Mr D take a look at this thread

Discussion in 'Projectors, Screens & Video Processors' started by ROne, Jan 16, 2002.

  1. ROne

    ROne
    Active Member

    Joined:
    Dec 21, 2000
    Messages:
    781
    Products Owned:
    0
    Products Wanted:
    0
    Trophy Points:
    18
    Ratings:
    +8
  2. Mr.D

    Mr.D
    Well-known Member

    Joined:
    Jul 14, 2000
    Messages:
    11,193
    Products Owned:
    0
    Products Wanted:
    0
    Trophy Points:
    133
    Ratings:
    +1,236
    OK looks interesting will report back.
     
  3. Mr.D

    Mr.D
    Well-known Member

    Joined:
    Jul 14, 2000
    Messages:
    11,193
    Products Owned:
    0
    Products Wanted:
    0
    Trophy Points:
    133
    Ratings:
    +1,236
    OK did some testing myself ( I did it at work so can't post the images but I'll see if I can replicate the tests in photoshop or after effects or such like at home when I get the time)

    What I did : generated a 768x576 image of antialiased dots ( measure em tomorrow it was quick) and resized it up using a variety of filter types ( I was working 16bit but I'l shove em through at 8bit tomorrow). I sized it up using integer and float values ( the actual image size has to be rounded to a nearest whole pixel value at the end obviously)

    To cut a long story short the results were pretty much what I was expecting ( used to) nearest was OK with regard to artifacting at integer scaling factors ( 2,3 etc) (however it produced quite awful aliasing the higher the scaling went) it was terrible at float values ( 1.5 , 2.3 ) with lots of artifacting where the dots were mapping to subpixel. Effectively destroyed the linear dot structure of the image.

    Bilinear was good at integer values and softened the aliasing issues compared with the nearest filter but it produced a little artifacting at float values :nonlinearity issues in certain areas of the dots themselves nothing hideous though would expect this filter to squash and elongate certain features at certain scaling factors but nothing you'd notice on an isolated viewing most likely.

    Bicubic was again similar(identical) to the bilinear at integer values and also similarly it didn't exhibit aliasing type artifacts. However at float values it did not exhibit the non-linearity type artifacting of the bilinear and in this aspect was again indestinguishable from the integer version of itself or the integer version of the bilinear.

    I then tried a minify and here I'd forgeting some issues with resizing images : some filters are better for sizing down than up. So when I was sizing up with the minify it was terrible at all factors be they integer or float : either as bad as the nearest at float or actually worse at integer. Then I remembered that minify and sinc (which I didn't try) are really designed for downsizing images.

    Pulled in a 4K ( 4096x3112 10bit log film frame) and downsized it to fit in a 768x576 frame ( wasn't actually fullap image so it needed a bit of padding to fit). Bicubic looked to have retained plenty of image detail and even had a try at maintaining grain patina to a certain extent but it was a bit aliasy and pixelly here and there ( technical terms) tried the minify and it effectively removed the aliasing and pixellation but at the expense of real image detail : loss of detail in hair and grain detail lost for good or bad; image didn't actually appear unsharp and if anything it looked cleaner than the bicubic.

    Now as far as I'm concerned I'lll stick by my earlier analysis/prediction that the integer scaling benefits the lower filter types but with the possibility of aliasing as the filters themselves are too primitive to interpolate adequately to avoid aliasing issues anyway. Bicubic produced visually better results with little appreciable artifacting whilst maintaining an alias free image.

    If I get time I'll either grab a real video frame or put a film frame through a telecine type process to replicate video material and have a butcher's and see if the bitdepth and colourspace play a role as suggested by some of the people on the AVSforum : I read the forum I just don't plan on posting until I get a projector: money where mouth is and all that.
     
  4. ROne

    ROne
    Active Member

    Joined:
    Dec 21, 2000
    Messages:
    781
    Products Owned:
    0
    Products Wanted:
    0
    Trophy Points:
    18
    Ratings:
    +8
    Food for thought, I just thought it was interesting that there were people pulling in all different directions, the question that sticks in my mind though is do the Radeons/geforces of the world use bi-linear filters as suggested the forum debate. Is it a fair comparision to compare a graphics cards scaling ability to software scaling ability in such programs?

    Would be interesting to see any visuals you could come up with.
     

Share This Page

Loading...
  1. This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
    By continuing to use this site, you are consenting to our use of cookies.
    Dismiss Notice