The charts are interesting, but the difference in RGB Luminance values for both the 40D and 5D are small between ISO 100 and ISO 640. I wonder what the real world impact is for that type of noise level.
A couple years ago I suggested a test of noise very similar to what AdamJ suggests in his CPG post. It was suggested that measuring noise in an intentionally underexposed image was not an accurate way of measuring. I forget the logic behind the argument, but at the time it seemed presuasive.
I wonder if a better test might involve a properly exposed image of an evenly lit pure solid color and then looking for variations from that color, or variations in luminosity across the image.
I'm the one who originally published that 5D chart. this is really splitting hairs, despite the various slopes in the graph. I've profiled my 5D extensively for its Lightroom default noise reduction, and with real photos there is a monotonic increase in NR values with ISO to achieve visually similar results. IE, each higher step in ISO required a corresponding higher step in NR. The increase in 5D noise with in-between ISOs is simply not visible in the real world without EXTREME manipulation (at least by my eyes). Other cameras may vary.
The thing I didn't do with that 5D chart is to relate how those arbitrary noise numbers scale in comparison to real data in a real scene. They're so small as to be completely buried.
DubRepublic wrote:
Does anybody have an idea of how where the 1dMK2n would fall on this scenario?
Yes I'm wondering about this too. My guess is that the 1D cameras are probably similar to the 30D/40D - but I haven't used one enough to say this with any certainty.
Tentacle wrote:
Wrong conclusion. The 5D simply pushes both the 1/3rd and 2/3rd stop broken ISO values. It does not use analogue sensor gain for the in-between steps because if that were the case then the noise profile over the whole ISO scale would be continuous. Now it shows dips at 'native' or full-stop ISO values.
Actually, Doug has it exactly right. The xxD and xxxD series cameras implement intermediate ISO's by software multiplication, as can readily be checked by observing the pattern of gaps/spikes in the raw data -- for instance, ISO 125 is obtained by multiplying the digitized raw data of ISO 100 by a factor 1.25, and so 25% of the raw levels will be unpopulated (similarly, a pulled exposure will bin adjacent raw levels together when the raw levels are divided by 1.25, resulting in spikes in the otherwise smooth raw histogram).
On the other hand, the 5D and 1 series cameras do not exhibit such spikes or gaps in the raw data at intermediate ISO's, indicating that they are implemented by analog amplification in hardware rather than software amplification in firmware. Measurements of camera read noise indicate that there are two stages of amplification, one for the main ISO's of 100-200-400-800 etc, and another auxiliary amplifier for the intermediate ISO's which boosts the signal by 1.25 or 1.6. This amplifier is sufficiently noisy that there is almost no benefit to using the intermediate ISO's (if one is shooting raw) -- very slightly lowered read noise (when referred to its equivalent in photons/photoelectrons) for ISO 125 and 160, and beyond that the camera does no better than a pushed exposure from the next lowest "main" ISO.
ejmartin wrote:
[...] On the other hand, the 5D and 1 series cameras do not exhibit such spikes or gaps in the raw data at intermediate ISO's, indicating that they are implemented by analog amplification in hardware rather than software amplification in firmware. Measurements of camera read noise indicate that there are two stages of amplification, one for the main ISO's of 100-200-400-800 etc, and another auxiliary amplifier for the intermediate ISO's which boosts the signal by 1.25 or 1.6. This amplifier is sufficiently noisy that there is almost no benefit to using the intermediate ISO's (if one is shooting raw) -- very slightly lowered read noise (when referred to its equivalent in photons/photoelectrons) for ISO 125 and 160, and beyond that the camera does no better than a pushed exposure from the next lowest "main" ISO.
Ok, I stand corrected, but not entirely convinced. I can't clearly distinguish between amped and pushed intermediate ISO steps. Is the existence of the extra amp a documented fact or an inferred one?
Tentacle wrote:
Ok, I stand corrected, but not entirely convinced. I can't clearly distinguish between amped and pushed intermediate ISO steps. Is the existence of the extra amp a documented fact or an inferred one?
