[MPlayer-dev-eng] [CFTWWI] VF logodetect/logoremove

[Shachar Raindel]

On 8/12/05, Ivo <ivop at euronet.nl> wrote:
> On Thursday 11 August 2005 23:28, Shachar Raindel wrote:
> > On 8/11/05, Ivo <ivop at euronet.nl> wrote:
> > > You're right it's not usable for realtime processing. It needs
> > > thousands of itterations.
> >
> > Can't one use estimations based on the fact we are processing a video?
> > i.e.  do the full processing in the first round, than use these values
> > as the initial values for the next round, or maybe even as the final
> > values (after applying mc), saving precious time?
> 
> I assume by round, you mean the encoding pass of an n-pass encoding? Having

No. Round = each frame. Mening use the values from the first frame as
the initial values for the second frame, etc.

> to run the regularization pde algorithm three times would not be necessary.
> You could run the pde filter once and write the output to disk in a
> lossless format. Then, you use that file as input for your n-pass encoding.
> You'll need some diskspace ofcourse :)
> 
Ofcourse you run filters only once. When doing multiple encoding
passes, one wants to get exactly the same input every time.

> > Also, can't one make the process parallel, so that one could get a
> > nice small swarm of computers, each processing one frame for an hour,
> > and if we have 1000 of them, it is about 3.6 seconds/frame, which is
> > not that bad if one wish to get high quality.
> 
> I suppose you could, especially if you write the output losslessly to disk
> as described above (all I-frames). You could split the source file in 1000
> portions of frames, have each node work on that and write it all to the
> same NFS partition. The main computer then merges all these files back to
> one and starts encoding it as usual.
> 
Or hack MPI into mencoder and make it able to use a nice big cluster
for video :)