F707's 5mp vs 3mp again....

[Howdy]

Thanks for the explanation, Shay. The argument is based upon green being the primary (sole?) basis for determining luminance in most interpolation algorithms. So, bobbo had TurboTed are referring to luminance as 'detail,' and from this 'fact' proposing that since a 5mp mosaic gets 2.5mp of luminance data, it is really only a 2.5 mp sensor, and thus can be treated as a 3mp sensor without losing detail.

Still don't agree with the argument, but at least I have an understanding of the basis on which the argument is being made. As I understand interpolation algorithms, the argument would be 'truer' for earlier, less developed algorithms, as they apply a straightforward 'interpretation' of the luminance data. It becomes less true for the more developed, more complex algorithms, which (as I think I understand, though I may be wrong) massage the luminance data in the context of color data and vice versa.

And if it helps, I do recognize that a 5mega-sensel (ms) mosaiced sensor (such as the 707) will have less resolution than a 5ms foveon-type sensor. Whether the 3mp foveon will have better resolution than a 5ms or 6ms mosaiced sensor remains to be seen, but I understand that the math would suggest it will be close.

But, back to the original proposition. That luminance is derived from the 2.5million green channel/sensors, and is then interpolated into a 5mp image does not lead to the conclusion that reducing the raw data to 3mp size before interpolation (which, I believe is the suggested procedure being followed by Sony in the 707, although I have no knowledge) will result in a 3mp image with as much 'detail' or luminance information as a 5mp image.

Howdy

 

[Howdy]

That is interesting, I've never considered going in a different
direction. I would propose, though, that someone who wrote an
algorythm to start at the northwest corner would in fact average
the pixels to the east, southeast and south so to avoid averaging
non-existant pixels. In this case, it would end up with
alternating red/green pixels.

The intent of the example was to (1) illustrate that an algorithm to resample down won't necessarily (in fact, will rarely) complement the algorithm to resample up.
down, and vice versa.

Perhaps, a better example, and one more relevant to the instant discussion, would be to begin with a row of 2x2 squares, alternating between 4 pure red (255r0g0b) and 4 green (0r255g0b) EXCEPT that the northwest pixel in each 2x2 square (whether red or green) also has 20 blue (ie, 255r0g20b). Reduce each 2x2 square into a single pixel, by averaging each color channel of each of the four pixels in that square. For a red 2x2 square, the result will be a singe pixel with the value of 255r0g5b. The green 2x2 squares will become a single 0r255g5b pixel. Then, enlarge this row of alternating pixels into a row of 2x2 squares. Since the only information provided by the 'parent' red pixel is 255r0g5b, each of the four pixels will also be 255r0g5b. Thus, the result of this two step reduction/enlargement, is to lose the three pure red pixels, and the one red pixel shaded slightly blue, with these original pixels being replaced by four 255r0g5b pixels. In short, a loss of detail has occurred in this simple example.

With more complex algorithms, in real world situations (and unlike the above example), repeated reduction/enlargement will cause increased degradation with every iteration. Your test #1 demonstrated, in a real world setting, that perceptable degradation does occur. Whether this perceptable loss of detail is 'acceptable,' as bobbo appears to conclude, or not is an individual decision for each photographer.

I say you can have fractional pixels. Take an image, use a
sophisticated algorythm to scale it up 20%. You'll end up with
more pixels each of which will have a value equal to its original
value and some fraction of its adjacent pixels. Technically, each
pixel contains a little more information than it used to, not more
detail but it does contain some component of its neighbors. You
now have no more information than you used to but it is taking more
pixels to display it. That is what we are saying is happening when
the RGBR mosaic is interpolated resulting in a camera's native
image size.

