reach zero!). It is very hard to imagine the perceived effects of 0% contrast when other effects in real life situations easily overshadows even 10% contrasts. I would sure like to see an independent scientific experiment using the MTF0 data!
Secondly, bicubic interpolation is not the same as spline interpolation in 2D. Of course any interpolation based on a cubic polynomial in 2D is a bicubic interpolation (and hence splines are a subset) but the subtle (and computationally immense) difference that distinguishes splines from most other types of bicubic interp. is that bicubic uses the continuity of a function and its derivatives to fit a polynomial and requires the solution of a very large set of coupled linear equations which however efficient on modern computers, is far from ideal in this case. Splines and in particular B-Splines with limited support are different in the fact that they are constructed by the n-th order convolution of a kernel with itself which are then easily and efficiently convolved with the image and thus do not require the solution of any equation (the diagonalization of matrices and etc.). While this may all seem too technical, it is extremely important it is the key reason for the widespread use of splines in commercial software. As such, in DSP literature, splines are usually designated just “splines” and simple cubic interp. In 2D is called “bicubic” with no ambiguity whatsoever. I suggest reading the papers by Michael Unser to get more familiar with B-Splines if you are interested.
Thirdly, I was surprised to see a mention of the Richardson-Lucy method!!! As you probably know, the use of this method requires a priori knowledge on the system function (sometimes referred to as the point spread in this case) of the entire camera system which is highly unlikely to be known in real life conditions i.e. when the shot is taken (you know the camera system is not entirely LTI). And as is the case with most iterative algorithms, if you’re going in the wrong direction, further iterations would only make things worse. of course Richardson-Lucy is a rather simple method and more sophisticated ML and MAP methods can be used for data recovery, but I doubt it that any photographer would use such methods to salvage detail and I think mentioning this would be going off topic.
Lastly, In regard to TCs, this has always been an issue for me and I didn’t quite understand your argument. Are you suggesting that in non-diffraction-limited cases, the use of TC should not significantly degrade the image quality? I have been using TCs for quite a while but I am used to getting significant image quality degradation with TCs (and I can assure you the focus is spot on and there is no camera shake but I don’t have my archive on my laptop here.)


Cheers,

sergey

A few things that might clarify this thread for people. Things are not always so black and white. First there isn't a simple sharp cutoff of line pairs per mm (lpm), and there is still perceived detail beyond MTF50. There is detail that people do perceive in images close to MTF0 (that's 0% contrast for those not familiar with MTF). Next, while putting on a 1.4x TC does reduce light by a stop, the magnification is only 1.4x so the resolution loss due to increased diffraction is only 1.4x. Depending on the quality of the TC, the loss could be a little greater, but beyond that, it will not significantly change the angular resolution of the main lens, which is what we are concerned with. In general if the use of a 1.4x TC meant a 2x loss in resolution, most here would not use TCs. Clearly this is not the case.

So now let's look at the 7D case. At the 4.3 micron pixel spacing, MTF (contrast) is down to about 20% (about 35% at f/4) pixel to pixel with perfect optics (diffraction limited). The blur filter will reduce that a a little more. But as has been discussed on BPN with macro photography working at f/16 and slower with some cameras, the mere fact that there is a hint of contrast means that the contrast can be boosted. and some of the image quality restored using tools like unsharp mask. If one uses more sophisticated tools, like Richardson-Lucy image deconvolution, more can be recovered. The cost is generally more noise, and of course if pushed too far, one gets algorithm artifacts.

So, I find the premise of this test/comparison flawed and the specifics of the reasons the test came out this way did incorrect.

I feel your tone in your response earlier this evening to Alan was a little harsh. Please try and tone it down.

Some notes on that response:


Note bicubic is a cubic spline in 2 dimensions. Every resampling algorithm in my experience does well on some subjects and not so well on others. Cubic spline generally does a good job.



