[MPlayer-users] pullup behaves badly on this clip

[D Richard Felker III]

On Thu, Sep 02, 2004 at 12:38:48PM -0400, James Crowson wrote:
> http://jc5.ath.cx/shared/utena.mpg
> 
> This is a 7.5 second clip from the first DVD of the anime series
> Revolutionary Girl Utena.  pullup seems to behave very oddly on this
> source material.  This clip is probably the worst example.  It drops a lot
> of frames, making the panning extremely jerky.  (The other inverse
> telecine filters seem to behave more sanely wrt frame drops, but they're
> bad enough on *other* parts of the source material that I just want pullup
> to work...)

after more investigation i can tell you a little about this clip, but
unfortunately the news i have is bad. during the pan, the pattern of
fields is as follows:

A | Ab | Bc | C  | D  | E | eF | fG | G  | H  | I
A |  B |  C | Cd | De | E |  F |  G | gH | hI | I

where the letter indicates the original progressive frame and the
line indicates the field (top/bottom). whenever two letters appear
together, it means fields from those two frames are blended, with the
capital letter being the stronger one in the blend (roughly 3:1
strength ratio, or perhaps 2:1).

according to the info i'd read before, i was under the impression that
utena was encoded with a pattern like this:

A | AB | B | C  | D | DE | E | F  | G
A |  B | C | CD | D |  E | F | FG | G

while disgusting, and while it confuses most naive inverse telecine
engines, this latter pattern is ok for pullup because it's able to
match the fields that fit together and throw away the blended ones.
notice that for each frame, there's at least one nonblended top field
and at least one nonblended bottom field somewhere in the stream.

but for the pattern in the clip you provided, frame B does not have a
nonblended top field anywhere, and frame D does not have a nonblended
bottom field! this makes it very difficult to repair.

in theory, a specially-written "utena inverse telecine" engine could
operate by suptracting out the blended crap (notice that when you have
a bad blended field, the "sister field" it's blended with always
exists in a clean form) or by performing some sort of averaging
process on the two blended fields (e.g. using Ab and Bc to estimate
B). however, this doesn't even account for the fact that it's very
difficult for software to find the patten with horrible material like
this, which makes the prospects of such an engine being written very
slim... :(

if you have some money to invest in the problem, i'd try importing one
of the region 2 (japanese) utena discs and see if they're better. the
r1 release is known to be horrible. alternatively you could try
getting pal discs (maybe from australia) and see if the studio did a
proper job making them 25fps progressive, but i somehow doubt it... if
you try any of these alternatives, please let me know the results,
since i'm looking for a clean copy of utena myself.

rich