This is part of a long series of posts about the Sony a6300. The series starts here.
I missed making a post yesterday because I was rewriting my photon transfer function analysis code so that I could characterize a camera in fewer exposures. I’ve reduced the exposure count by an order of magnitude by going to a gradient target:
I am sampling the target with a 4 row by 6 column grid of 400×400 pixel patches. This doesn’t really give me a reduction in the number of exposures I need by a factor of 24 because of the way the double gradient is organized, but it does give me a big reduction.
You’ll notice that there are no explicit patches, just the gradient. That’s to avoid having to register the images or use some kind of machine vision code in my analysis program to find the right places on the target.
Because the photon transfer function program (which I co-wrote with Jack Hogan a year or two ago) calibrates out correlated errors by subtracting pairs of images, the gradient nature of the target is not a problem at all.
I am still working out the details of the protocol, and I’ll publish it in a week or two in what I’m sure will be an extremely nerdy post. Today I just want to show you results for the a6300, when used in single shot shutter mode.
First the full well capacity:
Full well capacity doesn’t change with ISO in the standard camera model, and, in spite of the fact that the a6300 does its conversion gain changing trick, the modeled FWC doesn’t seem to change much. It shouldn’t change with raw channel, either, and for some reason that I haven’t figured out, the blue channel seems to model out to have a higher FWC.
Also, because of the conversion gain change, the FWCs above really only apply to ISOs between 100 and 320. After the a6300 raises its conversion gain by approximately a factor of 4 at ISO 400, you should mentally divide the FWCs above by that number.
Second, the read noise. These curves are similar to the ones I got from RawDigger and posted a couple of weeks ago, but they are different in that the read noise plotted is only the part of the read noise that is uncorrelated between frames, and also that the vertical axis is in electrons, which I can now do since I know the FWC.
Because of the a6300 conversion gain changing trick, this graph needs a little explanation, too. The a6300 doesn’t really have read noise below one electron. You should mentally multiply the read noise graphed here by a factor of four to account for the conversion gain change. There’s also an asterisk associated with the point at ISO 12800. The gratuitous, non-defeatable lowpass filtering that Sony saddled this camera with kicks in at ISO 12800. This throws off the FWC modeling, which throw off the read noise modeling. I will probably leave it out in future posts, but I though some of you might be interested in seeing how it messes things up.
The modeling program also spits out Engineering Dynamic Range (EDR) data:
Again, these are very similar to the curves that I posted earlier, but slightly different tn that, by design, they ignore fixed pattern read noise.
Well done, Jim. What do you mean by:
” You should mentally multiply the read noise graphed here by a factor of four to account for the conversion gain change. ”
If you use the SNR method (as opposed to the standard deviation method) gain should not enter the picture and FWC and RN should come out correctly?
Jack
I don’t think so, Jack. My model assumes that FWC is obtained at base ISO. Looked at another way, there is one Unity gain ISO, not two. But I could be wrong; this conversion gain changing stuff makes the traditional sensor metrics confusing to talk about. I suppose I could modify the model to allow conversion gain changing, but then it wouldn’t be the model we’ve known all these years and would be confusing to most people.
Jim