And too simple math is not the answer to refute what others are trying to tell you. Lenses don't deliver, for example, 50 lpm and lower. It is a slow fall-off. You should quote MTF 0% if you want to talk limited by optics. And note it is also possible to detect isolated subjects that are much smaller than the diffraction disk; such subjects also impact image perception.

It is the basic fact that lenses roll off slowly in MTF why cameras like the 7D can squeeze out a little more resolution, and do it with a 1.4x TC, and why citing single MTF values is incomplete.

Roger

[Alan Stankevitz]

Nice discussion here,

Alan,

I checked the EXIF for samples #28 and you used f/4 for 5D and f/10 for 7D, why is that? In those samples I don't really see any extera detail in the 7D shot but my eyes might be lousy....

Dan

Hi Dan,

Well...I stated I wasn't going to post any more on the subject, but since you are asking a question, I will respond.

It was 1/10th of a second, not f/10. Please recheck your EXIF viewer. All comparisons of the two sets were taken wide open. Either with or without a TC. (Either f/4 or f/5.6)

Alan

[Roger Clark]

Roger,
I very much enjoy reading your posts as they sometimes introduce new perspectives and topics for discussion. I understand this is not a signal processing forum but since you too seem to be interested in technical details, I figured I might throw in a couple of ideas.
Firstly, human perception is most closely modeled by a logarithmic scale, not linear. I have personally never seen a reporting of perceived detail even close to MTF0. That would mean –infinity on an MTF chart!!! (For those unfamiliar with the logarithmic scale, the distance between 1 and 10 is the same as that between 0.1 and 1 or 0.01 and 0.1 on a logarithmic scale, so it would take till infinity to reach zero!). It is very hard to imagine the perceived effects of 0% contrast when other effects in real life situations easily overshadows even 10% contrasts. I would sure like to see an independent scientific experiment using the MTF0 data!

Sergey,
MTF charts do not necessarily need to be plotted on a log scale. MTF does go to zero at the diffraction limit. See this web page which plots MTF on a linear scale:
http://www.cambridgeincolour.com/tut...resolution.htm

The look at Norman Koren's MTF page: http://www.normankoren.com/Tutorials/MTF.html
He shows modulation images down to 2%. Put one of those into photoshop and boost the contrast with an unsharp mask. Or have more fun with Richardson-Lucy.

Another example: scanning film. This 4x5 velvia transparency was scanned on a drum scanner:
http://www.clarkvision.com/galleries...996_01_L4.html
On the far side of the lake are barely resolved grass blades. That area is shown in Figure 2 on this page:
http://www.clarkvision.com/articles/...ng1/index.html
The image was sampled at various MTF levels. You would need to see the full size print, e.g. 30x40 inches to see that the structure in the grass gives the image wonderful fine detail. Sampling at the Rayleigh limit (9% MTF) has degraded the image detail, as does sampling at 25% MTF and 50% MTF.

Secondly, bicubic interpolation is not the same as spline interpolation in 2D. Of course any interpolation based on a cubic polynomial in 2D is a bicubic interpolation (and hence splines are a subset)

I understand that. I have coded splines. My point was that bicubic is a reasonable interpolation algorithm and its use should not be used to say an image test is invalid because it was used. It seems to me such an argument is kind of like the film versus digital wars when film people would claim you couldn't scan all the detail film contained so you couldn't show on the web all the detail film contained. While technically true, the difference in the real world images was meaningless.

Thirdly, I was surprised to see a mention of the Richardson-Lucy method!!! As you probably know, the use of this method requires a priori knowledge on the system function (sometimes referred to as the point spread in this case) of the entire camera system which is highly unlikely to be known in real life conditions i.e. when the shot is taken (you know the camera system is not entirely LTI). And as is the case with most iterative algorithms, if you’re going in the wrong direction, further iterations would only make things worse. of course Richardson-Lucy is a rather simple method and more sophisticated ML and MAP methods can be used for data recovery, but I doubt it that any photographer would use such methods to salvage detail and I think mentioning this would be going off topic.

