White paper: Foveon 2.4 - 5.0 times higher resolution than same MP Bayer

Discussion in 'Digital Photography' started by George Preddy, May 15, 2004.

  1. George Preddy, May 15, 2004
    #1
    1. Advertisements

  2. George Preddy

    SteveJ Guest

    I like the idea of the operation of the chip, but test show that the color
    fades at the higher ISO setting above 200.
    Give it time to mature.
    I need full color, LOW Noise, at the 1600 ISO setting.
     
    SteveJ, May 15, 2004
    #2
    1. Advertisements

  3. It took you a long time to find that document...

    Its conclusions are biased, and misleading. They compare pixel resolution
    while avoiding to mention the number of discretely sampled output pixels.
    According to their reasoning, a camera with one pixel is better than one
    with 6MP, obviously ludicrous.
    That 'comparison' was specifically construed to "exploit" certain feature
    differences and "fool" the Canon sensor. The carefully chosen worst-case
    color combination, is also extremely unlikely to be encountered in a normal
    scene.

    Again you are carefully avoiding to mention the difference in using a
    low-pass filter on the Canon and the lack of a low-pass filter on the Sigma,
    leading to aliasing artifacts with the Sigma.
    Also avoiding the fact that the low fill factor of the sensors in the Sigma
    leads to a further need for such a low-pass filter to avoid aliasing and
    stair-stepping.
    Also avoiding to point out that the Field of View between cameras was
    different, <http://www.xs4all.nl/~bvdwolf/main/downloads/AoR.png> because a
    fair test would have given something closer to this result:
    http://www.xs4all.nl/~bvdwolf/main/downloads/SD-9_identical_FoV_to_1Ds.jpg

    And for a more meaningful resolution comparison, compare:
    http://www.dpreview.com/reviews/orig.asp?/reviews/samples/rescharts/sigma_sd9.jpg
    which
    starts to misrepresent the finer lines above the 12 marker (even counts only
    4 or 5 black
    lines at the 20 marker instead of 9), with:
    http://www.dpreview.com/reviews/samples/rescharts/canon_eos1ds.jpg which has
    little trouble
    resolving the finest detail at the 20 marker (we need more detailed test
    cards for this
    camera), and the Fuji S2 at 12MP output setting on the same page
    http://www.dpreview.com/reviews/canoneos1ds/page21.asp also reaches a
    reasonable 18+ mark.

    Bart
     
    Bart van der Wolf, May 15, 2004
    #3
  4. George Preddy

    Bryce Guest

    Typical preddy bullcrap.
     
    Bryce, May 15, 2004
    #4
  5. Interesting tests. At least they matched the field of view of the
    cameras, which the outbackphoto test failed to do.

    The main thing the resolution tests show is the lack of an anti-aliasing
    filter. They show considerable contrast above the Nyquist frequency for
    the X3, which is BAD because it causes aliasing. The 10D's contrast
    drops much lower at the Nyquist frequency because of the presence of the
    AA filter, which is necessary. This also reduces contrast below
    Nyquist, but that can be restored by a bit of sharpening.

    You can see the unpleasant side effects of no AA filtering in the
    diagonal line test shown in Figure 7. The Canon slanted line looks like
    it is the same width everywhere with smooth edges. The Sigma SD10 image
    of the same line looks like the edge of the line is wavy and the
    width of the line varies. (And the SD9 would have been worse yet!).

    Is there anyone other than George who would prefer the wavy line on the
    right to the clean line on the left, particularly when a bit of
    sharpening will eliminate most of the sharpness difference while nothing
    will fix the waviness of the line on the right?

    Also, they are quoting resolution in cycles/pixel terms, not cycles per
    image height. What the data says is that the X3 sensor has higher
    contrast and thus higher apparent sharpness with the same number of
    pixels. Unfortunately for Sigma/Foveon, no other current DSLR does have
    as few pixels as the SD9/SD10, and so the comparison is not accurate.

    To compare the 10D to the SD10 on a lines per picture height basis, the
    data for the 10D would have to be stretched horizontally by a factor of
    1.4 relative to the SD10, giving about the same contrast at lower
    frequencies and considerably higher useful resolution. To compare the
    11 MP 1Ds, the Canon graph would have to be stretched horizontally by a
    factor of nearly 2. It would then beat the SD10 at all frequencies even
    without any sharpening.

    In other words, the paper pretty clearly demonstrates that a Foveon
    sensor without an anti-aliasing filter is badly flawed for photography.
    Those have been discredited many times in this newsgroup.
    Finally, note that George seems to be misinterpreting the test results.
    When the paper says "the frequency response of the Foveon is 2.4X that
    of the Bayer", what they mean is that the amplitude (contrast) is
    higher at a particular frequency. They do *not* mean that the upper
    frequency cutoff (the more usual meaning of frequency response) is any
    higher for the Sigma. That's set by the Nyquist limit, and the Sigma is
    worse than all other DSLRs in this respect because it just has fewer
    sensing sites than anything else.

    Nothing in the paper quoted supports George's math above.

    Dave
     
    Dave Martindale, May 15, 2004
    #5
  6. Only if color sampling is taking place, which with Foveon it never
    does, it is a full color sensor. That is why there is never color
    aliasing in Foveon shots, in fact color moire is absolutely
    impossible.
    By two grade schoolers. Those tests simply show raw data. The fact
    that you are offended by raw data is proof that you are acting
    irrationally.
    "The results show that the SFR for Foveon X3 sensors is up to 2.4x
    better. In
    addition to the standard SFR method, we also applied the SFR method
    using a red/blue edge. In this case, the X3 SFR was 3–5x higher than
    that for Bayer filter pattern devices."

