Great Bustard
wrote:

FrankyM
wrote:

Well, I guess this is where the diversion into semantics begins. How about this: you don't
compare
(in terms of the final photo) a single 2x2 pixel to a single 1x1 pixel -- you compare to four 1x1 pixels.

SNIP

For the record, downsampling is a
horrible
way to compare the IQ
potential
of two systems -- the proper method is upsampling one, the other, or both, to the same display dimension.

However, if the purpose of comparison is web display, then, of course, downsampling is the proper method of comparison, since the final photo is necessarily downsampled.

And, yes, the method of resampling is key. Using a "nearest neighbor" downsampling method is a bad choice, as is using an upsampling method that merely scales, rather than interpolates.

You could argue that nearest neighbour retains most
true
detail of any method. But my point is that if we are discussing images, rather than sensor data, the post capture processing is rather important.

SNIP

DxOMark uses the 100% NSR for the noise floor (DR100), which, by definition, includes photon noise. If you wish to use a different noise floor, that's entirely your preogitive, but, like DxOMark, you need to clearly spell out what noise floor you are using. In addition, DxOMark gives a "screen DR" (DR / pixel) and a "print DR" (DR / pixel of photo resampled to 8 MP).

The point is, to me, as a photographer, it isn't interesting.

Well, if you have no idea, then you have no idea. But I can tell you,
for a fact
, that more pixels for a given sensor size and efficiency results in more IQ all the way around (although this is subject to diminishing returns, of course). The only question is if the pixels can be made smaller without adversely affecting efficiency. However, the
overall
trend is that pixels have been getting smaller
and
more efficient. Of course, that's not to say that when a new technology comes out, that it might not have to begin with larger pixels.

I said I have no idea whether or not Canon has the capability to make the tech with smaller pixels. Do you have a mathematical proof for this? I would be interested if you do.

I can mathematically prove that smaller pixels do not result in less DR / pixel for equally efficient sensors (but greater DR / area), but not mathematically prove that smaller pixels can be made with the same efficiency. Please let me know if you want me to do so -- I'll be happy to oblige.

But I cannot mathematically prove that smaller pixels can be made as efficient as larger pixels. However, I can cite evidence, that,
as a general rule
, as a function of time, pixels have gotten smaller and more efficient.

What I'm interested in is mathematical proof of your statement above (academic interest only).

".. I can tell you,
for a fact/, that more pixels for a given sensor size and efficiency results in more IQ all the way around."/

However, I note in your reply above that you agree with Dr. Martinec's LL post:

http://www.luminous-landscape.com/forum/index.php?topic=42158.0

Excellent! It is basically a discussion of what DR measure you wish to use (DR100, DR50, DR25, etc.). However, as I said, more useful still, in terms of the visual properties of the final photos, is to compute the DR / area rather than DR / pixel.

Yes, I certainly do. But as I stated above, DR to a photographer is not the engineering DR
of the sensor
since it is affected by other things, not least of which is what the photographer finds acceptable. Martinec states;

"A final caveat: The useful DR to a photographer can be limited by more than indicated by the S/N figure of merit; for instance Canon DSLR's have a base ISO plagued by a lot of pattern noise in shadows, which can be visually much more objectionable than the random grain of unpatterned noise while not showing up in the noise standard deviation. On the other hand, pattern noise seems very well controlled on the D3x. The pattern noise can limit the useful DR -- how much one is willing to push shadows -- more than might be indicated by the S/N graphs."

I guess we've hashed this to death.