You might want to check out this thread:
Important Sharpening Information!
http://www.birdphotographers.net/for...ght=Richardson
In panel 14, I show the use of Richardson-Lucy on the bird image. I do Richardson-Lucy on many of my images; it works much better than unsharp mask. Of course it would be ideal if one had a perfect description of the blur, but in real images of 3D scenes, the blur is variable due to depth-of-field and out of focus areas. So in practice, one simply uses several different models of what the blur might be, tuning it to the local conditions and then blending the results together. Most of the time a simple 3x3 or 5x5 pixel Gaussian model and a few dozen iterations does an overall good job. So no, you don't need prior knowledge of the system function.

Here is an article on its use on a fox image:
http://www.clarkvision.com/articles/image-restoration1/

Lastly, In regard to TCs, this has always been an issue for me and I didn’t quite understand your argument. Are you suggesting that in non-diffraction-limited cases, the use of TC should not significantly degrade the image quality? I have been using TCs for quite a while but I am used to getting significant image quality degradation with TCs (and I can assure you the focus is spot on and there is no camera shake but I don’t have my archive on my laptop here.)

Perceived image quality and angular resolution resolution are different things. For example, lets say a lens that produces a fuzzy (not diffraction limited) but resolves feather detail in a particular situation. Add a TC and you still resolve the same feather detail, but the image is magnified so looks less sharp. It may not resolve more detail. I have never seen a TC that results in less angular resolution, even cheap poor ones. Arash's statement said lpm drops a factor of 2 with a 1.4x TC. That will never be the case. And the test in this thread was about which camera with the same lens resolved more detail on the subject. That is angular resolution, not lpm in the focal plane.

So if you get significant image degradation in the use of a TC, why do you use one? I would think your answer would be because you resolve more detail on the subject. Arash's statement would mean you resolve less detail.

Roger

[John Chardine]

Roger- Is there software out there which allows a mere mortal like myself to play with and use Richardson-Lucy?

[Dan Avelon]

I have to agree with Sergey, I worked in a precision optics company here in San Jose in the past 20 years, before I retired and started photography again. We constantly used MTF50 and MTF30 charts as a standard benchmark for our optics, here is a good reference www.imatest.com/docs/sharpness.html of course you can plot MTF on a linear scale and go to 0 but I don't think that is meaningful as far as sharpness is concerned.

BTW loss of resolution with high magnification optics is not something unusual, if you have a microscope you can test this, the highest mag objective usually degrades the image more, the image is more magnified so you can see features easier but it is not as sharply defined. I think this is what the OP was saying, when you add the teleconverter the image is more magnified but it is blurrier, when you downsample and sharpen the image to recover it is essentially close to the image with the larger pixel camera, in other words the extra pixels are just resolving the blur.

[Roger Clark]

Roger- Is there software out there which allows a mere mortal like myself to play with and use Richardson-Lucy?

John,
Yes. I use ImagesPlus http://www.mlunsold.com/ (about $210 now)

The user interface is a little funky but quite usable. For the more adventuresome, there is IRIS, which is free, at http://www.astrosurf.com/buil/us/iris/iris.htm by Christian Buil (Christian also does sensor analyses and has many technical reviews of sensors on his site).

Both Iris and ImagesPlus are written for astrophotography processing, but many functions are the same as photoshop tools, like curves, levels, etc. I have not used iris, but know people who do.

Roger

[Roger Clark]

I have to agree with Sergey, I worked in a precision optics company here in San Jose in the past 20 years, before I retired and started photography again. We constantly used MTF50 and MTF30 charts as a standard benchmark for our optics, here is a good reference www.imatest.com/docs/sharpness.html of course you can plot MTF on a linear scale and go to 0 but I don't think that is meaningful as far as sharpness is concerned.

