I’ve been working on improving the presentation of the SNR versus ISO setting data, and on improving the precision of the data itself.
On the latter point, the most effective thing that I’ve done to remove systematic errors is to keep the shutter speeds constant throughout a series of exposures, and make all the required exposure changes using the lens aperture. Less exposure range is possible when using the aperture than when using the shutter, but f/2.8 to f/22 gives me enough range to encompass the whole ISO range of interest, from ISO 100 to ISO 6400. The “artificial” low ISOs on the Nikons are a casualty, but I see no use for them anyway. In an effort to increase repeatability, I’ve started making 16 exposures for each data point, and averaging the results from those exposures. As a side benefit of that approach, I can compute the standard deviation of the underlying single-exposure data for each data point, and plot that information on the graphs to give an idea of the possible sampling errors involved.
With respect to the presentation, I’ve made a few changes.
The first is how the signal-to-noise ratio (SNR) is expressed. I had been using decibels, but that way of looking at ratios may be unfamiliar to many photographers. I wanted to keep using a logarithmic measure to allow the large spreads of data to be plotted on the same graph, and to preserve the easy to interpret nearly-straight lines that result when a logarithmic vertical axis is combined with a logarithmic horizontal axis. The log-log presentation is also familiar to photographers from the H&D or D-logE curves of the film era.
I considered measuring the SNR as density ratios. This would have some familiarity to greying photographers, but would probably confuse those who never used film even more that decibels. I finally arrived at what I consider to be an excellent solution: I will present the SNR data in f-stops. Mathematically, this is a logarithmic measure. Since the f-stop difference equals log base 2 of the ratio. The straight lines of the decibel plots are preserved. All that changes is the slope, and with the right selection of the vertical axis divisions, that can be fixed. Photographers from the old and the new eras are both used to saying things like, “That highlight is 5 stops brighter than that shadow,” so the transition to, “The signal is 5 stops greater than the noise,” shouldn’t be too difficult a transition.
I decided to drop the nearly-horizontal line that indicated the SNR that would be achieved it the analog-to-digital converter (ADC) resolution was the limiting factor in the SNR. In the cameras that I’ve been testing, at the raw bit depths I’ve been using (14 bits, for the most part, although the Sony cameras take some liberties with this when they compress the raw files – and you can’t turn that off), ADC resolution is far from the limiting factor. Also, should you ever care, it’s pretty easy to imagine a straight horizontal line, which is all you need thanks to the next change.
I had been including the SNR data uncorrected for the small exposure errors made by the camera, preferring to show the two guidelines as corrected instead. I thought it a benefit to see the exposure errors directly. I no longer think that’s of any use, so I’ve corrected the data instead.
I’ve now included results for the red, blue, and the average of the two sets of green pixels, rather than just the one set of green pixels that I previously measured. So far, I haven’t learned much from this additional data, but I may find differences in the behavior of the different pixel groups as I test more cameras. I have not corrected the SNRs for the difference in exposure between the three groups of pixels. I could go either way on this, and I’ve followed the rule to not manipulate the data without a clear purpose here. I may revisit this decision.
Now that the data is corrected, there doesn’t seem to be much use for the 1.5 decibel/octave line, which would have to be relabeled the half a stop per stop line thanks to the change to the vertical axis units. With many of the cameras that I’ve tested the slope of the actual data is very close to that slope, so that the added line just gets in the way. In all of the cameras I’ve tested, the slope of the data at the high ISOs is half a stop per stop, indicating that increasing the ISO setting is not giving any advantage over cranking up the exposure slider in Lightroom.
Here’s an example of the new SNR presentation for the Nikon D4:
I’ve also added a new graph that shows departure of the data from the half stop per stop line. In this graph, as in the one above, the mean, mean-plus-two-standard-deviations, and mean-Minus-two-standard-deviations lines are plotted. If the data is Gaussian, for a large sample, less than five percent of the results lie outside of the two lighter lines. Here’s what that looks like for the D4:
This shows the (pitifully small, in this case) advantage in post-Lightroom SNR to be gotten by cranking up the ISO knob on the camera: less than 0.2 stops from ISO 6400 to ISO 100.