InpaintDelogo - advanced logo removal & hardcoded subtitles extraction

[real.finder]

Originally Posted by StainlessS

Not sure but I think that ExInpaint is less temporally stable than AvsInpaint, better for still image but not video.

so that mean ExInpaint use temporal dimension (The previous and next frame)? in stab3 using stab3(mirror=15,FillBorder=2,use_exinpaint=true) will give better stability than stab3(mirror=15,FillBorder=2) (which use AvsInpaint)

[StainlessS]

I think that ExInpaint is purely spatial, it predicts missing parts based on spatial structure elsewhere in the same image.
If video, then spatial predictions can be quite a lot different to adjacent frames [but indiviual frames, might look good].

I've no idea what Stab() does.

[real.finder]

Originally Posted by StainlessS

I've no idea what Stab() does.

http://avisynth.nl/index.php/Stab

it made black borders by default, so the Stab2 and Stab3 has some solutions for it

[VoodooFX]

Originally Posted by real.finder

hi,

what about make the mask parameter as val? so user can feed it with any mask clip to allow moving logo mask

also is ExInpaint instead of AvsInpaint will be better? from what I see ExInpaint is faster and give better output in some usages

What I have seen from examples, ExInpaint is good for the images, that would make it bad for video. I could add separate function for it, for images.

Not sure what the use scenario would be with "val". Is there new code for an object tracking or an object recognition? How is that half python Avisynthtrackin code? Anyone tried it? Can it be incorporated to something useful?

I remember trying Stab, but couldn't manage to get the results, so dunno what it actually does...

[StainlessS]

I tried Avisynthtrackin a little bit,
[original StarWars] on C3PO walking towards camera alone over the sand dunes.
Worked reasonably well, but IIRC required occasional user nudge/tweak to keep it working.
For above to have worked, it must continuously vary tracking stuff as C3PO got bigger and bigger as well as moving in frame.

Probably too much bother to be useful though.

[real.finder]

Originally Posted by VoodooFX

What I have seen from examples, ExInpaint is good for the images, that would make it bad for video. I could add separate function for it, for images.

Not sure what the use scenario would be with "val". Is there new code for an object tracking or an object recognition? How is that half python Avisynthtrackin code? Anyone tried it? Can it be incorporated to something useful?

I remember trying Stab, but couldn't manage to get the results, so dunno what it actually does...

val in avs is not new, with it you can make the parameter accept any type of things, you can see how SMDegrain deal with prefilter parameter for example

[VoodooFX]

While I'm at this thread, one comparison for the history; somewhere here in forum is semi-manual "Delogo" avs script which was aiming to do better than vdub plugin, I didn't tried it (don't like manual scripts), but from examples I have seen it did better job than other methods at the time. Here few comparison screenshots:

Original:


DeLogo (avs script):


InpaintDelogo:


Original:


DeLogo (avs script):


InpaintDelogo:

[VoodooFX]

Originally Posted by real.finder

val in avs is not new, with it you can make the parameter accept any type of things, you can see how SMDegrain deal with prefilter parameter for example

I meant what to do with those "any type of things"?

[real.finder]

Originally Posted by VoodooFX

I meant what to do with those "any type of things"?

as I said "so user can feed it with any mask clip to allow moving logo mask" since now it only accept image mask path

and maybe I will add InpaintDelogo to stab3 since ExInpaint crash in x64 after refresh in avspmod and AvsInpaint give flickering borders

test codes

last.ExInpaint(FillBorders_stabi(pad=2,maskonly=true,thr=30).mt_expand(U=-255,v=-255))
last.InpaintLogo(FillBorders_stabi(pad=2,maskonly=true,thr=30).mt_expand())

so maybe your script will help if it work with mask clip

[VoodooFX]

Ah, that's why. I would look into it if there was generally working script/plugin for an actual moving logo detection/tracking. Filling borders doesn't give an incentive when that's not a small tweak.
Script is using same AvsInpaint, I wouldn't expect different results. How wide those borders are and why not cut them off?

[real.finder]

Originally Posted by VoodooFX

Ah, that's why. I would look into it if there was generally working script/plugin for an actual moving logo detection/tracking. Filling borders doesn't give an incentive when that's not a small tweak.
Script is using same AvsInpaint, I wouldn't expect different results. How wide those borders are and why not cut them off?

for moving logo, I think if the move not random then user can make mask using http://avisynth.nl/index.php/Animate and the logo

the borders width depend on the shaking fix and the stab settings, cut them off not always a good idea especially when you deal with 480p input that you have to cut the black borders in it and then use stab(), and it's worst if it have a big black borders up and down (16:9 in 4:3)

[VoodooFX]

Originally Posted by StainlessS

I tried Avisynthtrackin a little bit,
[original StarWars] on C3PO walking towards camera alone over the sand dunes.
Worked reasonably well, but IIRC required occasional user nudge/tweak to keep it working.
For above to have worked, it must continuously vary tracking stuff as C3PO got bigger and bigger as well as moving in frame.

