D100 and Rebel WTF??? 6MP vs 3MP Opinions please....

Discussion in 'Digital Cameras' started by The Dude, Jan 6, 2004.

  1. The Dude

    The Dude Guest

    I've shot over 10,000 shots on my D100 since I got it in Aug '02. I just
    bought my wife a Canon 300D (she's always loved her film rebel). I have
    always shot at large/fine (that's 6MP, low compression jpgs). Having looked
    at the review of the Foveon based SD9 at dpreview, I wanted to test
    something. If you look at p.20 of the review, you can see that Phil shows
    us that the ressed-up 3MP images of the SD9 carry at least as much detail as
    the native 6 MP inages of the D60. Now, I know that Foveon has other issues
    (noise, moire, contrast, raw only etc) and I'm not trying to start a Foveon
    vs Bayer flame war here. Rather, it made me curious about shooting with the
    D100 and 300D in 3MP mode. Would those 3MP images, ressed-up, carry the
    detail of the native 6MP I've been shooting? Well, damn it if the answer
    isn't YES!

    I started with some of my sharpest 6MP pictures. Well exposed, from a
    tripod, shot through 28-70mm 2.8 AF-S... an awesome lens. I did a Photoshop
    resize of the 3000x2000 image to 2240x1498. I then saved it as small.jpg.
    I reopened it, ressed it up, and compared it, per pixel, to the original 6MP
    image. I was startled by what I saw. No discernable change in image
    quality?!?! I proceeded by shooting at medium/fine and large/fine. Again,
    no meaningful difference once the medium file was ressed up. I repeated the
    exact same observations with the Canon 300D. Please repeat these tests
    yourselves and let mw know your opinions. This revelation will save lots of
    HD space and permit far more images on CF.
    The Dude, Jan 6, 2004
    1. Advertisements

  2. The Dude

    Michael Guest

    If you downsize an image by 25%, save it in a lossy format like jpg, then
    load it and size it back up, you will have a signifficant loss of
    information compared to the original image. My guess is something like 50%
    loss, depending on the jpeg quality setting (in your example 44% of all
    pixels are gone, plus another 5-10% in the jpeg compression).

    Photoshop does not resize simply by eliminating 1 out of every 4 pixels. It
    applies a special algorithm to reduce aliasing and preserve the original
    "look" of the image, even when it has less pixels to do it with. It does the
    same thing when it sizes back up from 3 to 4 pixels. It is not unlikely that
    it can do this so well that your eye is fooled into thinking there is no
    difference. But there is - at least 44% of the image is now synthetically
    generated by a computer program, whereas the original image was captured
    through the lens.

    Try to create a new image in Photoshop - 5x3 pixels in size. Draw a tiny
    pyramid, 5 pix at the bottom, 3 in the middle and 1 on top. Then ask
    Photoshop to scale it up to 640 x 400. You won't simply get 9 huge blocks
    stacked on top of eachother, you'll get something that "looks" more like the
    small image. Photoshop is adding what it thinks is missing to draw a big
    pyramid, based on the information it can extract from the small image. It
    doesn't look very nice but you get the idea.
    The fact that scaling a 3 mpix image up to 6 mpix in Paintshop or Photoshop
    makes the image look almost as good as a 6 mpix image is hardly a
    revelation. It's more like a testament of the effectiveness of the
    algorithms these programs apply to help people like me who are short on real
    megapixels .. :)

    Michael, Jan 6, 2004
    1. Advertisements

  3. I will always shoot in the finest jpg mode or raw. One might want to crop
    later. Best to start with as much detail as possible. Also, printing your
    shots 8x10 or larger might change your mind.
    Charles Schuler, Jan 6, 2004
  4. The Dude

    The Dude Guest

    How silly. Of course I understand that most simple concept. Let me be more
    clear in my assertion. The logic of this is as follows... the Bayer
    interpolation yields a precise 3MP image for a 6MP sensor. It seems that
    little is gained in actual resolution in the jump from 3 to 6. Those
    missing "actual" pixels that adobe guesses, are as good as the Beyer
    interpolation. Bayer requires it's own guess work. It looks like Bayer
    interpolation yields all it's detail in half of the actual CCD or CMOS pixel
    count. I have 2 cameras I have proved this with. Please give it a try
    before writing it off. Take a picture of that newspaper at 3MP and 6MP.
    You will see a nearly perfect match.
    The Dude, Jan 6, 2004
  5. The Dude

    The Dude Guest

    I frequently print at 13x19'. Believe me, I understand. That is why, up
    til' now, I always shot at large/fine on my D100. However, if all of the
    detail of the 6MP image can truly be resolved in the 3MP, it makes no sense
    to use the large setting. My theory is that a 6mp Bayer mosaic pattern
    yields 3MP of true detail. Likewise, a 4MP sensor would yield 2MP of true
    detail. And so on. Again, see here...


