Pages: [1]   Go Down

Author Topic: Comparing 7D and 5D2 limiting sensor resolution.  (Read 4522 times)

AJSJones

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 357
Comparing 7D and 5D2 limiting sensor resolution.
« on: February 23, 2012, 10:58:15 pm »

(Geek alert)
Following an earlier thread on lens vs sensor resolution, I learned of Bart van der Wolf's test image and some interesting discussion of images of it from various situations.  Since it is so easy to use (it is not "distance sensitive") I decided to give it a go and report anything interesting.  There have been a lot of discussions of 7D vs 5D2, with a lot of "noise" and I have both cameras.  What follows is in the format of a techincal report (that was a skill developed in my non-photography related career :D )  It was a fun experience for me, a bit like some of my scientific work from before I retired, so I hope someone finds it interesting - or have some suggestions for improving technique - I'm getting a bit behind the times on some of these things.
Cheers
Andy

Comparing resolution limits of 7D and 5D2.
Summary
Both cameras showed "true" resolution to about 92% of Nyquist frequency. Adobe Camera Raw provides that resolution (107 cy/mm for 7D and 71 cy/mm for 5D2), while DPP gives only 83% but notably better color moire control.  For fabric/fashion, repeating things (distant bricks etc) DPP might be preferred, while for landscape and fine, irregular detail, ACR might be preferred. (I've not used any other converters)  5D2's AA filter seems to be a little weaker, but this did not seem to affect the “limiting on-sensor” resolution.

Background

I don't have Windows (neither on a PC nor on an emulator on my Mac), so I can't use Imatest – I volunteered long ago as an alpha tester for Mac but Norm never seemed to proceed with that side of the project.  Bart's star seemed like a fairly easy way to test things for myself..

Methods

I printed Bart's “Siemens star” on Epson premium luster paper on my 7600 and all but the central 8-9 mm diameter was “resolved” - with moire artifacts and grey inside that.  I added it to my aging test target close to the center, inside the yellow rectangle.

 

I shot at f/8 with a 7D and a 5D2 from about 20 feet using a 100mm f/2.8 L macro with IS off.  Focusing was manual with Live View at 10X – the star made that very easy – tripod, 2 sec timer and remote release (mirror already up).  Lens optical axis was at the height of the center of target and they were essentially at right angles - camera double bubble leveled, target vertical. Even with a 4 µm CoC (~the 7D pixel size) the DoF should be 6” (±3”) and inspection of all the focus attempts showed they were indeed indistinguishable.  Raw files were converted by ACR (5.7 I think - in CS4, no sharpening) or DPP at “default settings” (in DPP all neutral, sharpness zero DPP 3.8.2 ).  WB was not an area of concern so “as shot” was accepted for both.  In PS CS4, white point and black points were set in a curves layer using the ends of the step wedge.  My typical “sharpening” is a “1 pixel radius “Generic” deblur in Topaz In Focus” and this was applied to all images. (The blur circles didn't change as a result of this). Composited images were converted to 8bit, sRGB and saved for web at 80% (this preserved all the relevant moire details).  One image (1100x811 pixels) contains the original pixels from each camera, so the size of the images is different on the screen, consistent with the pixel size difference



The other image has the 7D image downsized (simple bicubic) to the same size as the 5D2's image, for those who like equal area comparisons. (It is 850x535 pixels) How one might uprez the 5D2 image to the same "size" as the 7D raises its own issues which I decided not to address :D

.


Observations

It took a bit of experience to place the right size of circle to define the “central blur” area and it remains a bit subjective, but I think I got to be internally quite consistent.  All the radii hit the circle even if there is some evidence of aliasing outside the circle at that point.  Some radii go inside the circle and that varies by angular position (not surprising given a fixed array and asymmetrical lens aberrations) but mostly it was either extreme moire/aliasing or just blurred. These may be easier to assess when viewed at 200 or 300% on your monitor. I guess the precision of  the diameter of the circles I created would be ±2-3 pixels.

The cool thing about Bart's star is that there is no need to fuss over precise distances and focusing is a dream.  The key pieces of information (for me)  from this comparison

1) The confirmation of an oft-asserted but rarely-demonstrated difference in the AA filter between the 5D2 and the 7D.  I had always been skeptical of such claims but this set of actual data  seemed to support it.  Looking at the star from the 5D2 it is clear that there is false information outside the circle (i.e. not just the radii but quite a lot of weird lines – aliasing coming down below Nyquist to irreparably, if subtly, damage the real data) and the clarity of the aliasing and the intensity of the colors inside the blur circle all point to a weaker AA filter – by how much, who knows?!  The other indicator is the vertical line of 12 pt type from the ACR/5D2 data where the cyan/orange repeat is actually quite “pretty”!  Some aliasing is also seen in the 9 pt type from the 7D, with the same color repeat pattern.  While the AA filter seems somewhat weaker, based on the moire patterns and clear aliasing below Nyquist, it seems that the overall resolution limits of the two sensors are similar, when related to their own Nyquist frequency.