Dan,
I agree it is great to have a standard, like MTF50, to compare optical systems, but like any single metric, it is an incomplete description. I've attached an example sine wave chart and the arrow shows the 50% MTF point. Then I ran the chart through unsharp mask to boost the MTF. At the right edge of the lower chart, the MTF is about 10%. So do you really think none of the information below MTF 50 or 30 is important? I do. That is the fine detail in images that shows texture. It would be interesting to design a filter that would make a sharp cut at MTF 50 and MTF 30 and run that filter on real world images. I bet they would appear significantly degraded.

The fine detail and its loss is well illustrated in the MTF chart at:
http://www.cambridgeincolour.com/tut...resolution.htm
Go to the second MTF chart down the page. The blue line is a perfect lens, and below the chart you can move your mouse from a high quality lens to low quality lens. While the MTF50 point moves to lower frequency, it is the loss of MTF at the highest frequencies that also robs sharpness as illustrated by the mouse-over.

Roger

[Dan Avelon]

Dan,
I agree it is great to have a standard, like MTF50, to compare optical systems, but like any single metric, it is an incomplete description. I've attached an example sine wave chart and the arrow shows the 50% MTF point. Then I ran the chart through unsharp mask to boost the MTF. At the right edge of the lower chart, the MTF is about 10%. So do you really think none of the information below MTF 50 or 30 is important? I do. That is the fine detail in images that shows texture. It would be interesting to design a filter that would make a sharp cut at MTF 50 and MTF 30 and run that filter on real world images. I bet they would appear significantly degraded.

Roger

Roger,

I don't dispute that there is some spatial information past the 50% point, but not close to 0% as you originally suggested by 'MTF0', which BTW I had never heard before. The chart you included is a loglog plot so the right edge is 10% at which there is already no detail left.



Another point that you have to consider is that this data is not from real optics, it is matlab simulation. The combination of camera optics have a much sharper roll off about the 50% point, as noted in the reference too. In practice salvaging the attenuated tones in presence of imperfections and noise becomes very difficult and of little practical benefit, that is why industry has chosen to adopt the 50% standard for a wide variety of applications.

[Dan Avelon]

Dan,


The fine detail and its loss is well illustrated in the MTF chart at:
http://www.cambridgeincolour.com/tut...resolution.htm
Go to the second MTF chart down the page. The blue line is a perfect lens, and below the chart you can move your mouse from a high quality lens to low quality lens. While the MTF50 point moves to lower frequency, it is the loss of MTF at the highest frequencies that also robs sharpness as illustrated by the mouse-over.

Roger

This is very nice demonstration Roger, this is actually exactly the point the OP was making, you can think about the Canon prime as the ideal lens (purple curve), when you add the TC MTF50 point moves to the left (taking your own 1.4 number from earlier) this will be exactly the green curve which no longer resolves meaningful detail at high frequencies, in practice I also thin the loss of MTF due to current Canon TCII is more than 1.4X, if not 2X at least 1.6X. I searched the intertnet but did not find any actual charts from Canon so this is based on my own quick measurement.

A equivalant to prime lens



B after adding a teleconvertor with 1.4X shift in MTF50P

[Roger Clark]

Roger,

I don't dispute that there is some spatial information past the 50% point, but not close to 0% as you originally suggested by 'MTF0', which BTW I had never heard before. The chart you included is a loglog plot so the right edge is 10% at which there is already no detail left.

Another point that you have to consider is that this data is not from real optics, it is matlab simulation. The combination of camera optics have a much sharper roll off about the 50% point, as noted in the reference too. In practice salvaging the attenuated tones in presence of imperfections and noise becomes very difficult and of little practical benefit, that is why industry has chosen to adopt the 50% standard for a wide variety of applications.

Well, this sounds like a degradation of a film versus digital wars of yesteryear. The effect I illustrated shows detail well below 50% MTF and it really matters not if it was done with a real lens. This weekend if I have time, I will generate an example using a real lens. I am confident the results will show the same thing, and that a lot of what people discuss on BPN regarding perceived sharpness is at frequencies above MTF 50 (to lower MTF which gets boosted in post process sharpening).

