Newbie-type questions

Discussion in 'Projectors, Screens & Video Processors' started by NicolasB, Feb 25, 2005.

  1. NicolasB

    NicolasB
    Well-known Member

    Joined:
    Oct 3, 2002
    Messages:
    6,726
    Products Owned:
    1
    Products Wanted:
    0
    Trophy Points:
    137
    Location:
    Emily's Shop
    Ratings:
    +1,090
    1) Why is frame rate conversion so difficult? Surely all you have to do is alpha-blend successive frames. Is alpha-blending really that computationally intensive?


    2) Why is deinterlacing so difficult if the source material is inherently progressive? (E.g. a cinema film).


    3) What is the point of trying to deinterlace an inherently interlaced signal (such as certain types of video)? Wouldn't it be better to leave it in its native state? I understand that (e.g.) plasma screens are inherently progressive displays, but does updating every other line 50 or 60 times a second really look that much worse than the result of attempting to deinterlace the signal?

    I can see that 720p has advantages, because you've actually got double the frame rate of an SD signal, and I can see that interlacing causes problems on a direct-view CRT display, because (for example) a horizontal line of white just one pixel wide would acquire 25Hz flicker. But if the pixels remain stably illuminated until told to change (as with plasma or LCD), why is interlacing a problem?
     
  2. KraGorn

    KraGorn
    Active Member

    Joined:
    Aug 30, 2003
    Messages:
    4,745
    Products Owned:
    0
    Products Wanted:
    0
    Trophy Points:
    68
    Location:
    Warrington
    Ratings:
    +27
    I'm surprised none of the experts have dropped in, though I know Joe and Gordon have been on the road.

    I know little of the theory, and nothing of frame-rate conversion, so can't offer anything on number 1.

    As far as deinterlacing is concerned, I'm not sure of the point you're making. In principle it isn't difficult as far as I can see, as long as during the encoding process the necessary flags are set to tell the decoder how to handle it then deinterlacing should be a doddle. The problem as I understand it is that many (most?) DVDs are not well coded and at times are wrongly coded, hence a 'good' deinterlacer has to be intelligent and not blindly follow the flagging.

    As for the need when displaying on solid-state devices, I don't know if these devices are capable of functioning in an interlaced mode, it may be that due to their basic design they can't do so.

    Moreover, unless displaying at 1:1 resolution scaling has to be done somewhere and even though I claim very little expertise I can't see how one could scale an interlaced video stream.
     
  3. Pottsy

    Pottsy
    Active Member

    Joined:
    Apr 15, 2003
    Messages:
    295
    Products Owned:
    0
    Products Wanted:
    0
    Trophy Points:
    18
    Location:
    Leicester
    Ratings:
    +8
    Yes.

    Because it ends up being interlaced. And all that hard work interlacing has to be undone again. Correctly.

    Ever watch interlaced on a progressive device? Try a DVBt PCI card (nebula or whatever) on your PC monitor. Looks terrible. Line flicker and motion artifacts as well as other nasties.

    If old fashioned CRT TV's didn't exist any more, then everything would be progressive and we'd have no problems. Bit like in the computer world.
     
  4. BertM

    BertM
    Standard Member

    Joined:
    Oct 3, 2004
    Messages:
    41
    Products Owned:
    0
    Products Wanted:
    0
    Trophy Points:
    6
    Ratings:
    +0
    Let's take an actor walking from one side on the screen to the other. To keep it simple, let's assume we have 50 progressive frames per second in this shot. To satisfy our plasma display, we should convert the frame rate to 60 progressive frames per second.
    In the input condition, we have 50 photos of the actor, a photo for each 20 milli seconds. When these photos are displayed in this 20 msec rate, we see smooth motion.
    In the output condition, we require 60 photos of the actor, each 16.667 msec apart to get the same smooth motion.
    In the original frame rate, there was no photo taken at 16.667 msec, so we have to estimate what happened at that time.
    Let's say the actor moves 20 cm per 20 msec, or 1 cm per msec. At 16.667 msec, he moved 16.667 cm. To make smooth motion, we'll have to generate a photo where the actor is 16.667 cm further than the previous photo. This takes very, very smart algorithms that use extensive interpolation and correlation calculations, using a series of 20 msec apart photos to calculate the photo at a 16.667 msec boundary/time stamp in between.
    Should we use alpha blending, in other words mix two of the input frames, we would see the actor at two places instead of at the correct position. Because the small displacement, this will show as a very blurry picture. Simple said, you would loose sharpness by simply using a blending algorithm, and you would never be able to generate time-accurate frames.