Probably too much bother to be useful though.

If it can track an object with the changing shapes, then it should track movement of the static shape object much better, no? What if such object jumps long distances? I personally don't have time to test it, left comment there https://forum.doom9.org/showthread.p...10#post1926710, maybe someone interested in such logos will give it a go.

Somewhere I've seen that you wrote some motion blending script to isolate a moving logo, did you tried to adapt it to InpaintDelogo? Every time I look at the code of ID it makes my head spin for a bit...
How reliable such isolation is?

Originally Posted by real.finder

...it's worst if it have a big black borders up and down (16:9 in 4:3)

That's way beyond of practical use. Imo practical threshold would be somewhere bellow ~20px.

Now I'm trying to remove a static animated logo, a running timer, that would be a bias to what you are asking for.

[real.finder]

Originally Posted by VoodooFX

That's way beyond of practical use. Imo practical threshold would be somewhere bellow ~20px.

Now I'm trying to remove a static animated logo, a running timer, that would be a bias to what you are asking for.

indeed, but I didn't mean that, I was mean after crop them and use stab

and thanks about trying support it

[VoodooFX]

Originally Posted by real.finder

and thanks about trying support it

I started tinkering around it; so you want to supply there a whole clip aka dynamic mask or just a one frame mask?

There is lots of changes accumulated in the script over the year, I'll publish it after I'll add all new stuff to the manual.
I hate writing manuals, but there is no point releasing without them...

[real.finder]

Originally Posted by VoodooFX

I started tinkering around it; so you want to supply there a whole clip aka dynamic mask or just a one frame mask?

There is lots of changes accumulated in the script over the year, I'll publish it after I'll add all new stuff to the manual.
I hate writing manuals, but there is no point releasing without them...

supply a whole clip aka dynamic mask, or it's will be better if both

[VoodooFX]

Added support for dynamic/static mask clips, just need to optimize the script for the dynamic clips.

I noticed that multi-threading now works, it runs x3-x4 faster with Prefetch(4), not sure why, is it because some changes in new LSMASH indexer or AVS+ 3.5?

[kedautinh12]

Originally Posted by VoodooFX

Added support for dynamic/static mask clips, just need to optimize the script for the dynamic clips.

I noticed that multi-threading now works, it runs x3-x4 faster with Prefetch(4), not sure why, is it because some changes in new LSMASH indexer or AVS+ 3.5?

Where your download link new ver??

[VoodooFX]

Encountered some weird random bug (LSMASH 20200728 + AVS+ 3.4 r2923), first I thought that's some frame inaccuracy but probably it's bug in AVS+/Overlay or AVSinpaint, need to investigate it more. Is there some script/plugin which would compare frames from two separate video files (expected to be identical) and return frame numbers of not identical frames? Or any ideas how to do such comparison?

Originally Posted by kedautinh12

Where your download link new ver??

It will be posted when it's ready.

[VoodooFX]

No ideas how to make such comparison, that would really help to pin point the bug?

Here I made another run with few steps on the output to see more what is going on and accidentally noticed this on one frame ("before" bug was different and much severe):
Proper:

Bug:


Here it didn't influenced much the end result (left/bottom), but how in the hell top/right corner mask is red, feathered and "12" visible when it goes through Greyscale.Levels(127,1,128,0,255)...

[poisondeathray]

Originally Posted by VoodooFX

Is there some script/plugin which would compare frames from two separate video files (expected to be identical) and return frame numbers of not identical frames? Or any ideas how to do such comparison?

WriteFileIf with a runtime function expression like lumadifference . But there was some bugs in avs+ 3.x with this. It works in 2.6 classic. There was some discussion about this somewhere IIRC. Not sure what changed or what the exact syntax change was

Something like this works in avs+ 2.6 classic

.
.
.

Subtract(clip1,clip2)
WriteFileIf (last,  "log.log",  "LumaDifference(clip1, clip2)>0", "current_frame", """ ":" """, "AverageLuma")

Other approaches would be psnr . You can print out per frame PSNR in ffmpeg (should be infinity for identical frames), but that prints out every frame value. Maybe some spreadsheet or grep function to reduce to only different frames