How to improve chances of SRestore blend detection?

[ndjamena]

I'm attempting to transcode He-man and the masters of the universe at the moment, it's an NTSC version I ordered from America. I'm assuming at this point that the original frame rate for the most part was 23.976 or something but the framerate has been propped up with field-blending.

Sample:
http://www.mediafire.com/?292c3wfw1im1mrq

 
SetMemoryMax(1024)
SetMTMode(2,6)
MPEG2SOURCE("Video 1.d2v",cpu=4)
QTGMC(Preset="Placebo")
sRestore(frate=23.976,speed=-25)

which does what is implied. The thing is, I've noticed some rather obvious blends getting through despite having perfectly unblended neighbours. I've also noticed this kind of output before in other animation I've use srestore on, so I thought it was about time to ask the best ways to minimise this from happening. It's not exactly clear from the readme but I'm assuming -25 is the slowest speed with the best detection (I'm not entirely sure because it doesn't mention 0 as an option so I can't tell if positive and negative values do the same thing). From general experience up to this point I'm doubting if QTGMC on placebo is the best feed to use for detecting Blends and am wondering if there was a better bobbing option to use as a dclip. I remember trying yadif and a simple bob at some point with less than stellar results. Any constructive input would be appreciated as I've been thinking about this for a while now and am hoping to put it all to bed once and for all.

[jagabo]

SRestore(frate=25) gave much cleaner results. With or without speed=-25.

[ndjamena]

Does that mean it picks up some kind of pattern if you get the frame rate right or what? Why 25???

[jagabo]

I don't know exactly how SRestore works internally. But if you don't pick the right frame rate you get problems.

This type of conversion is almost always from using a common studio analog PAL/NTSC converter that uses field blending. Hence the input was a 25 fps PAL video.

[ndjamena]

So the masters of an official American DVD of an American Cartoon animated in the American state of California are, in fact, PAL????? I'll have to check every episode and see if they're all the same!

[poisondeathray]

Jagabo is right: this version is definitely from a PAL source . You can count the fields and look at the motion

It's actually 12.5 fps animation (every 2nd frame is a duplicate)

Srestore does a good job at default settings in single threaded mode, processing linearly - I don't see any signficant blends i in the first 20-30 seconds , or where it chose the wrong frame (which is rare) . You can adjust the threshold settings of srestore if required

2 common problems with srestore: using MT, and non linear seeking . You have to check linearly, and although some people swear MT is ok for srestore, it's a big headache for others . I find MT a big headache for all temporal filters - frames get mixed up, you get blends and weird results when you shouldn't

The 3rd less common problem is QTGMC can actually make field blending worse in some cases (it processes temporally so can make some good frames blended) . An alternative in those cases is yadifmod+nnedi3

[ndjamena]

Are you saying QTGMC is a bobber designed to eliminate 'bobbing' effect, so if a pixel changes from one colour to another due to motion from on frame to another then changes to a similar colour to the first for the third one, you'll wind up with something akin to blending in the second frame??? Or is something more sinister happening?

SRestore Tips:
1: Run single threaded.
2: High Memory Max.
3: Try different Frame Rates.
4: Set Speed to -25?????

Is there a better bobber I can use for dclip than QTGMC? Or maybe a lower setting will warp the picture less?

Jabago gave me a script to convert 'Turtles in Time' from PAL to 23.976fps that left a lot of blends behind, I want to try that one again but am not sure how to go about it at this point.

[ndjamena]

In terms of speed do I want -1 or -25 for best detection? It's not clear from the readme. Or should I just stick to 1 for something like this?

[jagabo]

So the masters of an official American DVD of an American Cartoon animated in the American state of California are, in fact, PAL????? I'll have to check every episode and they're all the same!

