InpaintDelogo - advanced logo removal & hardcoded subtitles extraction - Page 6

StainlessS · 8th November 2020, 05:53

Quote:

Originally Posted by VoodooFX

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)...

Dont see why it should make it red, however should for YUV use

Code:

Greyscale.Levels(127,1,128,0,255,coring=false)...

VoodooFX · 8th November 2020, 18:50

Quote:

Originally Posted by StainlessS

Dont see why it should make it red, however should for YUV use

Code:

Greyscale.Levels(127,1,128,0,255,coring=false)...

It's random/non-reproducible bugs, when using Prefetch(4). Masking and Levels(127,1,128,0,255) happens in RGB32.

Quote:

Originally Posted by poisondeathray

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

Code:

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

Thanks mate, that works perfectly. I edited it a bit so I can copy/paste log to SelectEvery:

Code:

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

AvS 2.60(2.6.0.6) & AvS+ 3.5 (r3106) produced identical logs, AvS+ was 20% faster.

I run comparison on stacked output with 4 steps (like in screenshot in the previous page), so, Prefetch(4) vs no Prefetch:
Detected discrepancies in 157 frames from 134237 frames, almost all of them in the top/right mask, in 18 frames those mask are ok but discrepancies are in the final step(bottom/left, but no visible differences).

Here is some examples of the bugged masks (it should be B/W with white numbers, last two digits shouldn't be visible), stacked original clip & one mask:

Now I need to run more tests to isolate the problem.

poisondeathray · 8th November 2020, 20:20

What format was source video ? Did you try threads=1 for LSmash ? Or other source filters to rule out some issue upstream ?

To be clear, was the issue only with Prefetch(4) ? Is no prefetch ok ? Or other values ?

I recall some discussion about source filter threads and seeking accuracy issues vs. host CPU threads , but I think that was for ffms2

VoodooFX · 9th November 2020, 13:28

Like I suspected it's something with Overlay, tested and same on AvS+ 3.4, 3.5.1, 3.6.0, 3.6.1.
Here is minimized and ready to run test scripts (extract to C:\) or preview included test results https://anonfiles.com/F7E0hdo2p0 or https://we.tl/t-J95N1CsfUz (7 days)

Quote:

Originally Posted by poisondeathray

What format was source video ? Did you try threads=1 for LSmash ? Or other source filters to rule out some issue upstream ?

To be clear, was the issue only with Prefetch(4) ? Is no prefetch ok ?

Isn't "threads=1" parameter for ffms2!? No-Prefetch is ok. I dunno working alternatives for LSMASH, and whatever source filter outputs it shouldn't bypass binarization of the frames.

pinterf · 11th November 2020, 19:03

Quote:

Originally Posted by VoodooFX

Like I suspected it's something with Overlay, tested and same on AvS+ 3.4, 3.5.1, 3.6.0, 3.6.1.
Here is minimized and ready to run test scripts (extract to C:\) or preview included test results https://anonfiles.com/F7E0hdo2p0

"Bad gateway", expired perhaps?

StainlessS · 11th November 2020, 19:10

Same here, 502 Bad gateway. nginx.
Maybe that nginx thing is the block on your location for download, (cant remember what its called).
Can anybody launch successful download ?

EDIT: The Bad Gateway page is showing as Sweden in FlagFox. [dont know if relevant]

VoodooFX · 11th November 2020, 19:48

Quote:

Originally Posted by pinterf

"Bad gateway", expired perhaps?

Still works on my side. Here is another link: https://we.tl/t-J95N1CsfUz

poisondeathray · 11th November 2020, 20:08

Quote:

Originally Posted by VoodooFX

Isn't "threads=1" parameter for ffms2!? No-Prefetch is ok. I dunno working alternatives for LSMASH, and whatever source filter outputs it shouldn't bypass binarization of the frames.

LSmash has threads=blah parameter too. But non frame accurate runs will account for differences between repeated tests, when prefetch is used

I repeated test with x64 version - I think the problem is lsmash threads=auto vs. threads=1 for x64. I think it' s similar to the aforementioned ffms2 threads bug with prefetch

EDIT - multiple consec runs can result in different results, even with source filter threads=1 . One run can show PSNR inf, but another something lower. Something else is going on...

It even occurs when source is encoded to I-frame UTVideo, and using AVISource instead; multiple runs can show different results when prefetch is used. So not source filter related

VoodooFX · 11th November 2020, 20:46

Quote:

Originally Posted by poisondeathray

I repeated test with x64 version - I think the problem is lsmash threads=auto vs. threads=1 for x64.

