[mpeg2] Re: Kfir Mpeg2 to SVCD

[Dan Hollis <goemon@anime.net>]

On Wed, 31 Jan 2001, Andrew Stevens wrote:
> Neither matters! The whole point is to test transcoding and NTSC/PAL
> isn't an issue.  Mind you if you're transcoding it probably makes more
> sense to grab at fairly high resolution 720x<y> and then transcode
> down.  I presume this is 480 across the whole image not just sampling
> 480 from the middle?

It is 480 across whole image.

> Aside: how good is the MPEG the KFIR generates?  In principle, with
> fast enough hardware, you could do a stupendously good job of motion
> compensation etc.  They claim scene change and flash/fade detection
> which is pretty cool.  What about motion compensation: is it purely
> luminance based or is chrominance taken into acount as well?   In
> theory, with fast enough hardware, they might even be able to do
> super snazzy things like choosing motion vectors based on the MDCT
> coefficient sizes for possible matches rather than using variance or
> sum-absolute-difference as approximate indicators.

>From my simple short tests of kfir (wintv-pvr) under Windoze:

With noisy sources kfir does really badly (it has not really got any good
noise reduction filters -- just a simple aperture and bandpass filter). I
tried to capture 480x480 2600kbps from VHS tape and quality was not so
good. When source is noisy with low bitrate the kfir seems to have hard
time with scene change.

flash/fade detection seem to work nicely, better than tsunami mpeg at that
bitrate (2600kbps).

visiontech claims 100 horizontal 100 vertical pixel for motion
compensation search. I don't know what the limit of tsunami mpeg is. It
certainly appears to do a good job of motion compensation (better than I
hoped, for such cheap hardware).

For realtime vcd rate (352x240, 1152kbps) the result was horrible. I don't
know if it was because of noisy source or because of some kfir limit, but I
didn't try it much - i couldn't tolerate the horrible result for too long
before my eyes exploded in pain.

If you give kfir enough kbps, it does a marvelous job - 704x480 and
480x480 were both very good at 5mbps. I'm sure with some tweaking you
could get near broadcast quality at 480x480 with some lower bitrate.

It would be nice if you could feed data from CPU directly into the kfir
for transcoding, that way I could capture video, do some
super-noise-correction, cropping and filtering, so that kfir can do better
encoding.

kfir claims to have auto 3:2 pulldown detection, I wonder if this means it
can IVTC in hardware...? It would be cool if it could.

-Dan