[FFmpeg-devel] GSoC 2018
michael at niedermayer.cc
Mon Jan 15 02:42:04 EET 2018
On Sun, Jan 14, 2018 at 09:28:58PM -0200, Pedro Arthur wrote:
> 2018-01-13 23:32 GMT-02:00 Michael Niedermayer <michael at niedermayer.cc>:
> > On Fri, Jan 12, 2018 at 11:56:07AM -0200, Pedro Arthur wrote:
> > > 2018-01-12 0:06 GMT-02:00 Michael Niedermayer <michael at niedermayer.cc>:
> > >
> > > > if pedro is up to date on this stuff, then maybe he wants to mentor
> > this
> > > >
> > > > either way, links to relevant research, tests, literature are welcome
> > > >
> > > > I can mentor this.
> > >
> > > One of the first NN based method was  which has a very simple network
> > > layout, only 3 convolution layers. More complex methods can be found in
> > > , , .
> > > The important question is where we are going to perfom only inference,
> > > using a pre-trained net or we will also train the net. The first is more
> > > easy to do but we don't exploit the content knowledge we have, the second
> > > is more powerful as it adapts to the content but requires training which
> > > may be expensive, in this case it would be best to use some library to
> > > perform the training.
> > Iam sure our users would want to train the filter in some cases.
> > use cases for different types of content anime vs movies with actors for
> > example likely benefit from seperate training sets.
> > The training code could be seperate from the filter
> > Also another issue is the space requirements that result out of the
> > training.
> > This was an issue with NNEDI previously IIRC
> > >
> > > There are also method which does not use NN like A+  and ANR.
> > How do these perform in relation to the latest NN based solutions ?
> Comparing psnr the first NN method (SRCNN) achieves the same quality but
> evaluation is faster than A+, or better quality at same speed.
> Newer NN methods (, ) uses "perceptual loss" functions which degrades
> the psnr but the images are much more sharp and appear to have better
> quality than those that maximize psnr.
it seems PSNR does not work as a way to compare these filters
though its possible that with video instead of still images there could
be instabilities in the details created by the filters.
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
Modern terrorism, a quick summary: Need oil, start war with country that
has oil, kill hundread thousand in war. Let country fall into chaos,
be surprised about raise of fundamantalists. Drop more bombs, kill more
people, be surprised about them taking revenge and drop even more bombs
and strip your own citizens of their rights and freedoms. to be continued
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 181 bytes
Desc: not available
More information about the ffmpeg-devel