There is one point about the DXO dynamic range calculation as compared to engineering DR that has always confused me. A common method of determining the read noise of a sensor is to put the lens cap on and take a dark frame. Read noise and shot noise combine in quadrature and with the lens cap on there is no shot noise and the read noise is the output of the sensor (provided that the read noise is not clipped as with Nikon cameras). A SNR of one indicates that some signal is present. Does the signal refer to the input or output of the sensor?
That's how I do it if it needs to be done accurately. I even exclude the possibility of lens electronics interfering, by using the body cap. I also cover the viewfinder in order to prevent light leaking into the mirrorbox. I also use the shortest possible exposure time (1/8000th sec.) to reduce temporal/thermal noise build-up. To also eliminate pattern noise, one should subtract two such black-frame exposures, and correct the StDev by dividing by Sqrt(2).
Since we cannot know if and how analog gain is employed, we'll have to empirically/statistically determine the input signal in electrons (e-) from the sensor output in ADU or DN, corrected for the gain. However, this will only work if the Raw data (ADUs) is not clipped
to (presumably) zero by eliminating any Blackpoint Offset. For example Canon cameras allow to analyze the full black-frame noise, and e.g. Nikon cameras typically clip the black-frame noise. In addition, some manufacturers (e.g. Sony) have started adding additional processing to record the ADUs with a non-linear response-curve in some of their camera models.
This all complicates testing (it's not practical to switch testing methods between models), and is part of the reason that DxO use a somewhat different method, at least that is what I have read between the lines. Apparently
, they use the least exposed parts of the SNR curves to extrapolate and determine the black-point at dB=0 (SNR=1, because 20*Log(1) = 0 dB) or, sometimes if that is not a typo, dB=1 (SNR=1.122, because 20*Log(1.122)=1 dB). They do also calibrate for actual exposure by measuring the actual influx of light, and ISO sensitivity. They seem to use a lens on the bodies, but they only shoot one exposure level at a time (by blocking the other filters in their setup), so glare should not play a role.
My findings with the laborious method of subtracting pairs of images and deriving gain from the slope of the response curve to calculate electron input, corresponds to within approx. 0.1 EV with the DxO findings on Dynamic Range.