Bob, here's an example. At very high ISO, the NEX5 has a stop or more advantage over the G2/GF1/E-PL1/E-P2 (all featuring the same sensor). You can see that by looking at these side-by-side crops (RAW conversions with color noise removed):
This was a predictable result based on the DxO data:
If you look at the rest of my comparison images
here and compare them to the DxO results, they correlate very well. You can download my RAW files from that page, process them in your RAW processing app of choice, and I think you'll find that the DxO results still hold up.
DxO can be misleading though. For example, if you download my RAW files, you'll see some ugly banding in the GH1 files, which has a real effect on image quality. This is completely ignored by DxO, as are several other parameters of image quality including lens performance, JPEG engine, NR applied at the RAW level, susceptibility to aliasing, color fidelity, etc.
Amin,
I think the last point you make is the real point combined with the raw processing that one performs to the image.
<rant annoyance="high">
Here is another characteristic: the pattern of the noise.
Many times noise appears is bands or clumps or some other pattern that makes it more obvious than s/n measurements. Just to set the record straight, I am an engineer through and through and eat measurements for breakfast. I just happen to think that the DxO measurements are not only inadequate but misleading.
Even many raw files have been in-camera processed to one degree or another so it is not so much sensor but the total image chain.
I have to confess that for me, ISO 50 is here I usually shoot and consider iso 800 to be sort of the outer limits. If you want to really understand the performance of a sensor then it ought to be done at the base sensitivity of the sensor itself. Reading actual sensor data sheets is often instructive, since many of them, or almost all of them actually, have only ONE sensitivity, with the higher sensitivities synthesized by shifting the data and reducing the actual number of information containing bits. Sometimes zeros are forced into the low order bits, sometimes whatever stuff happens to be produced by the a/d converter. In any case the number of luminance levels (bits of significant data) is reduced resulting in banding or other visual manifestations dependent on the scene.
Mostly the what the DxO numbers OUGHT to be can be derived from sensor data sheet values. Where they differ is a result of firmware, post sensor processing chain, and measurement artifacts. The true test is in the taking. If you start with a iso100 base sensitivity 14 bit (really hard to do well) sample and create an ISO 3200 file, it is not going to have more than about 9 bits of data with the low order bit of those nine just as noisy as the 14th bit of the base ISO sample.
Another annoying bit...
There is no such thing as color noise.
So far bayer image arrays all produce monochrome sample data filtered to red green and blue (admittedly with different band pass depending on sensor model). usually there is twice the number of green pixels than red or blue. There is thus one more potential bit of data available for luminance information, but not on every pixel, on the collection of near neighbors that are included in the debayering algorithm. This difference is sometimes processed as "color" noise which is more an artifact of the image processing algorithm.
So give me a break, show me the images and the ways that they differ for use by YOU in the ways that you care about.
The imposition of a pseudo-scientific measurement methodology for the comparison of sensors does more to obfuscate than to enlighten IMO.
and finally, their characterization of qualities such as color depth as relating to use in portraiture and dynamic range a landscape is downright idiotic. When I shoot a model, I want good depth and texture in the darks of brown irises and the specular highlights of glossy skin and when I shoot a landscape us old timers have been dealing with techniques to control image dynamic range from the zone system through multiple exposures at different EVs and selective layering but we still want all the color nuance we can get.</rant>
-bob