Sign in to follow this  
Followers 0

Comparing Actual Camera Resolutions

37 posts in this topic

Posted (edited) · Report post

A camera list and resolution chart showing something like this would be helpful (and motivate manufacturers to build products trying to get higher up on this list):
 
After working on this list for a while and seeing more charts, understanding Nyquist sampling theory is helpful ( http://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem ). Take the sensor width in pixels and divide by 2 to get a fair estimate of sampling resolution before aliasing. While aliasing past Nyquist can help make an image, especially a still frame, look sharp, when in motion, an alias-free image, especially with sufficient grain and high-frequency detail, will look most like film. Thus, even though the ARRI Alexa is only a 1080p camera, it currently produces the most film-like image and explains why it is used most when cost is no object.
 
<all cameras below have pro audio quality. Listing SNR & noise floor would be helpful)
F65 - 3.9K (8K sensor with a few tricks, not true 4K) - 16-bit RAW
RED - 3.xK (6K sensor: Dragon) - Wavelet compressed RAW
RED - 2.5K (5K sensor) - Wavelet compressed RAW (see charts in this thread)
1DC - 2.xK - 8-bit MJPEG and H.264 ("4K" mode, need charts)
F55 - 2.xK - 10-bit, XAVC, RAW (optional, extra hardware)
F5 - 2.xK - 10-bit, XAVC, RAW (optional, extra hardware)
FS700 - 2.xK (estimated with "4K" upgrade) - 12-bit RAW (extra firmware and hardware)
ARRI Alexa - 1.9K (3.5K sensor) - 444 and RAW - pretty much zero aliasing and very film-like grain (see charts in this thread)
BMCC - 1.xK (needs a test chart: will be less the 2.7K, closer to 2K from what I have seen so far) - 12-bit RAW <1/4" audio in>
Hero 3 1.xK (needs a test chart) - 8 bit, 2.7K mode, H.264
F3 - 1.9K - 10-bit - Sony Pro CODECS, 10-bit 422 out
1DC - 1.9K - 8-bit H.264 Super 35 Mode
C300 - 1.9K - 8-bit, 50Mbps long GOP MPEG2 (422), 8-bit 422 out, minimal aliasing
C100 - 1.9K - 8-bit, Very high quality 24Mbps H.264:AVCHD (420), 8-bit 422 out, minimal aliasing
FS700 - 1.9Kv, >1.9k diagonal, ~1.6Kh (trumpet chart) - 8-bit - Very high quality 24Mbps H.264:AVCHD, 8-bit 422 out, some aliasing
FS100 - 1.9Kv, >1.9k diagonal, ~1.6Kh (trumpet chart) - 8-bit - Very high quality 24Mbps H.264:AVCHD, 8-bit 422 out, some aliasing
GH2 - 1.7K - 8-bit,H.264:AVCHD  <add hacked bitrates> <not pro audio, add audio details>
TM700 - 1.7K - 8-bit, 28Mbps H.264:AVCHD (60p), excellent IS, loud fan (audio issue)
 
<insert popular camcorders such as TM900, G10/XA10, NEX camcorders, XF105/305, etc.>
 
<all cameras below don't provide pro audio>
GH3 - 1.8-1.9K? - 8-bit, 90Mbps H.264 I-frame, ~40Mbps IPB <add audio details>
Hero 3 1.8-1.9K? - 8 bit, 1080p mode, H.264
1DC - 1.7K* - 8-bit H.264 full frame 1080p mode
1DX - 1.7K* - 8-bit H.264 full frame 1080p
5D3 - 1.6K (post-sharpened) - 8-bit, 90Mbps H.264 I-frame, ~40Mbps IPB <audio - 1/8" stereo mic, average noise floor>
6D - ~1.6K (post-sharpened) - 8-bit, 90Mbps H.264 I-frame, ~40Mbps IPB
5D2 - 1.5K - 8-bit, ~38Mbps H.264 IPB <audio - 1/8" stereo mic, average to poor noise floor>
 
This was a quick write up without verifying some of these items. Anyone interested: please correct and/or add detail to this list.
 
The resolution column is actual resolution of captured frames when displayed as RGB. Sensor resolution and manufacturer claims ignored in favor of actual hardware testing. Cameras such as the 5D3 may exhibit first artifacts around 1200 pixels (600 TV Lines), however detail extinction does not occur until around 1620 pixels (810 TV lines). Thus, it would be helpful to show both first artifacts as well as detail extinction.
 
 
*1DC, 1DX 1080p mode: estimated (looks slightly higher resolution vs. 5D3)
 
<TODO> Add supported frame rates.
 
