ISO on the R5

Quarkcharmed wrote:

John Sheehy wrote:

Quarkcharmed wrote:

With all due respect to Bill's analysis, there must be gaps in the raw histogram which I don't see, so potentially it's more complicated than just digital multiplication.

Say if you multiply ISO 400 values to get ISO 500, you have to multiply each value by 2^(1/3) = 1.256, basically 5/4.

With integer digital numbers there will be gaps each 5th value - it won't be contributing to the raw histogram and we'll see the regular gaps.

A camera could do "corrections" differently for different lines of pixels, and fill a histogram that way, too; not just by using analog gain only. That's why you need to look at highlight headroom, too, to determine what is being done.

I suspect such corrections may easily introduce moire in the shadows. It can also be done by adding some dithering, but that will be adding the noise. Such additional corrections would just add to processing time with very little benefit - or am i missing something?

I am talking about rescaling the RAW values; not adding or removing noise.

Quarkcharmed wrote:

bclaff wrote:

Quarkcharmed wrote:

bclaff wrote:

MarshallG wrote:

Disclaimer: I’m only the messenger.

I’m at a trade show all weekly so I stopped in the Canon booth and spent a long time talking with one of their experts about the R5.

He then me that the R5 sensor slightly boosts voltage as you increase ISO,

Not slight but in proportion to the ISO setting.

and it did so in one stop increments.

True of many Canon bodies including the R5

Intermediate ISO changes are made post sensor read,

True of many Canon bodies including the R5

How do we establish that? I don't see it in rawdigger. With the interim ISO values (e.g. 125, 160, 320, 500 etc.) there are no regular gaps in the histogram.

4ex:

There are strange regular 'dips' though but they present at any ISO setting including 100, 200, 400

Anyway digital amplification should have produced much more frequent regular gaps with bins at zero.

It not so obvious as with older Canon bodies but you can see the pattern in the Read Noise

I added the red lines to show how every whole ISO setting is the bottom of a group of three that have a straight line. The pattern is most obvious at low ISO settings.

Right, that's very interesting. It's prominent before ISO 1600 although a bit debatable after 1600. But maybe it's just the small scale that hides it.

However the question is, why doesn't it show in the raw histogram? Say if ISO 500 is ISO 400 digitally multiplied by 5/4, we should see clear gaps every 5th bar.

I suspect some mild filtering or dithering which hides the gaps you might expect to see.

Quarkcharmed wrote:

It could be that, but then it's not 'digital', and the Canon's representative from the OP said it was digital multiplication between the 1-stop ISOs.

I would not expect a public-facing rep to know the difference between 1/3-stop gain granularity in the first amp, "digital" (integer math) pushes, and 1/3 and 2/3 stop gain from a secondary amp. I would have to talk to that rep enough to know if he understands all these possibilities before I would assume that he knew what he was talking about.

Of course, it may be the case that with a combination of HTP, lighting with one weak color channel, and e-shutter, that ISO 160 may have more visible read noise than ISO 200 in shadows.

MarshallG wrote:

Well… my eyes tend to defocus and glaze over on the math, but mostly because a 50,000 foot view of the data shows a basically flat line, nothing spiky.

It's nice to see the data, and a nearly-straight-line based on standard deviations may prove a limited difference, but standard deviations have nothing directly to do with visible noise. It is a hollow currency which depends on the noise character being the same always (usually assumed to be 1/f), only varying in intensity.

Imagine that you are laying under a glass table, looking up. Somebody comes along with bottles of black, white and grey sand. They poor the sand into little piles, one at a time so that there are clusters of black or gray or white, with dozens of each clumped together. You clearly see the "noise" of these B&W grains. Then, they take their fingers and keep mixing all the sand together, until it all looks grey to you. What happened to the standard deviation of the sand? Nothing. What happened to the histogram of the sand? Nothing. Yet, as far as you can see the noise is mostly gone after it is all randomly mixed. This is a removal of spatial correlation.

Standard deviations do not describe noise in a spectrum, which is where we SEE noise.