I tested threads=1 like you did, only with 32bit and got one frame bugged, so bug is there with threads=1 too. It makes less chance for bug to occur, 5mins is too short test for this condition.

Looking at those bugged masks I think it's something with buffer/memory, just a guess, I'm not into lower level programming.

poisondeathray · 11th November 2020, 20:48

Quote:

Originally Posted by VoodooFX

I tested threads=1 like you did, only with 32bit and got one frame bugged, so bug is there with threads=1 too.

Yes, it's almost "random", see my edits above . One run can be perfect, but the next a few frames off, the next perfect again. It's not source filter related

poisondeathray · 11th November 2020, 21:05

If you change the script to use mt_merge (and full range YUV masks , PC matrix if converting from RGB), 10/10 runs are the same with prefetch=4. So that suggests it is something with overlay and prefetch

pinterf · 11th November 2020, 21:51

Quote:

Originally Posted by poisondeathray

Yes, it's almost "random", see my edits above . One run can be perfect, but the next a few frames off, the next perfect again. It's not source filter related

Here is my minimalized script.

Code:

# Load VirtualDub, Press F2, go to frame 2-3-4 and back, refresh, step between frames left and right
# suddenly a garbage is seen
DynamicMask = BlankClip(1000, 128, 128, pixel_type="RGB32")
WhiteClip   = BlankClip(DynamicMask,color=$FFFFFF)
BlackClip   = BlankClip(DynamicMask,color=$000000)
BlackClip.Overlay(WhiteClip, Mask=BlackClip)
Prefetch(4)

Thank you for giving me the opportunity to work on it

VoodooFX · 12th November 2020, 11:09

Quote:

Originally Posted by pinterf

Here is my minimalized script.

Thank you for giving me the opportunity to work on it

Yep, Vdub2 reveals this bug on the first frames. Have fun with it.

pinterf · 12th November 2020, 12:24

Please check a (hopefully) fixed Avisynth build

VoodooFX · 26th November 2020, 18:43

There is some bug with high resolutions like 1920x1080 and AvsInpaint, if a mask is close to the left/right side then AvsInpaint hangs at start (no error).

Here is the bug simulator:

Code:

x=BlankClip(length=1000, width=1920, height=1080, color=$000000, pixel_type="YV12")

mask = BlankClip(x, length=1).Subtitle("Sample", text_color=$FFFFFF, x=8).Levels(180,1,181,0,255, coring=false)
# mask = BlankClip(x, length=1).Subtitle("Sample", text_color=$FFFFFF, x=519).Levels(180,1,181,0,255, coring=false) # Working position, moving the mask 1px to left("x=518") - hangs.

# x.InpaintLogo(mask, PreBlurSize=0.9, PostBlurSize=0.9)  # Droping the blur kernels to 0.9 - no hangs.
x.InpaintLogo(mask)

EDIT:
Bug should be fixed in AvsInpaint v1.3

VoodooFX · 28th November 2020, 18:05

New version released (updated the first post):

Code:

# Version 1.11 : 2019/08/31.  First public release.
# Version 1.12 : 2019/09/06.  Various tweaks. New parameters. "Inflate" and "Deep" changed.
# Version 1.13 : 2019/12/08.  Allow -ve LogoX,LogoY, and +ve LogoW,LogoH in Loc string (by StainlessS).
# Version 1.14 : [non public] "Fr1"/"Fr2" are renamed to "FrB"/"FrW". "GrainPP" is off by default.
# Version 1.15 : [non public] New mode for "InterpM", (-1) takes mask from new "iTune" parameter.
# Version 1.16 : [non public] New "diPP" & "diPPm" parameters.
# Version 1.17 : [non public] New "iTinflate" & "oTinflate" parameters. Interp/iTune mask to "mShow".
# Version 1.18 : [non public] New "EdgePP", "EdgePos" & "EdgeWide" parameters. EdgeMask to "mShow".
# Version 1.19 : [non public] New "IntSpd" parameter for "Interp" & the slow modes for "EdgePP".
# Version 1.20 : [non public] "Turbo" presets moved from soft to hard. iTune & Edge masks to "Show".
# Version 1.21 : [non public] Added support for the mask clips and a dynamic mask. 
#                             Added basic auto generation of a dynamic mask (aka "DynMask=2").
#                             New parameters: "DynMask", "DynInflate", "DynTune", "KillNoise", "maskPatch".
# Version 1.22 : [non public] Optimized speed for a dynamic mask, MaskTools2 required.
# Version 1.23 : [non public] Added auto generation of a dynamic mask with ClipBlend (aka "DynMask=3").
#                             New parameters: "Dyn3Seq", "ClpBlend", "DynPostTune", "Dyn3buffer".  
# Version 1.24 : 2020/11/28.  Various tweaks. "Show" 3 & 4. Updated the manual with the new stuff.