Not stated in official Canon documentation, rather inferred from direct inspection of the raw data. As stated above, it's easy to diagnose a software push after quantization, as it leads to gaps in the raw data, and one can determine the precise factor by which the data have been multiplied by examining the pattern of gaps. Hardware amplification before digitization leads to a filled histogram with no uniform pattern of gaps (if any) in the raw data.
What would be nice would be that when threads like this come up we had a Chuck Westfall clone who could actually tell us the real, genuine, unambiguous answer, because once again, not everyone here can be correct.
Some people offer their answers as definitively correct when they just aren't (and I'm not pointing at anyone in particular here)
Here it clearly shows ISO 160 gives the lowest noiselevel on the 40D
It is important to interpret the data properly. As the ISO gain is increased, the average number of photons per raw level changes in inverse proportion to ISO. Since photons are the signal, but the raw data is denominated in quantized levels that mean different things at different ISO, is the noise going up or down relative to signal?
To illustrate the distinction, here is the electronic read noise of a 1D3 vs ISO, as measured in raw levels:
And here is the same plot, but with the vertical axis in photon equivalents instead of raw levels (again apologies, this site's messaging software doesn't understand .png's):
The point is that, while it looks as though noise in raw levels is oscillating up and down with intermediate ISO's, that is because the noise in photon equivalents is being convolved with an amplification that is changing with the ISO. In terms of photon (exposure) equivalents, the noise is monotonically decreasing with ISO, and this is what one should use to decide whether it makes sense to underexpose and push in postprocessing versus set a higher ISO in the camera. The fact that the noise plateaus at the intermediate ISO's means that there is no significant improvement in the raw data obtained from using them relative to the next lower main ISO, just a loss of highlight headroom (all this for raw shooting only; for jpegs the intermediate ISO's can be useful).
One also sees from the exposure equivalent plot that paradoxically image noise is improved for fixed Tv/Av by increasing the ISO (really a variant of ETTR, since raising the ISO pushes the histogram to the right); however, a more substantial increase in image quality results from lowering the ISO to increase the headroom, and increasing the exposure (longer Tv/wider Av) to make use of that extra headroom.
Now, what was just said applies to the 1 series cameras and 5D, which use analog amplification for intermediate ISO's. For the xxD series, intermediate ISO's are implemented in software. If I recall correctly, ISO 160 on the 40D is ISO 200 pulled by 1/3 stop in software; so what the camera is doing under the hood is to implement ETTR by applying +1/3 EC to the exposure, and then dividing the digitized data by 1.25 before writing it out to raw or generating a jpeg. That's fine, ETTR reduces noise if you have the highlight room. Personally, I'd rather have a fixed, known amount of highlight headroom for any ISO I choose to use, and then knowingly apply EC for ETTR as I see fit for better IQ, so I turn off the intermediate ISO's.
kodakeos wrote:
I wonder if ISO noise in this sense is linear with time - IE: 1/500th @ F22 and cap on would change these results.
The electronic read noise tests, which one does by analyzing the raw data from images taken with the lens cap on (and the viewfinder covered), are best done at the fastest available shutter speed (1/8000sec). There is a third component of noise, which is dark current or thermal noise, which increases with the exposure time. Its contribution is less than the readout noise for exposures of less than a few seconds duration.
Lemme see if I have it right...
40D: 1/3 pull back from full ISO stop decreases noise, 1/3 push from full ISO stop increases noise. (IE ISO 160 good, ISO 125 not so good.)
5D: 1/3 push from full ISO stop increases noise, 2/3 push from full ISO stop is even worse. (IE ISO 500 not good, ISO 640 bad.)
Conclusion: 5D, stick to the full stops, 40D, OK to use -1/3 of a stop. (IE ISO 160, 320, 640, 1250.)
How does the 1D Mark III compare? It was mentioned earlier in this thread that 1 series cameras use the same system for intermediate ISO's as the 5D, but I wonder if that changed with the Mark III release, since it seems to have a lot of other features in common with the 40D. I also wonder which pattern the new 5D will follow.
FWIW, I have been up until now sticking to the native full stops on my cams, but if there is a benefit to using the 1/3 stops on my Mark III, it would certainly be good to know.