Roger

[Dan Avelon]

Well, this sounds like a degradation of a film versus digital wars of yesteryear. The effect I illustrated shows detail well below 50% MTF and it really matters not if it was done with a real lens. This weekend if I have time, I will generate an example using a real lens. I am confident the results will show the same thing, and that a lot of what people discuss on BPN regarding perceived sharpness is at frequencies above MTF 50 (to lower MTF which gets boosted in post process sharpening).

Roger

Roger,

I am also sure and agree with you that with advanced sharpening you can salvage detail above MTF50, maybe all the way down to MTF10 if noise is not too much. But what I and maybe some other people like to see is howmuch the entire MTF curve shifts to the left, when you add the TC. Depending on the amount of degradation the MTF10 point may or may not fall close to cut off frequency of the small pixel camera. So if you can, measure and plot MTF50, MTF30 and MTF10 points versus frequency with and without TC for a sharp prime so we can see how the numbers compare and will have a better picture.



Cheers,

Dan

[Roger Clark]

Roger,
I am also sure and agree with you that with advanced sharpening you can salvage detail above MTF50, maybe all the way down to MTF10 if noise is not too much. But what I and maybe some other people like to see is howmuch the entire MTF curve shifts to the left, when you add the TC. Depending on the amount of degradation the MTF10 point may or may not fall close to cut off frequency of the small pixel camera. So if you can, measure and plot MTF50, MTF30 and MTF10 points versus frequency with and without TC for a sharp prime so we can see how the

OK, here are MTF curves for a 300 mm f/2.8 lens on a 5D Mark II. The 5D2 is the camera I have with the smallest pixel pitch. The results are considerably better than I thought they would be. I chose the 300 f/2.8L IS because it is reportedly the sharpest lens in the canon line or at least near the top if not the top. The plots are for the center of the field of view. I used my custom bar chart based on the air force chart with additions. My chart was at 37 feet distance and all exposures were with mirror lockup with about 15 seconds settling time.

With no TC, MTF is limited by the blur filter at f/2.8.

With a 1.4x TC, MTF50 (50% MTF) frequency (line pairs per mm, lpm) drops by only 12%, and that is due to the increase in diffraction, but still limited by the blur filter. Note for those just entering the thread: Arash predicted 2x loss, Dan 1.6x loss, actual 1.12x loss.

With a 2x TC, we are seeing only about a 1.5x loss in MTF50 frequency. This is again because the no TC case is limited by the blur filter, not diffraction of lens aberrations. The MTF50 with the 2x TC is right where it would be expected due to diffraction. It doesn't get any better than that! That's an impressive lens and TC.

On a camera with finer pixel pitch, like a 7D, effects from diffraction and lens aberrations should be more prominent. If the TC were perfect and added no aberrations and the camera resolved the diffraction effects, one would expect that the drop in lpm with a 1.4x TC would be 1.4x and with a 2x TC would be 2x. The fact that we are seeing less that that and that the 2x TC is very close to the diffraction line is showing that the TCs are not degrading image quality in any significant way.

On a related topic, the half-Nyquist point with the 2x TC, where the images are diffraction limited, not blur filter limited, means image detail is spread over 4 pixels. The idea that there is little information finer than 4 pixels spacing means the destruction of fine detail.

One other note: the no TC case shows response above the Nyquist frequency and indicated the blur filter is slightly weak.

Roger

[Dan Avelon]

Roger,
Thanks for doing the test, AFAK when adding 1.4X TC light drops by a full stop but magnification increases by 1.4X so I would expect the MTF to drop by 2/1.4=1.43 relative to the prime, if the TC was perfect, in your test it drops by 12% so can you elaborate on the method you extracted MTF? perhaps images of the test strip if possible?