Sony FS100/700 have unusual resolution behavior. Diagonal resolution is 1000+, vertical is ~1000, however horizontal is ~800 (when using a trumpet chart). A line chart shows 1000 or so for horizontal resolution, however aliasing has started. These cameras can also alias on fine horizontal and vertical lines. C100/C300 can also alias, however the cameras provide 1000 lines of resolution on both h and v.
Edited by jcs

Share this post


Link to post
Share on other sites

Posted · Report post

add gh2 :)

Share this post


Link to post
Share on other sites

Posted · Report post

Pretty good list. Couple of things - I'd put the F65 at the top, it is the only true 4K camera sampling from an 8K sensor! GH2 is above FS100 for resolution.

Share this post


Link to post
Share on other sites

Posted · Report post

Updated. Regarding the GH2- how do you know the GH2 provides higher resolution vs. the FS100? Any links to tests/reports etc.?

Share this post


Link to post
Share on other sites

Posted · Report post

RED Epic has 5120 x 2700 Bayer pattern sensor. As Andrew noted, we need at least 8192 photosites (minimally, in green, horizontally) to actually capture 4096 pixel resolution (Nyquist- math and physics). Right now only the F65 has 8K sensor resolution (which provides for the most part full 4K resolution (but not quite: download the brochure from this page: http://pro.sony.com/bbsc/ssr/show-highend/resource.solutions.bbsccms-assets-show-highend-F65.shtml )).

 

The goal of this list/table is to show actual, measured, real-world performance (which can differ greatly from advertised specs).

Share this post


Link to post
Share on other sites

Posted · Report post

The goal of this list/table is to show actual, measured, real-world performance (which can differ greatly from advertised specs).

Not that I doubt your figures, but what is your source? It be nice to see the measurements/comparisons.

Share this post


Link to post
Share on other sites

Posted · Report post

Thanks for that list I've thought something like that would be nice for a while. GH2 is possibly about 830 lines with the hack I reckon.

Share this post


Link to post
Share on other sites

Posted · Report post

guys, where is gopro ?

Share this post


Link to post
Share on other sites

Posted · Report post

markm- if the GH2 is 830 TV Lines (horizontal resolution), that would mean the FS700 is less (per Andrew), however the FS100 is rated at 1000+ TV Lines. Are there charts, lab data to support these numbers?

 

Found an FS100 chart: http://www.dvuser.co.uk/images/img/reviews/camcorders/sony-fs100/chart.jpg

http://www.dvinfo.net/forum/panasonic-lumix-gf-gh-series/503321-how-many-lines-resolution-gh2.htm

http://www.dvxuser.com/V6/archive/index.php/t-232565.html (GH2 rated around 800 lines).

Looking at the FS100 and GH2 charts, the FS100 is pretty solid all the way up to 1000 lines (with some artifacts). The GH2 has artifacts at 800 lines. Is there a chart showing the GH2 doing better than 800 lines? (see FS100 chart link above)
 
This result may be due to compressor issues:
In other words, GH2 might have done better in this test due to better compression, where detail was lost for the FS100. From the charts I have seen, the FS100 is doing close to advertised 1000 TV Lines (see FS100 chart link above).
 
Unless there is a GH2 chart showing better results vs. the FS100, I'll move the FS100 back up (FS100 should do much better than GH2 with an external recorder to match/exceed hacked GH2 bitrates).
 

I have a GoPro Hero 3 Black Edition. 1920x1080 looks decent, 2.7K looks OK but suspect it's not really 2.7K (haven't tried 12fps 4K; would guess closer to 3K). I don't have any data from tests showing actual resolution. If someone has a source/link I can take a look and add it to the list.

Share this post


Link to post
Share on other sites

Posted · Report post

It seems quite odd that "post-sharpening" has increased resolution on the 5dmarkIII. If you are gonna do something cool and get cameras into a resolution list, then they would actually need to be measured. Changing sharpening in post does not change results from a real resolution chart.

Share this post


Link to post
Share on other sites

Posted · Report post

JCS TV lines described in advertising literature I don't believe has any relevence at all so I would disregard that.

I initially thought the GH2 had the same resolution as the Panasonic AF100 at about 630 lines which is an awful performance that doesn't even match 720p. I think it is fairly common for the large sensors to have about 800 including the Sony F3.  Really you need tried and trusted experts like Adam Wilt or Alan Roberts to do these tests..

 

Also sharpening cannot increase resolution.

Share this post


Link to post
Share on other sites

Posted · Report post

Updated. Regarding the GH2- how do you know the GH2 provides higher resolution vs. the FS100? Any links to tests/reports etc.?

