SD1 Digital SLR Camera 46Mp 1.5x crop Foveon X3 sensor (4800 x 3200 x 3 layers)

[Bart_van_der_Wolf]

you can do that yourself

I know, that's why I wondered if Bob had, and how he came to his conclusion ...

When I downsample an A900 image, even with just regular bicubic resampling (which is not optimal either), I get a much better image quality than the DP1 produces at its native resolution (which should be very obvious due to the amount of oversampling, less so with the larger pixel count of the SD1). Maybe my Photoshop does something different than Bob's, or he has a different example which reacts differently to downsampling?

Cheers,
Bart

[deejjjaaaa]

I know, that's why I wondered if Bob had, and how he came to his conclusion ...

When I downsample an A900 image, even with just regular bicubic resampling (which is not optimal either), I get a much better image quality than the DP1 produces at its native resolution (which should be very obvious due to the amount of oversampling, less so with the larger pixel count of the SD1). Maybe my Photoshop does something different than Bob's, or he has a different example which reacts differently to downsampling?

Cheers,
Bart

don't you think that you need to take

4.6mp * (24mp / 15.4mp) = 7.2mp camera to compare something with DP1 ? do you think that if you will downsample your 7.2mp camera it will beat 4.6mp of DP1 ? in the absence of 7.2mp cameras we can try 8mp Canons like Canon 30D ( http://www.imaging-resource.com/PRODS/E30D/E30DhSLI0100.CR2.HTM ) 

here is the Canon 8mp downsampled ("Nearest Neighbor", CS5) to DP1 ( 2640 x 1760 )

http://img217.imageshack.us/img217/3576/e30dhsli0100.jpg

[feppe]

don't you think that you need to take

4.6mp * (24mp / 15.4mp) = 7.2mp camera to compare something with DP1 ? do you think that if you will downsample your 7.2mp camera it will beat 4.6mp of DP1 ? in the absence of 7.2mp cameras we can try 8mp Canons like Canon 30D ( http://www.imaging-resource.com/PRODS/E30D/E30DhSLI0100.CR2.HTM ) 

here is the Canon 8mp downsampled ("Nearest Neighbor", CS5) to DP1 ( 2640 x 1760 )

http://img217.imageshack.us/img217/3576/e30dhsli0100.jpg

I'm not sure why you chose Nearest Neighbor as the downsampling method, as it produces unfairly poor results. Using the recommended Bicubic Sharper method the results are much better. Even then the DP1 is still significantly sharper - very impressive. Can't wait until the new Sigma is released to see if it retains the same per-pixel IQ.

[deejjjaaaa]

I'm not sure why you chose Nearest Neighbor as the downsampling method

That was suggested by BartvanderWolf, not by me as the way to get the similar visual qualify (@ 100% on screen) from existing 24.4mp (6048 x 4032) bayer w/ AA as in planned 15.4mp (4800 x 3200) foveon in SD1... I was just trying to show that this way (Nearest Neighbor) does not work even for 8.2mp (3504 x 2336) bayer vs 4.6mp (2640 x 1760) foveon and 8.2 / 4.6 is > than 24.4 / 15.4... certainly if we will start to cheat by printing (and not pixel peep) it might be another story.

[BJL]

certainly if we will start to cheat by printing (and not pixel peep) it might be another story.

That's funny! The idea that it is cheating to examine images the way that they are actually going to be viewed anywhere outside forum debates, and instead the honest approach is to take a tiny fraction of the image (say 2MP out of 15MP or more), blow that fragment up to about the size of a 19"x13" print (23" monitor) at a pathetically low resolution of about 100ppi, without even using upressing as one would if making a serious effort to print so big from so few pixels.

My gut feeling is that, given that such IQ comparisons are likely to be made by looking from as close as the viewer can comfortably focus, no display resolution under about 200PPI is of much relevance as a predictor of how the final displayed images will look to the "end user". Lower res. like 150PPI can work well in practice for large prints, say 19"x13" and up, but only to the extent that such prints are viewed from a comfortable distance, 15" or more, not at "print sniffing" range.

[deejjjaaaa]

That's funny! The idea that it is cheating to examine images the way that they are actually going to be viewed anywhere outside forum debates

most of the world population do not view their images printed nowadays... you might not like it, but at best they view them resized, etc... and it is very easy for them to zoom exactly 100% at particular interesting part of the image, unlike with the printed image... and even those who print do not print everything that they still keep and look at occasionally.

[Robert Roaldi]

If it's true that prints no longer matter (or soon won't), it seems pointless to build sensors with more than 3-4 mpix, given that most monitors don't display more than about 2 megapixels or so. Actually then, two mpix would be enough but I am allowing for the occasional need to crop, even though that's frowned on by some. But maybe I'm missing the point. Maybe it's not about looking at pictures anymore.

[BJL]

most of the world population do not view their images printed nowadays

Agreed on that; but those who view mostly or entirely on-screen also mostly view at about 2MP, needing at most about a 4MP sensor with a Bayer CFA. I doubt that most screen viewers (outside of us forum geeks) care much about scrutinizing crops to 20% or less of the total image area.

P. S. I just read Robert Roaldi's reply: I could have left it at that!

P.P.S There is none thing that screen viewing can reveal but prints miss: dynamic range, which is rather pathetic in prints. That is why about 200PPI or "50%" on screen makes sense to me for assessing IQ, which allows significant zooming in to only part of the image.