In all likelihood it was originally animated at 12 fps then each frame was doubled for 24 fps film. For PAL the film was sped up 25 fps and recorded interlaced (but both fields from the same film frame). That PAL tape was then converted to NTSC with an analog frame rate converter to make 59.94 fields per second NTSC video with field blending (stored digitally as 29.97i frames per second). Ie, whoever made the NTSC DVD didn't have access to the original film so they used the best they could find, a PAL video tape.

[manono]

In terms of speed do I want -1 or -25 for best detection? It's not clear from the readme. Or should I just stick to 1 for something like this?

I've never found adding anything to the basic SRestore(Frate=???) made any difference at all. As jagabo mentioned earlier, the best results will be for the correct framerate. Something you discovered by using the wrong framerate initially.

[ndjamena]

In case anyone comes across this thread looking for answers I'll add what I've learned up 'til now. In my initial processing of the Australian PAL version of Turtles Forever I used the default settings to get a framerate of 23.976 and it left a lot of blends behind. Lowering the Threshhold to 12 and setting the Speed to 1 pretty much reduced the number of blends left to Zero. Setting Cache at 50 also helps quite a lot for everything I've thrown at it and it doesn't really use much memory at all. I just Transcoded my Ghost in the Shell: Stand Alone Complex DVDs and although most of them were fully progressive, I discovered after I'd wasted a few hours on each episode that for no apparent reason a few random episodes in season 2 were interlaced and blended. It's not relevant to the thread but at this point I'm wondering about Madman Australia... Anyway, I tried:

 
 Interleave(TFM(field=1),TFM(field=0))
 sRestore(frate=23.976,speed=1,cache=50,thresh=12)

but yet again I was left with interlacing in the final output. I don't know why it keeps doing that, could I have the wrong version of a plugin or something? I was left with no option but using the QTGMC animation setting:

 
 QTGMC("Placebo",TR0=1,TR1=1,TR2=0)

(I really need that in a preset so I don't have to keep googling for -Vit-'s post.)
I then ran that through sRestore:

 
 sRestore(frate=23.976,speed=1,cache=50,thresh=12)

and that did a rather good job but there were still some rather obvious blends left. I could see why srestore was having trouble with those few frames but the only way I could get better detection at this stage was by pushing 'speed' into negative territory, which leaves the question:

In terms of speed do I want -1 or -25 for best detection? It's not clear from the readme. Or should I just stick to 1 for something like this?

which up to that point no one had answered. So I referred to the SRestore avsi file (or avs file - depending on how the user names it) and a quick search for 'speed' using notepad showed only 2 references to the variable at the beginning of the script, one reference was within an abs statement and the other checked whether it's a negative value or not, then sets another variable to 16 or 32 depending on the answer. So from that I guessed that the negative check determines whether Chroma is used for detection and the abs reference determines the strength: which would mean a speed of -1 is the Chroma equivalent of speed 1. I ran the QTGMC'd file through SRetore again using these settings:

 
 sRestore(frate=23.976,speed=-1,cache=50,thresh=12)

The Result: It's not feasible to check through every single frame of every single episode but as far as I can tell WITHOUT EXEPTION, EVERY SINGLE BLEND IS GONE!!!
So that's about it, unless someone has something else to add, those are the ultimate sRestore settings for use with clean animation that has been QTGMC'd. The threshold may need to be adjusted to account for noise or a different bobber, but that is to be expected and is why the setting exists in the first place. I'm done, and hopefully someone else finds this information useful.

[ndjamena]

OK, I may be being a pain, but I needed a thresh of 8 to get rid of this one:



There's another frame in the sequence that failed to be removed even at thresh=4 - it amounts to just a slight discolouration that isn't even noticeable when looking at just the frame itself. I'd have to post three pictures to show what I'm talking about and so I won't but since I'm sure there must be consequences to setting the threshold that low, much less setting it even lower, I'll just have to leave it there.