 

Seriously?

Share this post


Link to post
Share on other sites

Posted · Report post

Buuurrrrrrrrrrp.

Hoo rah cameras..

I'm gonna make a full frame Iphone that shoots raw and fits in your pocket.. I i "AM" form factor.. Siri tells you if the shot was worth a damn also.. my camera is the first to instill concidence with verbal validations. Makes the Dragon sensor look cute. 4k is so 2013.. Real innovation doesn't start until you can see a cold developing on the surface of his fingernails. In a bat cave around midnight at 2,365,432 iso with no noise. 12k for $12k. Where you at now Canon? I hear the ball dropping.

Share this post


Link to post
Share on other sites

Posted · Report post

Have a look at the details for F65, I think that although a 16-bit process is used, colour information is actually 12-bit by the time you get the footage. Or something like that... maths...

Share this post


Link to post
Share on other sites

Posted · Report post

I realized that we could reasonably simulate test chart results in software. A camera system (Gaussian) blurs light before the sensor (anti-aliases) then decimates (A to D samples pixels) into discrete photosites.

 

I downloaded the ISO12233 test chart and converted the PDF into a 1200 pixels/inch bitmap (Photoshop CS6). I then Gaussian blurred it 3 pixels and downsampled (bicubic smoother) to 1920x1080, then sharpened it to make the results more clear (to increase accutance):

post-16978-0-05657200-1359600005_thumb.j

Note aliasing starting above 900 lines.

 

I went back to the blurred 16K image and did the same process down to 1620x910 (+ sharpened):

post-16978-0-55358800-1359600049_thumb.j

Note that aliasing starts to become apparent around 600 lines.

 

Blurring and decimating is a common method to resample pixels. However, there are better resampling methods than bicubic such as the sinc-based Lanczos (Premiere Pro uses Lanczos).

 

In the 1920 example, we see it is pretty clean all the way up to 1000 line pairs.

Compare to an actual FS100 test chart shot: http://www.dvuser.co...fs100/chart.jpg

post-16978-0-59331600-1359601660_thumb.j

Looks like a solid 1000+ TV lines (with two interesting diagonal aliasing bands).

 

Now compare to the GH2 :

post-16978-0-28245100-1359601306_thumb.j

 

Different chart; Barry Green rates this as around 800 lines perhaps could be rated higher, but not as high as the FS100 chart result. We could say 900 or so before detail extinction.

http://www.dvxuser.com/V6/archive/index.php/t-232565.html

 

Philip Bloom's test compared the hacked GH2 to the FS100's 24Mbps codec. It appeared to me that the FS100 low detail/resolution could be due to overcompression. If using an external recorder, the FS100 should be able to do much better in that test case.

 

The 5D2 was rated at 440lp/ph (vertical resolution), so that would be 880 vertical pixels and if horizontal resolution is similar, we'd get 782lp/pw or 1564 horizontal resolution. Thus approximately 1564x880 pixel resolution. It's possible horizontal resolution is less- if anyone has a reference, please post it.

http://provideocoalition.com/awilt/story/nab_2011_-_scce_charts/

 

The 5D3 providing around 1620x910 resolution (the bitmap size before rescaling to 1920x1080 per ML folks) or around 810 TV lines seems reasonable:

https://vimeo.com/39517721 (890 line estimate was perhaps a bit too generous).

 

Closeup of the 5D3 still scaled to 1920x1080 (represented full 1920x1080 resolution on the device):

post-16978-0-81552800-1359603430_thumb.j

 

Closeup of 5D3 1080p video post-sharpened:

post-16978-0-64598500-1359603539_thumb.j

 

We can see decent line pairs out past 800. 810 would be fair and matches the 1620x910 pre-scaled bitmap reported by the ML team.

 

As always, if there are better chart or other real-world test results, please post them.

Share this post


Link to post
Share on other sites

Posted · Report post

Pretty good list. Couple of things - I'd put the F65 at the top, it is the only true 4K camera sampling from an 8K sensor! GH2 is above FS100 for resolution.

 

 

Andrew, I am guessing, you meant the AF100, instead of the FS100. Please correct me, if I am wrong.

Share this post


Link to post
Share on other sites

Posted · Report post

JCS I appreciate what you're trying to do but tests like these have to be done by the book and with proper test charts otherwise your results wont match other tests and are meaningless.

