GenMotion and DePan plugins for global motion compensation

[scharfis_brain]

okay it does work.

but:
It seems to work very unprecise
I always blend fwd & bckwd compensation within my scripts for better results, but fwd & bckwd do not always match each other, thus introducing false blending.
depanestimate seems to deliver only quantisized motion information.
Example: if a pan slowly accelerates (pan is getting faster and faster), the motioncompensation doesn't adapt smoothly. It compensates in 'steps'

Depan doesn't compensate zooms.

deshakers analyse is much better compared to v0.2 of depan.

okay, enough critisism

the concept of master/slave is pretty cool!

I've made some tests with my smoothpan standards conversion.
It works! But it isn't very unusable for now

can't wait the next version of depan

would it be a copyright violation just using deshaker's analyse code within depan?

[Fizick]

The motioncompensation doesn't adapt smoothly. It compensates in 'steps'

Try subpixel=2 (bicubic). Default is =0 (nearest pixel), in next version it will be =2. It is document bug. (also i know about Depanstabilize bug with input paarameters - will be fixed later).

Depan doesn't compensate zooms.

Yes, but it may be implemented later.

deshakers analyse is much better compared to v0.2 of depan.

Yes, but it often has error with rotation.
And depan method (phase-shift) is more fast.

would it be a copyright violation just using deshaker's analyse code within depan?

What code? No any source.
May be, I shall add reading of Deshaker log, as in genmotion, but it only useful with zoom added. I do'nt want to add rotation.

[Fizick]

New DePan version 0.3, May 15, 2004 - fixed bug with DePanStabilize input parameters, set subpixel=2 (bicubic) as default, add zoom estimation and ZoomMax parameter, zoom and rotation compensation, log file input, MatchFields parameter (for best denoising).

[Fizick]

New DePan version 0.5:
improve DepanStabilize and some bug fixed.

Anybody test the plugin? Or it is fully unusable?
Any proposal for next version?

[scharfis_brain]

I do not have access to my PCs to test all the new stuff.
For at least the whole next week

please stay tuned. I'll post my results asap, I can get back to the faster machine...

(or: recompile for avisynth 2.08 and Pentium 1, than I can test it on my 120MHz Notebook)

[Fizick]

Version 0.6, May 28, 2004 - minor changed and documented client-server format.
DepanEstimate: add stab parameter, scenechange at sharp trust changes, range may be 0, show correlation, crop output.
DepanStabilize: Freqmax is renamed to Cutoff, add adaptive zoom, Fill borders by neibours frames.

Version 0.7, May 30, 2004 - DepanEstimate: add improve zoom estimation.

[Fizick]

I release a new version of Depan, developed about month ago.
Version 0.9, June 13, 2004 - all clients: added "Mirror" parameter to fill empty borders.

mirror parameter - fill empty borders with mirrored from frame edge pixels (instead of black):
0 - no mirror (default);
1 - top;
2 - bottom;
4 - left;
8 - right;
sum any of above - combination (15 - all ).

Mirror mode is unique but slightly strange
It can give remarkable results for example for sky.

[Fizick]

Serg Belyansky,
You wrote:

Unfortunately one problem persists: any solid surface with a large number of frequent simular elements (e.g. brick wall or polka-dot dress) will exhibit an erratic motion estimation with elements disappearing and reappearing randomly. It can be partially remedied by increasing the "range" parameter of Depanestimate, but it slows the process and introduces artifacts on other surfaces. So I'm still scratching my head.

It is a hard problem with no simple and general answer.
But I can give one idea to you (seems i omitted it in doc).
Input clip for DePanEstimate may be not strictly same as source,
but prefiltered, or brightness-contrast adjusted, or masked by some method, or cropped, and so on.

If you get some new ideas from your head after scratching, please post here they, or at least the your scratching method.
And do'nt use my DeScratch filter until !

P.S. It seems, i found some nice (and big!) Russian forum thread about Avisynth.
http://forum.ixbt.com/topic.cgi?id=29:9331

[scharfis_brain]

thanks for that great tool, it helps a lot denoising shaky/panning analogue camcorder video.

but there seems to be one issue:

everytime I try to enable zoom compensation, using zoommax > 1, I receive the following error MSG from AVISynth:

AVISynth read error: Avisynth: caught an access violation at 0x025d80f5, attempting to readfrom 0x025ffb10

it seems that there is a memory leak, cause, the value of setmemorymax() is in direct corellation to the framecount that is rendered, until avisynth crashes.