Also taking the above data, when adding 1.4X TC at nyquist limit MTF is about 5%, so the part that is blur filter limited is between MTF10 and MTF5 which I argue and I think you agree is negligible for practical purposes. so it is very clear that for a tighter pixel camera Nyquist is way past MTF0 point, and unless the 7D has an unusually strong blur filter that I doubt based on the samples I have seen, it will be diffraction limited. And this was with 300 f/2.8 lens which I believe if not sharper, is as sharp as 500 and 600 primes. Based on this data there isn't much improvement in the small pixel camera when you add the TC. Of course pixels resolving diffraction blur are still better than interpolated pixels, so I believe if you want to upsample the photos, it has an advantage but not otherwise.

Thanks for doing the test, very informative

[Emil Martinec]

Roger,
Thanks for doing the test, AFAK when adding 1.4X TC light drops by a full stop but magnification increases by 1.4X so I would expect the MTF to drop by 2/1.4=1.43 relative to the prime, if the TC was perfect, in your test it drops by 12% so can you elaborate on the method you extracted MTF? perhaps images of the test strip if possible?

I'm not understanding your logic here; it seems to based on sort of model for the way contrast reduces that I don't follow. Are you saying that there is a rough scaling relationship for the MTF curve, that if you look along it by increasing the resolution by 1.4x, that the value of the curve drops by a factor 1.4x? That is the specification for the function f(x)=1/x, not the straight line that is observed at the tail end of the MTF curve.

The same intuition seems to underly Roger's thinking too...

With no TC, MTF is limited by the blur filter at f/2.8.

With a 1.4x TC, MTF50 (50% MTF) frequency (line pairs per mm, lpm) drops by only 12%, and that is due to the increase in diffraction, but still limited by the blur filter.

With a 2x TC, we are seeing only about a 1.5x loss in MTF50 frequency. This is again because the no TC case is limited by the blur filter, not diffraction of lens aberrations. The MTF50 with the 2x TC is right where it would be expected due to diffraction. It doesn't get any better than that! That's an impressive lens and TC.

I would have said, the pixel pitch and the blur filter (and of course, the two are correlated; the strength ie MTF50 point of the OLPF is set by the pixel pitch). But if we treat the pixels as a box blur of an incoming analog signal, they also degrade MTF relative to infinitely fine sampling.

On a camera with finer pixel pitch, like a 7D, effects from diffraction and lens aberrations should be more prominent. If the TC were perfect and added no aberrations and the camera resolved the diffraction effects, one would expect that the drop in lpm with a 1.4x TC would be 1.4x and with a 2x TC would be 2x. The fact that we are seeing less that that and that the 2x TC is very close to the diffraction line is showing that the TCs are not degrading image quality in any significant way.

Again this logic leads to an MTF curve that goes like 1/x, so I am puzzled where it is coming from. I would have said that a perfect 1.4x TC would slide you along the MTF curve by a factor 1.4 by simply magnifying everything in the image by that factor on the sensor. To know what that does to the MTF in lpm one has to know the slope of the MTF curve. Note that this seems to differ between the three curves, indicating some other factors at work.

The test of whether the images are diffraction limited is to stop down a stop and see that the resolution has dropped by 1.4x, not to change the magnification by 1.4x and stop down a stop and see that the resolution has dropped by 1.4x. But that is a different test than the one that was performed.

On a related topic, the half-Nyquist point with the 2x TC, where the images are diffraction limited, not blur filter limited, means image detail is spread over 4 pixels. The idea that there is little information finer than 4 pixels spacing means the destruction of fine detail.

Roger

That may be because the lens is wide open. Most people argue that results are better with the 2X TC when stopping down some. Again it would be interesting to see the above graph repeated at one stop and two stops down from wide open to see what diffraction is really doing, vs say the effects of lens/TC aberrations.

My take is the following: MTF50 goes from about 60 lpm bare to about 55 with the 1.4x and 40 with the 2x. That means that with the 1.4x TC a given object is magnified by 1.4x with a 1.2x loss in resolution at the sensor, meaning that the object is resolved 1.2x better. With the 2x TC, the gain is 2x magnification vs 1.5x loss in resolution on the sensor plain, for again of 1.33x in resolution of the subject. I'd be interested to see if those numbers improve if one stops down, as many people seem to say.