Graeme Nattress
RED Problem Solver
SO yes, in linear sensor data there are more code values per stop in each and every brighter stop. The number of code values per stop doubles on each increase of stop brightness. However, this is NOT why we want to use the brighter stops. We use the brighter stops because sensor read noise is the predominant form of noise in the raw image that matters. Because sensor read noise is essentially a fixed value that doesn't vary with the brightness of the pixel read, read noise predominates as a source of visible noise the darker the value you're reading. Say we're talking about a sensor that has 1% read noise and we read a really bright value pixel, say one at 90%. 1% read noise out of a 90% value is not going to be noticeable. Now let's read a really dark value of 5%, and our read noise is still 1%. Now read noise is contributing to 1/5th of the signal we've read.
Unfortunately some photographers who promulgated ETTR did so for the wrong reasons. They knew about images and linear light from the sensor, and knew an ETTR exposure lead to a better image, and they correlated that to the code values per-stop. But they didn't know sensors, and didn't grok that it's read noise we're trying to avoid by exposing to the right. In other words they gave the right advice for the wrong reason.
The other source of noise is shot noise which increases as the number of photons that hit the pixel. In other words shot noise predominates at bright exposures. This means that although you have vastly more code values at brighter stops, you also have vastly more shot noise. But as shot noise scales in line with brightness, we hardly ever see it as a source of image noise as it's fairly distributed per stop. What we see is read noise when we try to look at the darker stops. Because of read noise it makes no sense to have a great bit depth on the AtoD of the sensor than is necessary. You could have a 32bit AtoD and give umpteen code values to the darkest stops, but because read noise predominates you're wasting all those extra bits and super-expensive AtoD on recording read noise! As long as the bit depth is enough so as quantization noise is lower than read noise, you need not go any further. Because most practical AtoD are linear, this leads to an over-abundance of code values in the brighter areas of the image, but again, we're not doing ETTR to use these code values, we're doing ETTR to reduce the visual appearance of read noise.
Graeme
Unfortunately some photographers who promulgated ETTR did so for the wrong reasons. They knew about images and linear light from the sensor, and knew an ETTR exposure lead to a better image, and they correlated that to the code values per-stop. But they didn't know sensors, and didn't grok that it's read noise we're trying to avoid by exposing to the right. In other words they gave the right advice for the wrong reason.
The other source of noise is shot noise which increases as the number of photons that hit the pixel. In other words shot noise predominates at bright exposures. This means that although you have vastly more code values at brighter stops, you also have vastly more shot noise. But as shot noise scales in line with brightness, we hardly ever see it as a source of image noise as it's fairly distributed per stop. What we see is read noise when we try to look at the darker stops. Because of read noise it makes no sense to have a great bit depth on the AtoD of the sensor than is necessary. You could have a 32bit AtoD and give umpteen code values to the darkest stops, but because read noise predominates you're wasting all those extra bits and super-expensive AtoD on recording read noise! As long as the bit depth is enough so as quantization noise is lower than read noise, you need not go any further. Because most practical AtoD are linear, this leads to an over-abundance of code values in the brighter areas of the image, but again, we're not doing ETTR to use these code values, we're doing ETTR to reduce the visual appearance of read noise.
Graeme