[deejjjaaaa]

If it's true that prints no longer matter (or soon won't), it seems pointless to build sensors with more than 3-4 mpix, given that most monitors don't display more than about 2 megapixels or so. Actually then, two mpix would be enough but I am allowing for the occasional need to crop, even though that's frowned on by some. But maybe I'm missing the point. Maybe it's not about looking at pictures anymore.

certainly you are missing the point - people do zoom and look at certain parts of the image @100%... and they can do that w/ one click...

[hjulenissen]

Have you tried downsampling your A900 6048 x 4032 pixel images down to 4800 × 3200 with the Nearest Neighbor resampling method in Photoshop? That should give a reasonable impression of what can be expected.

Nearest neighbor downsampling should give you:
1. A very "sharp" aliased image
2. No noise suppression

I dont see the relevance of using that kind of resampling for comparing a high-resolution to a low-resolution camera. I'd rather use something that:
1. Does some spatial averaging (reduces noise and aliasing)
2. Coefficients could either be equal (rectangular filter) to simulate larger pixel sites, or somewhat smoothened to simulate lower cutoff OLPF
3. Allow negative coefficients or only positive? That depends. A camera sensor + optics will only have positive coefficients, so for simulating that you should have no negatives. But one benefit of having a high-rez sensor is that you can do processing in software instead of hardware...

I think that is more fair to the high-resolution camera. Now, I guess it all depends on what one wants to compare, and this comparision may tend to be an apples and oranges one.

[Bart_van_der_Wolf]

Nearest neighbor downsampling should give you:
1. A very "sharp" aliased image
2. No noise suppression

I dont see the relevance of using that kind of resampling for comparing a high-resolution to a low-resolution camera.

While downsampling with the nearest neighbor method is certainly suboptimal (as I've demonstrated here), it is a quick and dirty method to visualize aliasing artifacts of a camera without optical low-pass filter (OLPF) without the risk of false color demosaicing. It is not going to give exactly the same result as the Foveon sensor will produce, but it does create a sharpness/crispness that's somewhat similar, and which seems to be in a large part the appeal of Foveon images to many observers.

I've attempted to keep the suggested approximation method simple, so others with modest means can individually reproduce the effects. One can certainly create a more elaborate/accurate approximation by using a different downsampling method/filter but that seems a bit of overkill, and may give the larger image an 'unfair' benefit. Afterall, significant oversampling and optimized downsampling is in most cases going to give better results than sampling at a lower frequency (plus upsampling) without OLPF.

Cheers,
Bart

[hjulenissen]

While downsampling with the nearest neighbor method is certainly suboptimal (as I've demonstrated here), it is a quick and dirty method to visualize aliasing artifacts of a camera without optical low-pass filter (OLPF) without the risk of false color demosaicing. It is not going to give exactly the same result as the Foveon sensor will produce, but it does create a sharpness/crispness that's somewhat similar, and which seems to be in a large part the appeal of Foveon images to many observers.

I see that you are trying to help people do this for themselves.

But in the interest of doing it "right" (even if that makes it less accessible), I would suggest the following:
A OLPF-less Foveon sensor should (ideally) integrate an unweighted area of dimensions NxM and sample the photon count. We would like to see what that leads to in terms of IQ,

1. Find a good high-resolution image with pleny of details up to the Nyquist limit, but few debayering artifacts or moire.
2. Optionally, do image blurring (using e.g. Gaussian blur)
3. Do a integer decimation of either 2x2, 3x3 etc by averaging a neighborhood

By doing/not doing step 2, one could get an idea of what the benefit/drawback of OLPF would be for a Foveon sensor.

Of course, this depends on a range of idealised models. I assume that the accumulated blurring from motion-blur, optics and OLPF might as well be modelled by a Gaussian, but real-life may not be that simple.

I think that adhering to/breaking the Shannon-Nyquist sampling theory can be explained (somewhat exaggerated) by the following analogy:
When using a topography map to navigate the terrain. Would you rather that each elevation reading represented:
A) An infinitely small point in the centre of each square
B) A flat average covering exactly the representation square
C) A weighted average that is most depending on the center value, less to the edges of the square, but also of values outside the square

A) Means no prefiltering, maximum acuteness and maximum aliasing/moire. It also cannot be implemented in real sensors. C) can mean optimal sampling in the Nyquist sense, but that cannot be implemented fully in a sensor either. B) Is what ideal OLPF-less Foveon sensors would do, and contains characteristics of both A) and C). OLPF-equipped Foveons would be somewhere between B) and C).

-h

[jduncan]

I will buy an SD1 if its imager performance is in the same bracket as my DP1, or better.  I always found the pixel level appearance of shots with the DP1 to be superior to virtually anything else, it just lacked for ultimate resolution, and around 30mp bayer equivalence would make it a compelling choice.  It won't replace my H4D-50, of course, but it may give my Sony A900 a run for its money. 

Quentin

The funny part is that the new sensor will provide above 30 actual megapixels in a 36 x 24 sensor. Maybe is too late for sigma, or maybe the features are not there.  I  am inclined to believe that the foveon sensor will give diminishing returns as sensor density increases. But let's way and see.
I eager to see the pictures this sensor can do.