In yesterday’s post, I speculated that the Claff Photographic Dynamic Range Curves created by by direct search might be smoother if the searched data set was from a simulated camera rather than a real one.
I programmed up a Nikon D4 simulator that used the modeled read noise values from the data set. I picked the D4 because it subtracts out the black point in-camera, with the potential to chop off some of the signal at low mean values, and I wanted to see if that was a problem for computing Claff PDR. I told the simulator to not truncate the histogram.
I ran the simulator to produce a spreadsheet of means and standard deviations, then analysed that with my PDR program. The results looked just about as lumpy as the real camera results, Then I changed the exposure increment from 1/3 stop to 1/10 stop, and things smoothed out:
The data point at ISO 250 is more believable in the simulated result, but it makes the point at ISO 160 look suspiciously low. The systematic discrepancies at high ISOs are because I just simulated one raw channel, whereas the data points for the real camera are an average of all four raw channels. I just used one FWC for all the ISOs.
If I take the average read noise at each ISO as my target we get a little closer:
The discrepancies at ISO 160 and 250 remain.
My conclusions are:
Most of the lumpiness in the real camera curves is caused by the coarse exposure sampling, which makes the average difference between the searched-for SNR and the found SNR larger.
It is impractical to make more exposures, but it’s desirable to sample the mean values more densely. Therefore, I will experiment with using a gradient for a target, so that I can sample several different mean values in a single exposure.
In the case of the D4, the histogram truncation performed by the camera did not affect the Claff PDR results. In the case of other cameras, it might. I should probably run simulations to check whenever I test a camera that does histogram truncation. By the way, none of the three Sonys tested yesterday do that.