High-End Digital Vs. 4x5 Film

[shileshjani]

[QUOTE=Nitish Kanabar]
So up-sampling the image DOES produce new data/pixels, if this reasoning is correct.

New data, yes. New information, No.

[Nitish Kanabar]

New data, yes. New information, No.[/QUOTE]

What is the difference? In my first example, two gray pixels were added - isn't that new information?

[Ralph Barker]

Nitish - in my view, you are correct. Upsampling (obviously) creates new pixels, because the total number of pixels increases. The soft/firm ware for doing so looks at the adjoining pixels and makes a best guess at what the color value for the added pixels should be. I believe what shileshjani meant by, "New information, no" was that the added pixels don't contain additional real detail.

[Dan Wells]

I performed an interesting (and purely visual) inadvertant test on my own equipment. Neither my digital nor my film setup is up to what these guys are using, but it gives some idea. I use a Nikon D200 (brand new - I was actually primarily playing with the camera to test it), along with a Hasselblad and a Horseman VH "baby view" , which I chose because I have a Nikon 9000 scanner that likes 6x9 cm, but not 4x5. I just printed a couple of prints that seem sharp to my eye, but upon VERY close inspection may be reaching the limits of their format. Both were printed on the Epson 4800, which is a frighteningly sharp printer. The Nikon print is 15 inches in the short dimension and 22.5 the long way, and I feel that the Nikon makes it to this size for my purposes (I wouldn't hesitate to print a sharp D200 shot this big), and the Horseman shot was cropped in one dimension only to produce a print size of 15.5x36 inches (an artistic choice, the image happens to look good as a panorama). I wouldn't crop a Horseman shot much tighter than this and try to print it 16 inches on the short side, judging by the fact that the detail is, again, "just holding" in this shot.
If I take both of these as being pretty close to the maximum print size I am comfortable with for their respective cameras (of course, I could print the Horseman 24x36 without cropping if I had a printer wide enough), that puts the short side of the Nikon's sensor as basically equivalent to about 4 cm of film. The Nikon seems by this measure to just about equal 6x4.5 (which makes some sense from what other people have said as well). The 1Ds mk II might have enough pixels to resolve like 6x6 (in a different shape), or even 6x7. If pixels scale more or less evenly (10.2 mp cropped out of a 1Ds mkII or a Phase looks something like a D200), this puts the single-shot backs in the category of 6x9 or a little larger, but not quite 4x5. I wonder how evenly high quality pixels scale? Of course compact digicam pixels scale VERY poorly (I've used 4 MP SLRs I prefer to any 8 MP compact), but are good pixels basically good pixels, at least at low iso? It also says that a big Phase back just might make a 30 inch wide print, because it has about 40 mp, so twice the resolution of my D200 (which makes 15 inch prints) in each direction...Nothing scientific, just my eye (and yours may differ).

-Dan

[Nitish Kanabar]

Nitish - in my view, you are correct. Upsampling (obviously) creates new pixels, because the total number of pixels increases. The soft/firm ware for doing so looks at the adjoining pixels and makes a best guess at what the color value for the added pixels should be. I believe what shileshjani meant by, "New information, no" was that the added pixels don't contain additional real detail.

Thanks Ralph. I'm afraid I don't quite follow. Perhaps there is something subtle (or obvious!) that I could be missing here. What distinguishes real detail with respect to an interpolated pixel? Pixels added to an image have a color-value - as do the original pixels. Once an image is up-sampled, then there isn't anything that distinguishes an interpolated pixel from an original pixel. To an observer, the interpolated pixel does represent some image detail in that case, doesn't it?

Considering my original example of the crude interpolation (#000000,#555555,#AAAAAA,#FFFFFF). To an observer, this would indicate that the subject did have those two gray values when the image was taken. Granted this is bad interpolation, but no interpolation algorithm is perfect and any pixels an algorithm does add would become indistinguishable from original pixels, wouldn't they? So, to me, up-sampling the image, does modify it at some level.

Having said that it does seem, to me, that J Michael Sullivan has a valid point when he points out that it is erroneous to draw conclusions from comparisions between upsampled images and non-sampled images.

[Dave Moeller]

Nitish-

I think I can help with the "New information, No" interpretation by extending your thought experiment a bit.

Assume that you have a digital camera that has a resolution of two pixels. Now, imagine that you filled the veiwfinder with two squares of mat board, one black and one white. You carefully align the image so that the white square will fall exactly over one of your pixels and the black square will fall exactly over the other pixel.

When you make the image and then look at it, you see the information that was present in the world when you pressed the shutter. There were two squares, one black and one white, and that's exactly what's in your image.