2)  The raw converters make different compromises between perceived intensity of residual color aliasing and luminance resolution.  ACR gave 91 and 92% of pixel-pitch-based Nyquist (107 and 71 compared to 116 and 78 cy/mm for the pitch of 7D and 5D2 respectively) while DPP gave 83% for both.  Based on the earlier discussions, it's likely that this resolution limit represents around 10% MTF – although the comparisons rely on the consistency rather than absolute value. The ACR images had more residual color patterns close to the limit than the DPP ones.  I know nothing about the inner workings of the demosaicing and color interpolations and the compromises that are presumably involved, so someone else will hopefully be able to comment.

3) The confirmation that the lens I used seems to outresolve the sensor – why I originally got started on this little project to satisfy my curiosity. (Looking at the star placed at different distances from the optical axis will await a rainy day).



« Last Edit: February 23, 2012, 11:01:00 pm by AJSJones »
Logged

Bart_van_der_Wolf

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 8862
Re: Comparing 7D and 5D2 limiting sensor resolution.
« Reply #1 on: February 24, 2012, 06:41:16 am »

I printed Bart's “Siemens star” on Epson premium luster paper on my 7600 and all but the central 8-9 mm diameter was “resolved” - with moire artifacts and grey inside that.

Hi Andy,

Thanks for the feedback. I don't worry about people calling this type of tests geeky, they will not gain the insights that the geeks do. In fact these tests may reveal improvement areas in our technique or tools that would otherwise not have been identified.

I have a few observations that might require checking. The unresolved central diameter of 8-9 mm on your print might indicate something interesting as well. It won't invalidate the results, but it may require shooting at longer distances than 25x focal length (which you did at 60x FL, so no problem) to be able and blur the center. 8 mm equals only 291 PPI print resolution on paper. I think that your printer is capable of more detail (QED), even on Lustre paper instead of glossy.

This is how to determine, (144/pi)/8mm = 5.7 cy/mm, or ~11.46 lines per mm, which is 291 PPI. Again, no problem because that approximately equals human visual acuity at normal reading distance, but it could be better. I'm not sure if you used the "Finest detail" option of your printer driver, which allows to use the 720 PPI print resolution, if not then the driver will downsample crudely to 360 PPI. Combined with a little surface stucture that could end up a little bit lower, e.g. that approx. 291 PPI. To give you an idea about what is possible, on my simple Pixma A4 desktop printer I can achieve a blur center of +/-3.5mm on glossy paper which equates to 665 PPI, which matches the 600 PPI maximum resolution (within fractional mm measuring inaccuracy).

Quote
I shot at f/8 with a 7D and a 5D2 from about 20 feet using a 100mm f/2.8 L macro with IS off.

Nothing wrong with keeping all parameters the same, but do realise that the small sensels of the 7D will have lower contrast due to diffraction. This will suppress the inevitable aliasing a bit more, compared to the 5D2. I don't think it will drastically change the 7D's resolution limit that you've found, although it might reduce resolution for lower contrast detail and it will affect the aliasing. So maybe the difference with the 5D2 moiré is a bit less pronounced than what you found.  

Quote
Raw files were converted by ACR (5.7 I think - in CS4, no sharpening) or DPP at “default settings” (in DPP all neutral, sharpness zero DPP 3.8.2 ).

In DPP, make sure you also set the noise reduction to zero, as it seems to be on by default and the amount differs depending on source and camera settings.


Quote
Observations

It took a bit of experience to place the right size of circle to define the “central blur” area and it remains a bit subjective, but I think I got to be internally quite consistent.  All the radii hit the circle even if there is some evidence of aliasing outside the circle at that point.  Some radii go inside the circle and that varies by angular position (not surprising given a fixed array and asymmetrical lens aberrations) but mostly it was either extreme moire/aliasing or just blurred. These may be easier to assess when viewed at 200 or 300% on your monitor. I guess the precision of  the diameter of the circles I created would be ±2-3 pixels.