Share this post


Link to post
Share on other sites

Posted · Report post

Hey mark- a chart test is pretty straightforward, right? Align and focus correctly, shoot. Other than having the chart too close, errors in set up result in reduced measured resolution. In my day job I work with cameras, pixels, and math. For example, we could do an automated FFT/DCT for a groups of photographed(video) lines at various resolutions and look at the frequency bins to get an empirical measurement. For such a test, we'd want to clean up via sharpening the source frames to remove as much low-frequency information as possible (improving accutance (a form of microcontrast) as much as possible). I suspect the slashcam.de test did something similar and didn't post-sharpen the 5D3 footage, resulting in an abnormally low high-frequency measurement.

 

The charts I found looked pretty solid. Which one(s) do you feel weren't done by the book?

Share this post


Link to post
Share on other sites

Posted · Report post

Andrew, I am guessing, you meant the AF100, instead of the FS100. Please correct me, if I am wrong.

 

No - GH2 is better than AF100 on a test chart for resolution.

Share this post


Link to post
Share on other sites

Posted · Report post

Hey mark- a chart test is pretty straightforward, right? Align and focus correctly, shoot.

 

It is very straight forward, but many other factors can influence resolution/sharpness. The lens used is a big factor.

It would be a straightforward test if all the cameras were done at the same time with the same chart and preferably the same lens (sharpest possible).

Share this post


Link to post
Share on other sites

Posted · Report post

Hey mark- a chart test is pretty straightforward, right? Align and focus correctly, shoot. Other than having the chart too close, errors in set up result in reduced measured resolution. In my day job I work with cameras, pixels, and math. For example, we could do an automated FFT/DCT for a groups of photographed(video) lines at various resolutions and look at the frequency bins to get an empirical measurement. For such a test, we'd want to clean up via sharpening the source frames to remove as much low-frequency information as possible (improving accutance (a form of microcontrast) as much as possible). I suspect the slashcam.de test did something similar and didn't post-sharpen the 5D3 footage, resulting in an abnormally low high-frequency measurement.

 

The charts I found looked pretty solid. Which one(s) do you feel weren't done by the book?

The problem is what you print on Your printer resolution as well as probably ink density. I guess it doesn't matter that much of you're testing all cameras with your methods The problems arise when comparing to someone else who has used a different method. There has to be standards. 

 

http://www.aig-imaging.com/mm5/merchant.mvc?Screen=PROD&Store_Code=AIIPI&Product_Code=QA-77&Category_Code=Resolution-ISO-Optical-Test-Chart

Share this post


Link to post
Share on other sites

Posted · Report post

bruno & mark - good points, in a perfect world, however no test is ever perfect. Again, worst case the measured resolution will less than actual, so we know which way the trend can go. Looking at the lines of the ISO12233 chart, it's easy to see when aliasing starts whereas extinction of detail is less clear. In the case of the 5D3, I did a very crude estimate with the ISO12233 chart, and estimated 800+ lines (using a scaled still as a reference), whereas Jason's chart was full sized and shows at least 810 lines (which matches the 1620 bitmap frame reported by the ML team). I didn't see any clear problems with the other charts. Did you? Are there test results which contradict the ones in this thread (if so- link?)?

Share this post


Link to post
Share on other sites

Posted · Report post

Hi Jcs

 

Buy some proper tests charts! I'm sure you will excel in this area and become EOSHD's resident technical resolution expert.

 

The only problem then is access to cameras. :)

Share this post


Link to post
Share on other sites

Posted · Report post

jcs: People don't frame 100% the same. That immediately throws out some comparisons. They don't use the same lenses. That throws out even more. Or the same settings. Not to mention the printer resolution problem.

 

Also you can't just "estimate" that stuff like the 1DX is somehow sharper than the 5dmarkIII. I don't think it is. Does your thinking outweigh my thinking? That's why it has to be measured, to get rid of placebo and the human brain which is foolhardy. Some guys on the DVXuser forum think that the 5dmarkII is sharper than the 5dmarkIII which is just because the III lacks aliasing. That aliasing does increase "apparent" sharpness, same thing that is going on with the 1dx.

 

If you are going to do a proper resolution chart, do it properly. 

 

Alan Roberts from BBC rates the AF-100 at about 680/pvi

 

""Resolution is clean up to 63% of 1920x1080, where there are low-level null zones visible.  This 

means that the image has clean resolution only up to 1210x680, which is not good, given the 
strength of the aliasesResolution is clean up to 63% of 1920x1080, where there are low-level null zones visible.  This 
means that the image has clean resolution only up to 1210x680, which is not good, given the 
strength of the aliases."
 
 

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!


Register a new account

Sign in

Already have an account? Sign in here.


Sign In Now
Sign in to follow this  
Followers 0