{On the one hand, I may be obsessing over something everyone else already knows, on the other hand, if they already knew, why didn't they tell me when I asked?}

[jagabo]

You're right, the Interleave(TFM,TFM) bob doesn't work well with field blended material because it always uses an adjacent field to fill the frame. If the closest match is a blended field TFM is forced to use it (whereas a filter like QTGMC is free to ignore it). Its post processing usually removes (blends) the comb artifacts (you can use pp=0 to turn it off and see the full combing) but there's often some combing left behind. In the end, you end up with more blended frames than with other bob filters. The hope is that SRestore will remove those bad frames but sometimes there are just too many.

On the other hand, QTGMC sometimes creates its own blending artifacts. So sources with no field blending sometimes give better results with Interleave(TFM,TFM).

Good work with the SRestore tuning though. I'll keep that in mind when I use the filter.

[provato]

I'm sorry to bring up this old thread, but I'm experimenting with srestore, and wanted to report something strange, in case someone can explain it.

I have an NTSC X-men animated DVD, which I want to backup, and I tried ndjamena's qtgmc+srestore settings (sort of...) to get rid of the most doubled frames I can. Here is the exact script I used:

# Set DAR in encoder to 4 : 3. The following line is for automatic signalling
global MeGUI_darx = 4
global MeGUI_dary = 3
DGDecode_mpeg2source("C:\VTS_08_1.d2v")
QTGMC(TR0=1,TR1=1,TR2=0,edithreads=4)
LoadPlugin("MeGUI\tools\avisynth_plugin\TIVTC.dll")
sRestore(frate=25,omode=5,speed=-1,cache=50,thresh=4)
crop(6, 0, -2, 0)
Spline64Resize(720,576) # Spline64 (Sharp)
Undot() # Minimal Noise

Note 1: I used "frate=25" because jagabo had told me that this DVD was made following this procedure: FILM Source --> PAL Tape (field blended) --> NTSC DVD

Note 2: I loaded TIVTC plugin because "cache=50" wouldn't work if I didn't!!!

Note 3: I run this script in MeGUI (I'm writing this even though it's quite obvious...)

My Question:

Video turns out to have 9/10 less doubled frames and motion appears smooth enough, but when I check in Mediainfo, it says that the video is 59,94 fps progressive!!! (as in 29,97 x 2 if we had to do with fields...). Why has this framerate happened?

[ndjamena]

I'm still stuck on an iPhone so I don't have much capacity to help but I'd just like to point out that in the off time I took another look at the srestore script. The script is confusing and I realise now that a lot of what I surmised about how it worked was wrong. Setting speed to -1 should give the best results but using negative speed level does not activate chroma detection, it in fact doubles the size of a bilinear linear resize somewhere in the script, I'm not sure what it does exactly but it does help. To activate chroma detection you need to set MODE to negative, which would generally mean mode = -2. Srestore will actually covert the chroma to luma and combine it all into a single monochrome image using stack vertical/horizontal. The cache setting in fact calls a request to RequestLinear at the end of the script which does in fact require TIVTC and may or may not be useful. I know requesting frames from srestore out of order can be a problem and that it can give different results depending on where it starts. Using cache may for the most part be redundant, but I still think unless it's causing problems cache should always be used as a precautionary measure. The final thing is the speed parameter settings are kind of a mess. The srestore manual says that setting speed to 25 will result in one out of twenty five pixels will be tested. Presumably a setting of 23 will lead to 1 out of 23 being tested. Looking at the calculation being made in the actual script says differently. I don't have the script in front of me but from what I remember speed settings use significantly less pixels than the manual states. The only thing for certain is the a speed of one checks EVERY pixel, every setting above that results in rather severe loss of pixel count, and since the calculations must adhere to mod 4 dimensions a lot of the numbers between 1 and 25 are equivalent, even at 1080p. I'm not sure if that's a bug in srestore, a misunderstanding from the author of the manual, or a revision added to srestore but it's worth pointing out. At the moment I'm using speed -1, mode -2, cache 50, thresh 6. I can't really help beyond that.

[provato]

Ok, thanks for the info ndjanema, but I'm simply wondering, why when I use srestore(frate=25) on an NTSC field blended source, I get 59,94 fps progressive?