Interesting theory, but it doesn't comport with any of the interpolation algorithms implemented in current digicams, as best I can tell. First, interpolation of data obtained from a mosaiced 5mega-sensel ccd to produce a 5mega-pixel image is not the equivalent of re-sampling a 3mp up by 20% or 40% to 5mp. Second, reduction of the data from a 5mega-sensel ccd to produce an interpolated 3mp image will result in the loss of information. Without doubt, because the initial data set from a 5ms ccd is larger than the data set from a 3ms ccd, the 5ms ccd will produce a more detailed 3mp image than will the 3ms ccd. The fact remains, however, that the 3mp image, whether from a 3ms or 5ms ccd will have less information than a 5mp image from a 5mp camera, and that loss of information becomes perceptible in the effort to use the 3mp image as a 5mp image.

Anyway, got to run, off to go camping with my sons!

Are you avoiding my test #2??? ;-)

Have a great time camping, I can't wait until my kids are older and
we go on a camping trip.

Finally had time to do test 2. Image 2 is still the original 5mp, and image 1 is still the 3mp original resampled to 5mp, albeit now with sharpening artifacts. (Not that that is necessarily bad, it wasn't as immediately obvious on the screen, but it remains perceptible.)

And, camping was a hoot. Went to a naturally occurring oasis out in the middle of the desert with my sons and a bunch of friends. Very neat place.
Howdy.

 

[Simon Chapman]

I'm posting my own pair of test images below.

The test consisted of taking two consecutive pictures differing
only in the selected resolution, one in the F707's native 2560x1920
and the other in the F707's 3mp equivalent, 2048x1536. The camera
was set to F5.0, Outdoor WB, Infinity focus, shutter 1/400 second.

One of the images below is a 480x480 crop from the center of the
5mp image, saved at the highest PhotoShop JPEG quality and shown at
100%.

The other image started with the 3mp image original, resized in
Photoshop to 2560x1920, and then cropped and saved as above.

I would like to see this test applied to a subject which is more prone to JPEG artefacts, such as high contrast straight lines (eg roofline of a building, edge of a rectangular table, etc) and also finer lines than the examples given, eg human hair, cat whiskers, etc). I would also like to see green taken out of the equation since the F707 has a tendency to oversaturate greens and reds which would confuse the issue. Nor am I sure that infinity focus is the best base for the test.... neither of these images looks very sharp to me. And lest we forget, neither of them shows what a 5mp TIFF looks like out of the camera.

SC

 

[TurboTed]

OK, TurboTed you badgered me into it. Actually, I finally had
a moment to spare. Figures that I don't have that moment until
nearly midnight on Saturday.

OK, my first impression is that Image 1 seems to have several
distinct sharpening artifacts.

I was afraid of that, but thought I keep it as it was anyway.

Image 2 has a better 'feel.' Detail seems to arise from detail, ie
expected differences between neighbouring pixels, not artificial
differences.

So, my vote would be Image 1 started life as 3mp, then was resized
to 5mp and then sharpened (oversharpened?)

Yep, right again.

Image 2 came from camera as 5mp image.

Howdy

 

[TurboTed]

The intent of the example was to (1) illustrate that an algorithm
to resample down won't necessarily (in fact, will rarely)
complement the algorithm to resample up.
down, and vice versa.

Very true.

Interesting theory, but it doesn't comport with any of the
interpolation algorithms implemented in current digicams, as best I
can tell. First, interpolation of data obtained from a mosaiced
5mega-sensel ccd to produce a 5mega-pixel image is not the
equivalent of re-sampling a 3mp up by 20% or 40% to 5mp. Second,
reduction of the data from a 5mega-sensel ccd to produce an
interpolated 3mp image will result in the loss of information.
Without doubt, because the initial data set from a 5ms ccd is
larger than the data set from a 3ms ccd, the 5ms ccd will produce a
more detailed 3mp image than will the 3ms ccd. The fact remains,
however, that the 3mp image, whether from a 3ms or 5ms ccd will
have less information than a 5mp image from a 5mp camera, and that
loss of information becomes perceptible in the effort to use the
3mp image as a 5mp image.