One example of the dynamic logo/delogo is in the thread there -> Detect Position of Watermark in movie and remove with logo inpaint

Quote:

Originally Posted by real.finder

maybe your script will help if it work with mask clip

Check if it helps anything.

real.finder · 28th November 2020, 18:34

Quote:

Originally Posted by VoodooFX

New version released (updated the first post)
Check if it helps anything.

thanks! but it's still in "Attachments Pending Approval", can you upload it else where?

VoodooFX · 28th November 2020, 19:53

Here are screens for another dynamic logo removal:

12s GIF sample (stacked: original, dynamic mask, delogo output):

manono · 28th November 2020, 22:11

It's approved now. As I don't read every post in every forum, if you have something needing approval, you can PM me if you like.

StainlessS · 29th November 2020, 01:47

In your [@ VoodooFX] post #118 attachment [12s GIF sample], I earlier saw a white flash in bottom image part, it is now gone.
[maybe at end of repeated loop]
EDIT: attachment update, ? EDIT: your shit, is "The Shit" (well good man)

11th November 2020, 19:10	#106 \| Link
StainlessS HeartlessS Usurer Join Date: Dec 2009 Location: Over the rainbow Posts: 10,980	Same here, 502 Bad gateway. nginx. Maybe that nginx thing is the block on your location for download, (cant remember what its called). Can anybody launch successful download ? EDIT: The Bad Gateway page is showing as Sweden in FlagFox. [dont know if relevant] __________________ I sometimes post sober. StainlessS@MediaFire ::: AND/OR ::: StainlessS@SendSpace "Some infinities are bigger than other infinities", but how many of them are infinitely bigger ??? Last edited by StainlessS; 11th November 2020 at 19:13.

12th November 2020, 12:24	#114 \| Link
pinterf Registered User Join Date: Jan 2014 Posts: 2,314	Please check a (hopefully) fixed Avisynth build __________________ AviSynth+ on github, Other repos: RgTools, Masktools2, MvTools2, TIVTC, Average

28th November 2020, 19:53	#118 \| Link
VoodooFX Banana User Join Date: Sep 2008 Posts: 989	Here are screens for another dynamic logo removal: 12s GIF sample (stacked: original, dynamic mask, delogo output): Last edited by VoodooFX; 28th November 2020 at 22:36.

29th November 2020, 01:47	#120 \| Link
StainlessS HeartlessS Usurer Join Date: Dec 2009 Location: Over the rainbow Posts: 10,980	In your [@ VoodooFX] post #118 attachment [12s GIF sample], I earlier saw a white flash in bottom image part, it is now gone. [maybe at end of repeated loop] EDIT: attachment update, ? EDIT: your shit, is "The Shit" (well good man) __________________ I sometimes post sober. StainlessS@MediaFire ::: AND/OR ::: StainlessS@SendSpace "Some infinities are bigger than other infinities", but how many of them are infinitely bigger ??? Last edited by StainlessS; 29th November 2020 at 02:56.

Thread Tools	Search this Thread
Show Printable Version Email this Page	Search this Thread: Advanced Search
Display Modes
Linear Mode Switch to Hybrid Mode Switch to Threaded Mode

8th November 2020, 20:20	#103 \| Link
poisondeathray Registered User Join Date: Sep 2007 Posts: 5,371	What format was source video ? Did you try threads=1 for LSmash ? Or other source filters to rule out some issue upstream ? To be clear, was the issue only with Prefetch(4) ? Is no prefetch ok ? Or other values ? I recall some discussion about source filter threads and seeking accuracy issues vs. host CPU threads , but I think that was for ffms2

11th November 2020, 21:05	#111 \| Link
poisondeathray Registered User Join Date: Sep 2007 Posts: 5,371	If you change the script to use mt_merge (and full range YUV masks , PC matrix if converting from RGB), 10/10 runs are the same with prefetch=4. So that suggests it is something with overlay and prefetch

28th November 2020, 22:11	#119 \| Link
manono Moderator Join Date: Oct 2001 Location: Hawaii Posts: 7,406	It's approved now. As I don't read every post in every forum, if you have something needing approval, you can PM me if you like.