    Obviously B&W targets are meaningless, color targets count. The 2.4X
    figure (derrived from B&W targets, when no color interpolation is
    required) is always cited, and it is ridiculously conservative in
    favor of Bayer.

    Bayer is great for black, but black only because any sensor, R, G, or
    B can return a black reading. IOWs, as long as there is no light, you
    get the Bayer-advertised MP rating. If there is light, you're out of
    luck to the tune of a 75% drop in optical resolution. Though this
    papers claims 80%, probably due to the additional effect of the
    Foveon-unecessary blur filter.
     
    George Preddy, May 16, 2004
    #6
  7. George Preddy

    JPS Guest

    In message <>,
    The Sigma resolution images are aliased. The 10D ones look better, IMO.

    They didn't mention that frequencies near-but-below the nyquist on the
    10D are attenuated, but respond to sharpening very well.
    --
     
    JPS, May 16, 2004
    #7
  8. George Preddy

    Ray Fischer Guest

    All cameras sample color.
    Of course there is.
    Yeah, that's about all it would take to discredit your bullshit.
     
    Ray Fischer, May 16, 2004
    #8
  9. The problem with the Sigma camera is that there is lots of luminance
    aliasing. Look at the diagonal line in the paper you referenced: it
    looks like it has a scalloped edge and isn't uniform in width.

    The reason is that anti-alias filtering is needed in all systems that
    regularly sample a continuous function at discrete points. This is
    perhaps the most fundamental basis of digital signal and image
    processing. It applies to digital audio, to black&white cameras, and to
    colour cameras. This includes the ones with the Foveon sensor; to say
    otherwise is simply preposterous.

    The contrast graphs clearly mark where the Nyquist frequency is, and the
    Foveon-sensor camera clearly has quite high contrast at the Nyquist
    frequency. Any student who has taken one signal processing class can
    tell you that the result will be aliasing.
    What the tests show is that if you make test images using
    complementary-colour test targets designed to show Bayer sensors at
    their weakest, and shoot under conditions of unequal fields of view that
    discard all the extra pixels of the other DSLRs, the Sigma cameras can
    sort of equal other DSLRs. Those results are valid on their own, once
    you understand the test questions.

    However, you tend to quote that page as if it showed superiority for the
    Sigmas (it doesn't), and as if it was related to normal photographic
    subjects (it's not). Further, you normally just post pointers to the
    two images, without the explanation. It's *your* use of that page to
    support your argument that has been discredited.
    And what does that value of "2.4X" mean? If one reads the paper
    (you obviously did not read it, or did not understand it), that value is
    a contrast difference at a single spatial frequency. What it says is
    that if you shoot a picket fence at a particular distance with the SD10,
    and then recompose the shot so the same area fills the same number of
    pixels on the 10D (i.e. discarding half of its sensor area), then the
    raw data from the Foveon sensor will show more contrast than the raw
    data from the 10D. It absolutely does not indicate that the SD10 has
    any higher resolution - this is a measurment *at the same resolution per
    pixel on both cameras*, and it pretends that the higher pixel count of
    the 10D does not matter.

    But, in fact, the 10D does have more pixels, and does have a higher
    resolution as well, as clearly demonstrated by dpreview's resolution
    tests. And the high contrast of the SD10 at the Nyquist frequency,
    clearly visible on the graphs, mean that it's not suitable for general
    photography at all.

    There is absolutely no way that, if a camera has 2X as much contrast at
    a particular frequency, that you can say it is equivalent to one with 2X
    as many pixels. As the robot on a particular TV program used to say,
    "That does not compute".
    But that number tells you nothing at all about resolution - it's a
    contrast measurement. Read the paper! And B&W targets *do* count,
    because the human eye has 10 times higher resolution for seeing
    luminance changes than colour changes. That's why most camera testing
    uses B&W charts.
    That's not an argument; it's just babbling.

    Dave
     
    Dave Martindale, May 16, 2004
    #9
  10. Nonsense. Aliasing can even take place in a 'monochrome' sensor. Color is
    totally irrelevant.

    SNIP
    No, not in resolution, which is your claim.
    Which proves that you don't understand what they have written.

    A higher SFR means a higher modulation, and I may add at a frequency that
    seems to be carefully picked to maximize the difference with properly
    anti-aliased images.
    Yet many of the SFR tests described by the paper were done with a black and
    white slanted edge target. So you mean to say that those tests they
    performed were wrong?????

    SNIP of more misunderstood results.
    No they don't. They always carry a signal, because true black only exists in
    a theoretical Blackbody, and there is always noise. Black only exists after
    subtracting a certain Blackpoint level.
    SNIP of more misinformation and Preddy weird science.

    Even a reflective Black and White target on a Photographic silver image with
    the highest dynamic range will still reflect 0.5% of the incident light.
    That 0.5% is only 7.7 bits darker than Saturated White, and paper White is
    only 95% reflection, so the difference is even easier to catch than that.
    The conclusions you draw, are therefore completely baseless.

    What is more, the SFR tests do not allow to saturate the sensor, nor do they
    allow a blacklevel that leads to zero luminance, so a lower contrast target
    should be used.

    Bart
     
    Bart van der Wolf, May 16, 2004
    #10
    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.