Okay, I'm going to back off a bit and concede that the 3mp image has slightly less information than its 5mp big brother due at least to a difference in sampling/resizing algorithms.

Finally had time to do test 2. Image 2 is still the original 5mp,
and image 1 is still the 3mp original resampled to 5mp, albeit now
with sharpening artifacts. (Not that that is necessarily bad, it
wasn't as immediately obvious on the screen, but it remains
perceptible.)

Right, see my conclusion below.

And, camping was a hoot. Went to a naturally occurring oasis out
in the middle of the desert with my sons and a bunch of friends.
Very neat place.

Sounds like a neat place, I think it would take me a weekend just to drive to a desert.

CONCLUSION: I guess in hindsight this test was doomed to fail. Since we can ALL agree that the 3mp image will never have MORE information than the 5mp image, it would take a perfect combination of resizing/sampling and sharpening and anything else just to match the 5mp image. But as I believe you stated, it took a close look, one closer than I think it will take once the Foveon chip arrives on the scene. I think we'll see the 3mp Foveon give the 5mp CCDs some stiff competition in detail resolution. We'll also see that these users will lose substaintially more detail in their images if they choose to take pictures in 2mp size.

Having said that, I'm going to continue to use the 5mp size for all my shooting except where I'm really, really short on space. In this case I feel better knowing now that dropping down to 3mp on the F707 is pretty darn close to the 5mp photos.

 

[Howdy]

CONCLUSION: I guess in hindsight this test was doomed to fail.
Since we can ALL agree that the 3mp image will never have MORE
information than the 5mp image, it would take a perfect combination
of resizing/sampling and sharpening and anything else just to match
the 5mp image. But as I believe you stated, it took a close look,
one closer than I think it will take once the Foveon chip arrives
on the scene. I think we'll see the 3mp Foveon give the 5mp CCDs
some stiff competition in detail resolution. We'll also see that
these users will lose substaintially more detail in their images if
they choose to take pictures in 2mp size.

I don't think the test failed at all. First, it demonstrated to many people how well the 707 performs in 3mp mode. That's good to know. Might come in handy if I'm short on memory, or if I know the shot is intended for a 4x6 print, or as a screen saver. Second, it demonstrated that there is a difference in quality between 3mp mode and 5mp mode. That's good to know too. Your test, and bobbo's, provided good information and allowed people to reach their own conclusions. I don't consider that a failure at all. Indeed, seems quite successful to me. I'm glad you took the time. I know I learned a good deal, and feel more comfortable using 3mp mode.

Those 3mp Foveon's are going to be interesting, adding quite a new 'dimension' to digital photography at a 'prosumer' level. Before this, getting full rgb info for each pixel meant spending multi-multi-thousands to get a studio camera. I wouldn't be suprised at all to see a foveon 3mp sensor equalling, or even bettering, the resolution of the 707 5mp ccd.

Having said that, I'm going to continue to use the 5mp size for all
my shooting except where I'm really, really short on space. In
this case I feel better knowing now that dropping down to 3mp on
the F707 is pretty darn close to the 5mp photos.

I couldn't agree more. Again, thanks for taking the time to explore the issue.

Howdy

 

[Shay Stephens]

I believe the 3.4MP will be better than the F707 when it comes to resolution. It will probably perform as good resolution wise as a 6.8 MP mosaic sensor, and beat it when it comes to color and artifacts. Lack of noise looks to be a strong point with it as well.

 

[chiquinho]

Hi!

I'm a newbie - so I´m sorry if this question sound stupidly.

What's the best way to save MS Capacity. Shooting in 5 MP Standard Res. or better stay with the 3 MP Fine Res. ?

thx
chiquinho

 

[Peter Epstein]

I tested this on my 707 and found that 5 MP Standard was superior to 3 MP Fine. You can do your own test to see if the same result holds for your camera.