    The ressed-up 3MP images on the left provide at least as much detail
    information as the native 6MP images on the right. My tests with the D100
    at medium vs. large show no significant difference in detail. Try this to
    prove it to yourself. Take the exact same image at large and medium.
    Increase the size of the medium to match the large and decide how much the
    "guesswork" of Adobe is inferior to the guesswork of the native 6MP Bayer
    image. Very little change. For fun, size'em both up to 12MP. Is there any
    difference in the images. Try it!!
    The Dude, Jan 6, 2004
  6. The Dude

    Jon Noble Guest

    I too have always assumed the more pixels the better. And it is true.
    However, I have a story similar to yours. Relatives of mine recently came
    back from Africa with a full CD of pictures, shot with several cameras. One
    image caught my fancy even though it was 1.3 mp. I used Fred Miranda's SI
    plugin for Photoshop to resize this picture so it could be printed 13x19 at
    240 ppi using my Canon i9100. The results were stunning, far better than I
    ever expected.

    Everyone says you can't print pictures from a 1.3 megapixel camera at 13x19,
    but upping the rez with intelligent algorithms makes it worthwhile.

    OK, not exactly your situation, but another example of how good up rezzing
    can be.

    Jon Noble, Jan 6, 2004
  7. The Dude

    Chris Brown Guest

    Doubt it - you may want to consider using RAW mode, and/or applying some
    unsharp mask to your images.

    I just did this test with my 10D. Here's the result:


    The picture was converted from raw with no sharpening applied, duplicated.
    One of the copies was resized to 3 megapixels (2160*1440), and then back to
    6 megapixels (3072*2048) using bicubic interpolation. Each image then had
    the same level of (relatively aggressive) unsharp mask applied. Both samples
    are fragments viewed at actual-pixel size. The shot was handheld at 100 ISO,
    using a Canon EF 50mm f/1.4 lens and a 420EX flash. I find the difference in
    image quality quite discernable indeed!

    I think it should be pretty obvious which is which...
    Chris Brown, Jan 7, 2004
  8. The Dude

    JPS Guest

    In message <3ffb0a8d$0$7065$>,
    You're missing the forest for the trees; an image is not a fixed pixel
    spacing; it is a collection of pixels, which vary in number. You can
    not assess the detail of an image by looking at it pixel-for-pixel or at
    any PPI resolution, unless you are comparing two images of the same

    A camera that is properly filtering the light striking the sensor can
    *NEVER* have a black pixel next to a white pixel. It is impossible, and
    any system that can sample that way at its native resolution is
    cheating, and allowing garbage to get recorded. It *is* possible if you
    leave out the filtering, or if a quick'n'dirty downsizing algorithm is
    used on the data. "Nearest Neighbor" is the quickest and dirtiest, and
    6.3MP bayer images downsized to 3.43MP with this algorithm look
    strikingly similar to SD9 images, but 1Ds images with the same target
    size look even closer. I suspect that what the SD9 is giving you is
    what you would get if you used a properly-filtered 23MP image, and used
    nearest neighbor to downsize it to 3.43MP.

    Unless you were using the very sharpest of lenses (like expensive
    telephoto primes), a 23MP APS-size image would probably look pretty
    bland "pixel for pixel", compared to the garbage you would get with the
    3.43MP Nearest Neighbor downsize, but actually contains up to 6.7x times
    as much detail (depending, of course, on the optics).

    Printing both at the same size, the one that looked better "pixel for
    pixel" will look like utter crap next to the "pixel for pixel blurry".

    And, of course, if the lens is sharp enough, the properly-filtered image
    still contains detail at legal frequencies it has attenuated that can be
    boosted with USM at sub-pixel radii; the sharper the lens, and the less
    noise you have, the better you can restore them.
    JPS, Jan 7, 2004
  9. The Dude

    JPS Guest

    In message <3ffb178e$0$95111$>,
    It depends which algorithm you choose, but none of them do what you
    suggest. Photoshop's "image size" does not know or care that your
    pyramid has 3 edges. It sees 15 pixels, and nothing more.