    Unfortunately, many, many programmes (sports for instance) are recorded with intelaced cameras.


    That's a good point. I think feeding interlaced signal to a progressive display might yield good results, but also might intriduce flickering on plasmas and progressive crt's as these are very fast in building a picture. On LCD's this might work out wonderful.

    Now this is excatly what advanced deinterlacers do: find out the places in the picture that move, and deinterlace these, and leave the unmoving parts as they are, until they change, simply to not loose sharpness.

    Bert
     
  5. NicolasB

    NicolasB
    Well-known Member

    Joined:
    Oct 3, 2002
    Messages:
    6,726
    Products Owned:
    1
    Products Wanted:
    0
    Trophy Points:
    137
    Location:
    Emily's Shop
    Ratings:
    +1,090
    Maybe I'm not explaining myself very well in places. :)

    I guess by "difficult" what I really mean is "expensive". The iScan HD+ is (I believe) considered quite good value, and yet it sells for close to £1000. Other scalers are far more expensive again. I'm having some difficulty understanding what it is that a scaler or deinterlacer does that requires it to cost quite so much.

    For example, deinterlacing an originally-progressive source (such as a cinema film) ought not to be very complicated. Take a blank bitmap, 720x576. Grab the first field. BitBlt each separate line of pixels onto your bitmap at lines 1, 3, 5, etc. Take the next field, and BitBlt each line onto your bitmap, this time at lines 2, 4, and 6. Bingo: deinterlaced image.

    The only remotely difficult thing is making sure that you are combining two fields from the same original frame rather than the second field of one frame and the first field of the next. But a) I would have expected that to be flagged at source, and b) even if it isn't flagged correctly, once you get it right the setting remains the same for the whole of the film. So all you'd have to do is present a test image to the user and have him manually toggle between the two possible settings (if necessary), and you've solved the problem. How can this require thousands of pounds worth of hardware to do?

    (It's only marginally more difficult for an NTSC signal using 3:2 pulldown - all you have to do is skip 2 fields out of every 6. Determining which is likely to be relatively easy because they're consistently exact duplicates of previous fields.)

    I can see that deinterlacing something whose original source was interlaced would be much more difficult, yes. But I can't see why it's actually worth bothering with deinterlacing a source like that at all, unless you're going to display it on some sort of CRT display. So if it is this alone which makes scalers expensive, then why not offer a cheaper version that doesn't do it?

    I don't buy that. :) In fact, alpha-blending of successive frames is exactly how NTSC-to-PAL conversion is actually done when (for example) a TV program made in NTSC is converted for broadcast in Britain.

    Find a series made in NTSC but shown on British TV (and where the NTSC video has been converted directly rather than a new PAL version being generated from the original film source). Step through it frame by frame, and that is what you'll see: 6 frames alpha-blended down to 5. Of course, this won't look as good as the original, but it looks far better than the weird "jerky" conversion one tends to see in practice on certain DVD players. And again, it doesn't seem like it ought to require all that much computation to do this on the fly. Perhaps it would on a general purpose CPU, but certainly not if you're using purpose-built video hardware.

    As for actual scaling, again, bilinear filtering (which, as far as I know, is the most sophisticated algorithm used - I'm not aware that commercial scalers use bicubic or Lanczos3 filtering or anything fancy) isn't particularly difficult - this is precisely why an HCPC can do a good job with only about £150 worth of video hardware.

    So what am I missing? :)
     
  6. Pottsy

    Pottsy
    Active Member

    Joined:
    Apr 15, 2003
    Messages:
    295
    Products Owned:
    0
    Products Wanted:
    0
    Trophy Points:
    18
    Location:
    Leicester
    Ratings:
    +8
    This is particularly depressing when you consider that the guts of the machine (the SiL504 chip) costs only £7.
     
  7. BertM

    BertM
    Standard Member

    Joined:
    Oct 3, 2004
    Messages:
    41
    Products Owned:
    0
    Products Wanted:
    0
    Trophy Points:
    6
    Ratings:
    +0
    Hmmm, this sounds too easy. All correct timing will be lost, leading to jerky playback. Original NTSC material looks good on my PAL system, so the conversion must have been a lot more advanced than the frame rate conversion my iScan uses. But maybe you're right, and is it a simple trick of blending a certain percentage of consecutive frames, but I'm still wondering how this would correct for time domain errors. Blending two frames at 0 msec and 20 msec can impossibly produce the required frame at 16.667 msec.

    BTW, checking correct fields in originally progressive material can be very easy and fast performed by simple cross correlation routines.

    regards,
    Bert
     

Share This Page

Loading...