First off, sorry if it would have been better to revive this old thread
or even hijack a more recent thread. Secondly, please feel free to correct me anywhere. My assumptions/understanding may be wrong entirely!
I (think I) understand Guillermo's posts on the topic of ISO-less sensors, and I definitely understand the practical implications. On the other hand, it seems like a bit of work to figure out if a sensor is ISOless using the data provided by DxO (ultimately, via sensorgen) using the method he uses. I'm just wondering if the following approach is a valid and easier way to make the same determination using readily-available DxO calculations.
The intuition: the primary reason to increase sensor gain is (from a data-collection perspective rather than a viewing perspective, for which not having a dark image might be the main reason), for a fixed exposure, to have less noise in the shadows. As such, one is potentially sacrificing highlight data, or at least not making optimal use of the full-well capacity of the sensor (by maximizing SNR due to shot noise). In other words, the reason why we crank up the ISO is because we expect that when we sacrifice some headroom in the brighter part of the scene, we gain at least a little in the shadows.
So if, after applying gain to the sensor data and assuming sufficient light, the most luminous parts of the scene will be at (approximately) the same value as it would have been had the exposure been adjusted instead of the sensor gain, would it make sense to assume that a sensor's DR, for a given ISO, is limited by its performance in the darker parts of the scene? If so, instead of having to work with raw SNR data as Guillermo has done, if DR is measured by finding when SNR drops below a certain point, would it instead be easier to determine whether a sensor is ISOless using published DR measurements? Or, in other words, does sacrificing a stop from the high end buy us anything in the shadows, or have we simply lost a stop of DR?
The method would look like this: take two pairs of ISO-DR values, (ISO1, DR1) and (ISO2, DR2). If (log(ISO1/ISO2)/log(2))/(DR2-DR1) ≈ 1, then the sensor is basically ISOless for those pairs of ISOs and the lower ISO of the pair should be used for taking a shot. If, instead of being approximately 1, the left-hand side is, say, >1.3, then the sensor is not ISOless. Some values:
Canon 1D MkIV: ISO100 vs. ISO3200: 2.3
Nikon D3s: ISO200 vs. ISO6400: 2.1
Canon 5D MkII: ISO100 vs. ISO6400: 2.0
Pentax K5*: ISO200 vs. ISO1600: 1.2
Sony A77: ISO100 vs. ISO3200: 1.1
Nikon D7000*: ISO100 vs. ISO3200: 1.0
Nikon D300: ISO200 vs. ISO3200: 1.0
Leica M8: ISO160 vs. ISO1250: 1.0
The *-ed camera (sensors), I think, are ones that come up in ISOless discussions.
If this method is valid, then it looks like ISOless sensors have actually been with us for a while (the Nikon D2X, for example). For what it's worth, I shot a pair of RAW images on my D300 at the same aperture and shutter speed, varying only the ISO - ISO200 for one and ISO3200 for the other. I processed both in RPP and one of the full-resolution demosaicing algorithms, but using a +4EV compensation on the ISO200 image, and compared the output. They seemed to have roughly (visually estimated) the same amount of noise, as I would expect based on the DR values at DxOMark
. There was only a tad more chroma noise in the ISO200 image.