    If you want a real pyramid, use the vector tools.
    JPS, Jan 7, 2004
  10. The Dude

    JPS Guest

    In message <3ffb2612$0$87180$>,
    That depends on the quality of focus.
    There is no "bayer interpolation" of pixels. The only interpolation is
    the full RGB set for the pixel.
    Bayer does not "guess" anything. It "assumes" that chromatic values are
    not going to change at more than half the pixel rate. The luminance
    values are extremely accurate, if done correctly, and they are what
    count the most. We don't see RGB color with our eyes and brains; we see
    luminance with a ghost of color hovering around it, mainly fabricated by
    the brain.
    You are confusing a lot of issues. "Bayer" has nothing to do with any
    of this. It is just a coincidence that the only Foveon cameras
    available have no anti-aliasing filter (and the SD9 doesn't even have
    microlenses). All sensors should have some degree of anti-aliasing
    filter; at least with current sensor technology that is unable to
    out-sample the sharpest lenses used. Any number of reasons could make a
    3MP downsize seem to contain everything the 6MP contains. Motion blur
    of the camera often occurs in circles; preventing any real focus over
    the length of the exposure; zoom lenses are often soft at their max zoom
    and/or wide open. The AA filter suppresses detail that can be
    resurrected (if present) with a USM of about 0.3 to 0.5 pixel radius at
    200 to 500%, and this detail is hard to see without the USM.

    If you use a sharp lens, with a flash, focused precisely on the subject,
    you *will* lose detail downsizing to 70.7%. It's really a matter of
    degree, and it has as much to do with the lens and camera handling as it
    does with anything else.
    JPS, Jan 7, 2004
  11. The Dude

    JPS Guest

    In message <L2HKb.7424$>,
    Well, you know, halving the pixels only reduces the height and width by
    29.3%, which really is not as much as "50%" sounds like. It's not
    really that big of a difference.

    The higher we go in megapixels, the less any image is going to suffer
    tremendously from a 70.7% resize. The benefit of increased pixel count
    is not linear; it's not even linear to a single dimension (square root
    of pixel count).
    JPS, Jan 7, 2004
  12. When you reduce the size from 6 to 3 MP, the maximum possible linear
    resolution is reduced to 70% of what it was. Now, the full-size image
    may not have actual detail up to that limit, because the anti-aliasing
    filter necessarily removes some fine detail below the theoretical limit.
    When you downsize to 3 MP, you're using a digital filter that can be a
    bit better about keeping information than the physical AA filter. Thus,
    the 6 MP image may not lose as much as you'd expect, because it's
    somewhat weak on that high-frequency detail in the first place.

    However, there is *some* loss, if the original is sharp. For example:
    go to http://www.dpreview.com/reviews/canoneos300d/page21.asp and
    download the upper left full-size test chart. Look at the resolution
    test section just to the right of centre, a sort of sloping wedge.
    There's another wedge above that too, for testing horizontal resolution.
    Look and see how far along the wedges you can go and still see 9 clearly
    resolved black bars. You'll have to view the image at 100% scale at
    least to make any judgements. I'd say the limit on this particular test
    is somewhere between 14 and 15, which means 1400-1500 lines per picture
    height. (Note that this isn't the resolution limit, just a point up to
    which I know I can get clean images without artifacts).

    Now take this image and reduce it from 3072 to about 2172 pixels wide.
    Look at the resolution wedges again. Now the highest resolution clean
    pattern is at about 10 or 11. Resizing the image back up to 3072 wide
    does not improve that figure at all. In other words, detail that was in
    the original image has been permanently lost, at least in this example.

    Whether there is any such fine detail in your images depends on
    technique (use of tripod, shutter speed, focusing, depth of field).
    Whether you can *see* any loss of detail even if some happens depends on
    the subject of the photo. Maybe for what you shoot and the way you
    shoot, there won't be any noticeable loss. But it isn't *always*
    visually lossless.

    The other question which you haven't tested: is there any difference in
    quality between just setting the camera to 3 MP output, and shooting at
    6 MP but reducing later to 3 MP using Photoshop or Irfanview? The
    former is attractive because you can store more images on a given size
    card, but the latter gives you more control of the resizing process.

