Luminous Landscape Forum
Equipment & Techniques => Digital Cameras & Shooting Techniques => Topic started by: Guillermo Luijk on December 02, 2017, 07:56:12 am
-
I have prepared an infography to try to demystify the importance of sensor size regarding Dynamic range, vs sensor technology which has a much broader influence.
To do it I have taken the Sony A7R III sensor's dynamic range as a reference, and I have calculated which crop on this camera would yield the same SNR for a given output size than the rest of most popular FF cameras (DR figures taken from DxOMark).
Not meaning this crop can substitute the cameras under comparison by any means, to begin with the crops have lower pixelcount. It's just a visual way to represent in terms of sensor size (crop) what improved technology can bring to us.
So the interpretation is: a 0,9Mpx crop on the Sony A7R III would produce the same visible noise in the shadows as a complete Canon 6D II capture once both images are scaled to the same output size.
(http://www.guillermoluijk.com/misc/sensorsizesvsa7riii.png)
The table includes APS and M4/3 sensors, ranking from higher to lower pixelcount of the crop:
(http://www.guillermoluijk.com/misc/sensorsizesvsa7riiitabla.png)
The statistics to do the calculations is the usual quadrature binning assumption: rescaling 4 pixels into 1 doubles SNR, i.e. increases DR by 1 stop. I translated DxOMark's DR differences into linear scaling and hence equivalent crop factors for each compared camera.
Regards!
-
Thanks for the comparison, Guillermo.
If similar results were also to apply to SNR at 18%, then that would be absolutely fantastic. ;D
But I guess that would not be possible. It would be against the laws of Physics. ;)
-
You may want to have a look at my recent blog post here:
https://wp.me/P9yiea-19
-
Thank you for this visualisation Guillermo!
-
Sounds like unfounded BS to me.
-
Sounds like unfounded BS to me.
just the math assuming random noise (not FPN for example) and no fine details to be preserved ( works like a charm on uniform surfaces - think a colored patch on a target for camera profiling where we can bin everything in a patch into 2x2 sensels before demosaicking to a single "RGB" pixel ;D - no details necessary there, just the "color")
for example http://www.photonstophotos.net/Charts/RN_ADU.htm#Canon%20EOS%206D%20Mark%20II_14,Sony%20ILCE-7RM2_14
we can see that @ ISO 100 (nominal values dialed on camera) ~2 stops better /lower/ noise (0.62 vs 2.87) = ~4 times lesser (1.537 vs 7.311) noise in DN (digital numbers in raw files)
S/N when you average two = (S1 + S1) / sqrt (N1^2 + N2^2) = gains ~1.4 increase when you trade resolution
-
Thanks for the comparison, Guillermo.
If similar results were also to apply to SNR at 18%, then that would be absolutely fantastic. ;D
But I guess that would not be possible. It would be against the laws of Physics. ;)
It looks like Guillermo is using engineering dynamic range as a basis for comparison, which is not surprising since he is an engineer by profession. This DR is heavily dependent on read noise. SNR at 18% is largely dependent on shot noise. Between these two criteria, is photographic dynamic range (PDR) and using PDR would level out the playing field.
Bill Claff publishes PDR for various cameras and his determinations for selected cameras are shown below. Using PDR the differences are less pronounced. Canon has upped their game with the 5DM4 as compared to the 5DM3.
Regards,
Bill
-
...and thank you, Bill, for this clarification.
-
If similar results were also to apply to SNR at 18%, then that would be absolutely fantastic. ;D
But I guess that would not be possible. It would be against the laws of Physics. ;)
SNR at 18% depends only on photon statistics, not on sensor noise (negligible at this light level). SNR at 18% is entirely determined by the "full well" capacity of the sensor, summed over all pixels. SNR scales with sensor size -- cropping a given sensor to 1/4 the number of pixels will reduce SNR by a factor of 2.
-
Thanks for the comparison, Guillermo.
If similar results were also to apply to SNR at 18%, then that would be absolutely fantastic. ;D
But I guess that would not be possible. It would be against the laws of Physics. ;)
SNR18% would yield much smaller differences. I consider the DR parameter of much more interest in the real world though, since in most cases when noise ruins an image is because the contrast of the scene forces us to use the sensor at low exposure values, and that means read noise (DR), not photon noise (SNR18%). "My pic gets noise in the shadows", sounds familiar right?. DR is deeply related to technology, and this is what I wanted to point: size is not so important as technology regarding noise (just look at the mediocre 6D's performance).
You may want to have a look at my recent blog post here:
https://wp.me/P9yiea-19
Great analyse indeed Emmet! what package did you use? R? Python?.
Regards
-
SNR18% would yield much smaller differences. I consider the DR parameter of much more interest in the real world though, since in most cases when noise ruins an image is because the contrast of the scene forces us to use the sensor at low exposure values, and that means read noise (DR), not photon noise (SNR18%). "My pic gets noise in the shadows", sounds familiar right?. DR is deeply related to technology, and this is what I wanted to point: size is not so important as technology regarding noise (just look at the mediocre 6D's performance).
Great analyse indeed Emmet! what package did you use? R? Python?.
Regards
I am glad you liked it, unfortunately not on this one, I actually used a Norwegian software I used to work with during my PhD. I will be uploading a new post very soon on M4/3 comparison and later some comprehensive lens analysis as well.
-
Interesting results.
The key take away for me is that Sony buying over the Toshiba sensor division was a major disaster.
Indeed, when you see how tremendously good the sensor of the D7200 is compared to FF, you can only lament that this team didn't have the chance to continue their effort and to deploy their next gen in a FF sensor.
Cheers,
Bernard