[Fizick]

I found some memory leakage in Depan, but only at constuctor stage (not per frame).
I can not get such errors.
May you give me more info? (Avisynth and plugin versions, script, CPU, etc)

[scharfis_brain]

using this (or similar script)

i=converttoYV12()
mdata=i.DePanEstimate(zoommax=1,improve=false,pixaspect=1.094)
i.DePanInterleave(data=mdata,pixaspect=1.094)
temporalsoften(1,10,13)
selectevery(3,1)

when I set zoommax=2 and improve=true

avisynth crashes after some frames (depends on the value of setmemorymax) with above mentioned error msg.

I am using:
newest AVS build.
newest DePan release
AthlonXP 1600+ with ASUS A7VE-266 and 768 MB of DDR-RAM

[Fizick]

I tried similar script.
At what frame number it crashes (100, 1000, 10000, 100000 )?
And for what Setmemorymax ?
Newest AVS build (20.08.2004) is buggy. See development forum.

[Fizick]

I am sorry, but i really try version 0.8
0.9 is buggy, yes. (i almost did not test it )

So, good news:
New version 0.9.1 is out. Seems, zoom bug is fixed.

[scharfis_brain]

Many thanks for this bugfix

But: I finished my current project before you did the bugfix

nontheless, DePan becomes my standard aid for denoising handycam footage. it is just great, cause it is the only thing allowing low contrasted moving detail (like grey trexturized walls) being denoised while their texture remains untouched.

a blockbased motioncompensation without global motion compensation first, is not able to catch this correctly...

[Fizick]

I think about using more fast library (FFTW).

[Serg Belyansky]

Originally posted by Fizick
Input clip for DePanEstimate may be not strictly same as source,
but prefiltered, or brightness-contrast adjusted, or masked by some method, or cropped, and so on.

Yes, I tried it this way. And I must say I'm now think that I have no clear understanding of the process at all.

The motion compensation helps to match neighbor frames, right? But DePan works with global motion only, compensating only for "main" object (background or close-up), not both. This the local motion of "polka-dotted" dress should not affect it at all. And judging by undecimated debug output, it actually does not.

I tried MVTools, which is by design "local" and should compensate for _everything_, making the perfect match between frames. But that's NOT IT -- it actually makes DeSpot to work much worse, with a lot of nasty edge effects on moving subjects. The "disappearing polka-dot" problem is somewhat remedied, but I can't say that the difference is easily noticeable.

So I think I just don't understand the "motion compensation" part well. For example, what "motpn" parameter exactly do with local motion compensation performed?

Thresholding with "sign" parameter helps, but not much, since I can't set an actual threshold: the "polka dots" are not as bright as brightest parts of dirt blobs, but they are still falsely keyed and removed by DeSpot (my film dirt is bright white, scanned from negative).

(Bugreport: the "sign" parameter description in Russian manual is erratic; the values are mixed up.)

P.S. It seems, i found some nice (and big!) Russian forum thread about Avisynth.
http://forum.ixbt.com/topic.cgi?id=29:9331 [/B]

Yes, but my firewall prohibits me from posting there

[Fizick]

Serg Belyansky,
Your (our) problem is serious.
Yes, DePan works with global motion only.

And DesPot is not ideal solution still.
What "motpn" parameter exactly do with local motion compensation performed? - I wrote about it in doc, you also may see a C source.

BTW, I do not understand, what is "polka-dots" ?
May your post some example image? (Or in russian)

If the most question now is about Despot, please return to Despot thread. May be we can improve it with some ideas (not detect very small dots or very round spots ).

Thanks for discussion!

[hartford]

Polka Dots

[Dreassica]

My guess is that he means dot crawl and the like.. could be wrong thre.

[Serg Belyansky]

Originally posted by Dreassica
My guess is that he means dot crawl and the like.. could be wrong thre.

No-no-no-no. We are talking about regular structures in an image (such as dress fabric pattern or brick wall pattern) which are prone to be partially removed by DeSpot side-effect.

Fizick, I'll try to look at the sources. Maybe the "sign" parameter can be improved by introducing the thresholds for it.