MotionPerfect-like project

[bilu]

Background thread:

http://neuron2.net/ipw-web/bulletin/bb/viewtopic.php?t=167

I wonder if XVID developers knowledge about motion estimation could help developing an Avisynth filter with Dynapel's MotionPerfect-like functionality.


Bilu

[Lobuz]

That would be useful to doubling fps of movie like with Philips's system here (http://forum.doom9.org/showthread.php?s=&threadid=54192) . First slow version could be used in avisynth. And optimised version in real time playing dshow filter. So there are many interests in developing that algorithms. But we must find someone who can do it.

Reagrds
Lobuz

[bilu]

Both methods are also good for dropped frames and jerky recordings. Every video encoding (whatever the codec is) would gain from this.


Bilu

[FredThompson]

Fixing dropped frames would have to be a specific capability built-in. As I explained in the thread over at neuron2's forum, the ability to specify, programmatically, where to replace/insert frames could be very helpful.

One of the things I discussed with the DynaPel folks while trying to fix a messed up decomb was how this could allow the creation of accelartion and decelaration. I still have those emails and will provide them to anyonw who is interested in development.

[Joe Fenton]

There's already a thread discussing this.

http://forum.doom9.org/showthread.php?s=&threadid=54192&perpage=20&pagenumber=1

Tom Barry already has an AviSynth plugin to frame double using motion estimation. Go check out the thread and Tom's web page.

[FredThompson]

They're significantly different. MotionPerfect operates in 3D space and synthesizes frames at custom framerates. Tom's deinterlacer doesn't do anything near that, it's more like an intelligent bob without bobbing. Go to dynapel's site and look at the demo video and you'll understand.

[sh0dan]

... and Toms only has a maximum on a 1 pixel search radius (in a 3x3 kernel around the current pixel).

Doing motion estimation for encoding is one thing, but reconstructing frames are a whole other, and much more complex issue. I'm not trying to put you down, but using XviD's ME for frame interpolation would require a lot of additional work.

[FredThompson]

Oh, yeah. The reason I keep directing people to DynaPel's site is so they can understand MotionPerfect comes from terrain following in-flight guidance systems. This stuff doesn't come cheap, it really IS rocket science.

[Joe Fenton]

What you are talking about is a MAJOR undertaking. I can see that some people might want to work on it, but unless DynaPel releases some of their code, you're looking at a project that will be years. What Tom is doing is very simplistic in comparison, but it gives better encoding results NOW instead of later. I was just suggesting that maybe you work up to something like MotionPerfect from a more modest base. Help out Tom and eventually it acquires more of the same features as MotionPerfect.

[MfA]

Let us not excaggerate, it probably will need a lot of tweaking ... but it isnt a lot of code. It just needs someone with a fair amount of time, and who doesnt suffer from NIH syndrome.

I mean if you just said "Fuckit, Ill implement Philips's method" (from papers and patents) you could be done in a couple of days. I dont think a 10+ times slowmo would look as good as Motionperfect, but for simpler frame interpolation it should be fine. Looks good enough on TV anyway.

[bilu]

That's a good idea, to start with Philips method for educational purposes just like XVID does with MPEG-4. I bet Philips would like to see some user feedback from the video encoding scene :)

Then other motion estimation features could be worked out over this filter, which at least would be a work basis for the next chapter, Motion Perfect. :)


Bilu

[Joe Fenton]

Yes, but are there papers and patents for Philip's method, and do they give you enough info to be able to make an "educational" version? I agree that if this is available, all it takes is one motivated person to get the code out quickly. Now, where would this person find that info? How much is it? It's a fair question. When I first worked with MPEG 1, I purchased the specs as that was the only way to get them. It was close to a thousand dollars. I don't feel like doing something like that again. :D

[bilu]

Threads on Philips DNM:

http://neuron2.net/ipw-web/bulletin/bb/viewtopic.php?t=145
http://forum.doom9.org/showthread.php?s=&threadid=54192

Bilu

[MfA]

http://www.ics.ele.tue.nl/~dehaan/publications.html

[shlezman]

I believe that many methods and algorithms in that area can be found in the academic work under the name :
temporal super resolution
or
time-space super resolution
via google

Using these technics with ALOT of optimization work the smooth-rate playback filter can be written.

[morsa]

How about using a border detection code (like Msharpen has)on the first and following frame and then feed this borders as the mesh mask for Libmorph GPL library for morphing.There is also this new filter for GMC but don't know if it is of any use.
I guess this approach could be far better than Barry's Framedoubler.
Don't know, may be I'm a bit silly.
Tell me what you think.

[bilu]

@mods

Please move this thread to Avisynth Development.
I started it here at first only because I felt XVID's ME knowledge could become handy. So far it didn't, and on Avisynth Development it will bring more attention from users concerned with this kind of things IMHO.


Thanks in advance,
Bilu