[MEncoder-users] Re: framerate control question

[rob pfile]

Trent wrote:

> I found that decimate=2 will not work correctly.  It helps to raise  
> the hi
> threshold setting, to something like say 1000, but it will still make
> mistakes.
>
> The problem is that the duplicate frames that were created for the  
> 24 -> 60
> conversion are not 100% identical.  There are encoding artifacts  
> that make
> them different.  In high motion scenes where there are lots of  
> artifacts,
> the duplicate frames will be rather different from one another.  On  
> the
> other hand, in low motion scenes the *non*-duplicate frames will be  
> very
> similar.

> If you choose a low threshold for decimate, it will not drop frames  
> in high
> motion scenes that it should have dropped, resulting in very  
> noticeable
> stuttering.
>
> In you choose a high threshold, it will drop frames that it  
> shouldn't have
> dropped in low motion scenes, resulting in a less noticeable  
> stuttering.
>
> I extracted frame difference information for some clips and  
> analyzed it in
> a statistics package.  There are no settings for decimate that will  
> not
> make a signficant number of mistakes.
>
> That's what prompted me to come up with my own filter.  Just  
> following a 3/2
> pattern doesn't work when you have dropped frames from transmission  
> errors or
> commercials.

yes, i agree with everything you have said. furthermore at least ABC  
seems to render the titles and the TV ratings bugs at 60fps, which  
messes up decimate.

there is yet another person lurking around besides you and me who has  
written their own filter for 720p 60->24 conversion. so even though  
the peanut gallery has repeatedly said there's no reason for a new  
filter for this purpose, i think its clear that a new filter would be  
a good idea.

anyway, at this point i have 2 versions of my filter, one that first  
turfs all odd frames, then drops the most similar frame in each group  
of 5 frames. i have another one that just keeps the two most  
different frames in each group of 5 frames. unfortunately i see  
numerous breaks in the 3/2 pattern - not just reversal of the pattern  
but sometimes there are 2 unique frames back-to-back. whats worse is  
that these seem to be purely extra frames - there is no duplicated  
frame later that one could drop to make up for the out-of-sequence  
frame. the 2nd version of the filter does slightly better against  
this problem.

these filters work very well when i split the audio from the video  
and just transcode the video alone (with -noskip), but i'm running  
into issues with the v_timer_corr code unnecessarily skipping and  
duping frames if i transcode with the audio (which means i cant use - 
mc 0 -noskip, since there is an initial PTS difference that needs to  
be corrected). that's the subject of my more recent mail to the list.  
so far no solution to that problem without hacking mencoder.

rob