    Dave Martindale, Jan 7, 2004
  13. But not all of the "detail information" in the 3 MP images is actually
    real. There are a number of places where what you see in the images is
    very unlikely to be what was actually present in the original scene.
    This is a known problem with any image capture device which lacks
    adequate anti-alias filtering. The Sigma cameras record *something*
    that looks sharp when presented with too-fine detail, even if it's
    sometimes wrong. The Bayer cameras show you a blur when they can't
    resolve fine detail, which is at least theoretically the more correct

    It's like the difference between starting with a 12 MP image and
    resizing down to 3 MP using "nearest neighbor" or "bicubic"
    resampling. The first looks sharp, but has problems. The second
    looks less sharp but more faithfully records the original scene.

    If the Sigma camera had used an anti-aliasing filter, like the
    Canon/Bayer camera did, the true difference in resolution would be more
    apparent. In addition, the Sigma raw converter software applies more
    sharpening in software than the Canon in-camera software, so the Sigma
    images look sharper when the fine detail content is actually the same.

    Anyway, don't conclude that 3MP = 6MP based on *this* comparison.

    Dave Martindale, Jan 7, 2004
  14. The Dude

    The Dude Guest

    You would think. But look at the 7th picture down on this page...

    The vertical details are actually better resolved on the ressed-up 3MP

    But doesn't this ignore the guesswork of Bayer interpolation?
    The Dude, Jan 7, 2004
  15. The Dude

    The Dude Guest

    True enough. There is measurable loss.
    The Dude, Jan 7, 2004
  16. The Dude

    Chris Brown Guest

    I don't think it is, actually. What this is showing is just the different
    levels of sharpening applied by both sets of software, not very much (D60)
    vs quite a lot (SD9). Sharpness is not the same as detail. The detail is
    there in the D60 picture, it's just less obvious because it hasn't been
    subjected to the same level of aggressive sharpening that the SD9 image has.

    You don't have to take my word for this. Download the D60 image yourself,
    load it into Photoshop and have a play with unsharp mask.

    Phil Askey's reviews are pretty good, but he does have this bad habit of
    comparing images from different cameras which apply radically different
    amounts of sharpening, and then using these comparisons to derive things
    about resolving power. This is misleading.
    Chris Brown, Jan 7, 2004
  17. You mean the "picket fence" on top of the riverboat? If you look
    carefully at the left one, you'll see that not all of the "points" on
    the fence are equally well resolved. Some even seem half-missing. This
    is a clue that you're looking at detail that's too fine for the SD9 to
    resolve *accurately*, and that any 3 MP camera with a proper
    anti-aliasing filter would not resolve it at all.

    So the difference is mostly one of leaving out the AA filter, which
    isn't generally a good idea.

    Dave Martindale, Jan 7, 2004
  18. The Dude

    The Dude Guest

    That's a good point, but I can't fully agree. Try this...

    I downloaded both tower bridge images. I ressed-up the SD9 image to 6MP. I
    then zoomed into the top of the left tower on each image (@300%). No matter
    how much I played with USM, I could not get the D60 image to resolve the
    roof tiles that the SD9 shows pretty clearly. This is surely superior
    detail. I am aware that the reschart comparison discussed earlier in this
    thread does produce a measurable change when downsampled and ressed-up. But
    I still feel that the ressed-up 3MP med/fine jpgs are indistinguishable from
    the native 6MP images produced by the camera in every real world image I've
    tried. Have you tried it yet? You might be surprised.
    The Dude, Jan 7, 2004
  19. The Dude

    Chris Brown Guest

    Indeed, that's quite clearly aliasing. You can see a similar effect in
    Michael Reichmann's original comparison of the EOS 1Ds, 35mm and MF film.
    The 1Ds gave similar aliasing on the corrugated side of a building, which
    led Mr Reichmann to originally erroneously conclude (subsequentally
    corrected, IIRC) that the 1Ds outresolved medium format slide film, which of
    course it didn't.

    If you sharpen the D60 image so that it has a similar level of sharpening to
    the SD9 image, they end up resolving a pretty similar amount of detail, with
    perhaps the D60 having a very slight edge.
    Chris Brown, Jan 7, 2004
  20. The Dude

    The Dude Guest

    I suppose it's a subjective thing. But as I look at the "picket fence"
    image on the left vs. right, I prefer the less-than-perfect vertical info
    from the 3MP image to the utter blur of the 6MP image. Just look at the two
    crops. Can you really say you prefer the one on the right? Honestly?
    The Dude, Jan 7, 2004
    1. Advertisements

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.