The exact cut-off can indeed be a bit arbitrary and has to do with orientation (higher resolution diagonally) and accidental alignment with the sensel grid, but we can only fool ourselves. ;) That's why I suggest to work with the diameter and not the radius of the blur disc, it will require 2 decisions which tends to cancel out errors.  In practice, due to differences in contrast and less than optimal shooting conditions there will be some spread as well, but the bias will be in the larger diameter direction. Aliasing artifacts can indeed manifest themselves outside the blur diameter, they are by definition larger than at the Nyquist frequency, and inside the limiting resolution they become easy to detect due to their deviation from the straight line towards the center point. This also demonstrates that aliasing artifacts do not always produce a simulated coarser resolution pattern, they can stick out like a soar thumb, and be hard to cure.

Quote
The cool thing about Bart's star is that there is no need to fuss over precise distances and focusing is a dream.  The key pieces of information (for me)  from this comparison

1) The confirmation of an oft-asserted but rarely-demonstrated difference in the AA filter between the 5D2 and the 7D.  I had always been skeptical of such claims but this set of actual data  seemed to support it.

Indeed, although you might get a bit less of a difference when shooting with a wider aperture on the 7D. With a 4.3 micron sensel pitch, one can expect diffraction to soften the per pixel contrast (and thus aliasing) at f/5 and narrower. The 5D2 will only be impacted starting at f/7.1 .

Quote
While the AA filter seems somewhat weaker, based on the moire patterns and clear aliasing below Nyquist, it seems that the overall resolution limits of the two sensors are similar, when related to their own Nyquist frequency.

Yes, that's quite common for good lenses. Only when a lens has issues (used wide open, or defocused/decentered, corners) the distance to the Nyquist frequency will increase.

Quote
2)  The raw converters make different compromises between perceived intensity of residual color aliasing and luminance resolution.  ACR gave 91 and 92% of pixel-pitch-based Nyquist (107 and 71 compared to 116 and 78 cy/mm for the pitch of 7D and 5D2 respectively) while DPP gave 83% for both.

Do check the default noise reduction settings in DPP. Nevertheless there are different choices made in dfferent converters. These tests can reveal which one might be better suited for a specific task. Afterall, we have the option to combine conversion results in postprocessing should we encounter issues in our preferred (for it's workflow) converter that are better handled in another converter.

Quote
3) The confirmation that the lens I used seems to outresolve the sensor – why I originally got started on this little project to satisfy my curiosity. (Looking at the star placed at different distances from the optical axis will await a rainy day).

While sampling with a higher sampling density sensor will always benefit the resolution one can extract from a lens, the fact that the percentage of Nyquist frequency with the 7D stayed virtually constant compared to the 5D2 shows that there is still enough lens resolution to spare. Corner performance will always be less than in the center of the image circle, but that is often of lesser importance with 3-dimensional scenes where DOF defocus already reduces resolution. When testing corners, make sure to also focus in those corners because due to curvature of field it may require a different focus distance than for the optical center. Only a lens that is specifically designed for flatness of field will focus the corners at the same distance.

Cheers,
Bart
« Last Edit: February 24, 2012, 08:52:36 am by BartvanderWolf »
Logged
== If you do what you did, you'll get what you got. ==

AJSJones

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 357
Re: Comparing 7D and 5D2 limiting sensor resolution.
« Reply #2 on: February 24, 2012, 11:58:36 am »

Thanks Bart , for your considered feedback; the geek alert was only to warn off those who only assess these things subjectively and who are not fond of numbers and charts, not to waste their time on the post :D

I forgot to mention that I had indeed turned off both chroma and luma noise reduction in DPP.  The 8-9 mm estimate was completely eyeball (no loupe) of the area where the whole set of radii were clearly perfectly reproduced (no residual moire, in particular) and no attempt to assess the blur diameter for printer resolution! It almost certainly is a large overestimate compared to if I were to try to put the circle in there to assess the printer resolution.  I mentioned it only because I wanted to make sure what I shot had absolutely no imperfections related to the target.  I perhaps put more emphasis on a safety factor for DoF than worrying about diffraction, but it's easy enough to reshoot at f/5.6. - I think the target makes 10x LiveView focus acqusition pretty precise and the lower DoF will still be sufficient for peace of mind!

The patterns in the data below the Nyquist were particularly helpful to me to understand the nature of the problem and the irreversibility of undoing the aliasing, which is noticeable even in a grey image portion - I had been thinking the issue was primarily one evidenced by what folks call "color moire" that could be "corrected" post facto.  I rarely shoot fabric or regular patterns so I think I will stick with ACR  for my fine-detail landscape shots (that comparison held for both images) . 

As for testing lenses, I will sometime compare the target in the corners between focusing on the center target and the corner target - for flatness of field as a separate (?) issue from corner lens resolution.  We need some rain here, for that to happen, and we need rain for other reasons, too!

Thanks again - updates to follow
Andy
Logged
Pages: [1]   Go Up