If you now upsample (using the pixel colors you stated in your previous post), you'll get more pixels in your image, but none of those pixels will represent anything that existed in the real world. So, yes, you do have more data. But no, you don't have more information. In fact, what you have is false information. Your picture shows gray shades that didn't exist in the real world. Upsampling algorithms can take a pretty good stab at what's missing, but in reality they never add real information to an image.

Hopefully this makes sense.

Be well.
Dave

[Nitish Kanabar]

Nitish-

I think I can help with the "New information, No" interpretation by extending your thought experiment a bit.

Assume that you have a digital camera that has a resolution of two pixels. Now, imagine that you filled the veiwfinder with two squares of mat board, one black and one white. You carefully align the image so that the white square will fall exactly over one of your pixels and the black square will fall exactly over the other pixel.

When you make the image and then look at it, you see the information that was present in the world when you pressed the shutter. There were two squares, one black and one white, and that's exactly what's in your image.

If you now upsample (using the pixel colors you stated in your previous post), you'll get more pixels in your image, but none of those pixels will represent anything that existed in the real world. So, yes, you do have more data. But no, you don't have more information. In fact, what you have is false information. Your picture shows gray shades that didn't exist in the real world. Upsampling algorithms can take a pretty good stab at what's missing, but in reality they never add real information to an image.

Hopefully this makes sense.

Be well.
Dave

Thanks Dave. The distinction makes sense now.

Interestingly, an observer viewing only the interpolated image without access to the original will not know that there is false information in that image or at the minimum will not know WHICH information is false in that image. Now when that observer compares the interpolated image with an non-interpolated image, then that observer is actually comparing a data-set with false information with a data-set with no-false information. The results of such a comparision would definitely be suspect.

[David Luttmann]

Let me go on the record that I never call MR et al "liars" -- those are David L's nasty words.

I called their methodlogy flawed, which it is.



J Michael Sullivan

Sorry Michael, you said the images were "doctored." That directly suggests that by making any comparison, the authors are trying to pull the wool over your eyes. Doctoring an image would be intentional manipulation.

But of course, you still try to avoid admiting this. Of course, you also avoid explaining to us all how we should compare resolution differences between images of different sizes. We've alrreasy establsihed the point that you were incorrect in stating the images were downsampled. They were only upsampled to keep image sizes on the screen relative. But please, do tell us how we compare different sized images.

I'm still waiting for your explanation as to how we do this, and how the intentionaly "doctored" (your words, not mine) the images. Maybe you can actually stick to this point to clarify youserlf.

Oh well.

[Jack Flesher]

They were only upsampled to keep image sizes on the screen relative.

FTR, the 4x5 film scan was done at 2500 PPI on a Tango and had to be downsampled slightly to match the BL file size.

[MJSfoto1956]

Interestingly, an observer viewing only the interpolated image without access to the original will not know that there is false information in that image or at the minimum will not know WHICH information is false in that image. Now when that observer compares the interpolated image with an non-interpolated image, then that observer is actually comparing a data-set with false information with a data-set with no-false information. The results of such a comparision would definitely be suspect.

That is precisely my point.

However, there *IS* a corrollary: if your final print size is 16"x20" using inkjet, then yes, the IDs and other high-end SLRs are more than competent to make stunning prints (stunning for inkjet of course). Using a 4x5 or larger camera will *NOT* give you an advantage at this size with inkjet output (whose "resolution" is a fraction that of analog paper). Only if you were to expose a 16"x20 negative and then contact print it would the difference be noticable (and then mostly only close up).

So in the digital domain, digital SLR's satisfy 80% of the market -- including advertisers. Digital is this regard is clearly "good enough". It remains to be seen if 16mp is as big as SLR's get (there are many technical reasons why it may not get signifcantly bigger than this with current CMOS technology).

As for larger formats (either film or digital) -- to print them digitally means to print them larger than 16x20 so that the the benefit can be readily seen. If we were to scale the Betterlight 200% to 12000x16000 pixels and print that at the native 360dpi of the Epson inkjet printers we would end up with a tack-sharp image at 33"x44". Printed at this size, the 1Ds would pale in comparison and the softening effect of the extreme scaling necessary would become obvious.

But digital artists do have a fair response: at a print size of 30"x40" (or larger) most people don't view such an image up close. And further, such an image viewed at 50" away, would have a hard time distinguishing between the Betterlight and the 1Ds. Of course, this is anethema to most large format photographers who demand an image that "takes you in" all the way to a few inches away. Few would deny that such an image is superior in every respect.

Unfortunately, the mass market seems not to care.
Thus, you persue large format for other reasons...

J Michael Sullivan