[manono]

From the SRestore doc:

omode
Srestore can be used as blend-decimation-function, as simple deblend-function and for double-blend-removal
1 - deblend-mode 1 -> detected blends will be replaced with the previous frame
2 - deblend-mode 2 -> next frame is used instead of the detected blend
3 - deblend-mode 3 -> detected blends will be replaced with the neighbour that has the smaller difference
4 - deblend-mode 4 -> use the neighbour with the smaller blend-possibility
5 - deblend-special -> outputs the one of four frames with the smallest blend-possibility
>5 - blend-decimation -> for all decimation operations
The output-modes 1-5 are simple deblending modes, so the framerate will not change.

And since the framerate when it gets to SRestore is double the source, you get 59.94fps because FRate=25 does nothing. Remove omode=5 from the script and try again.

Note 1: I used "frate=25" because jagabo had told me that this DVD was made following this procedure: FILM Source --> PAL Tape (field blended) --> NTSC DVD

jagabo didn't suggest using an omode in there, did he?

[provato]

I never implied or told that jagabo told me to use omode. I just had better luck taking care of doubled-frames with omode.

Anyway, thanks for the solution-answer manono.

Still one question: any way to get the omode=5 result but with a framerate conversion 29,97 -> 25? (I know about selecteven(), but I still need frate=25 to work)

[manono]

Still one question: any way to get the omode=5 result but with a framerate conversion 29,97 -> 25?

You can use TDecimate afterwards.

[provato]

Thanks again Manono, could you please tell me if I'm wrong here? the plan is:

QTGMC() For deinterlacing
srestore(omode=5) for blend-removing
selecteven() for 29,97 fps
tdecimate() to convert to 23,976 fps
AssumeFPS(25) if needed to adjust/synchronize the video's length to the PAL audio's length

right?

[manono]

You said jagabo told you it was PAL To NTSC. I haven't seen any samples but if it's as he said then you don't want to decimate to 23.976fps unless you want a slight jerk every second from missing frames. You'd decimate to 25fps (or maybe 24.975fps) which will keep the audio in synch. You may (or may not) want to slow it to film speed afterwards (and also slow the audio to match).

[provato]

You said jagabo told you it was PAL To NTSC. I haven't seen any samples but if it's as he said then you don't want to decimate to 23.976fps unless you want a slight jerk every second from missing frames. You'd decimate to 25fps (or maybe 24.975fps) which will keep the audio in synch. You may (or may not) want to slow it to film speed afterwards (and also slow the audio to match).

Οκ, I'll try both ways then to see which one feels more "jerky" (the tdecimate to 24,975 or to 23,976), and I'll report back.
Thanks again manono, you've offered me great help

[provato]

Update: Both ways feel completely jerky.... I don't know how they made this DVD, but seems that in quick movement, there is not even one un-blended frame... It also felt a little jerky playing from the DVD itself.
Maybe this DVD source (X-men animated series, I had uploaded part of it in another thread) followed some weird road to be done, like Film->NTSC 3:2 tape->PAL field blended tape->NTSC double field blended tape....

Anyway for 24,975 fps I used Decimate(6), correct?

[manono]

It also felt a little jerky playing from the DVD itself.

Animations by their nature are much more jerky than live action film. Can you point to the sample on the other thread?

Anyway for 24,975 fps I used Decimate(6), correct?

Yes, that's one way.

[provato]

Animations by their nature are much more jerky than live action film. Can you point to the sample on the other thread?

I deleted the VOB sample I had uploaded in the other thread. I will upload a new one tomorrow.
I will be so greatful if there is a way to minimize both blends + jerkiness of these video episodes. This series has a DVD after 20 years of many people waiting, and Marvel + Clearvision screwed it up...

[provato]

Sorry for double post (again). Here is the VOB sample:

https://mega.co.nz/#!0ok1AAQR!Ln6qONZYi7zaxctkW5eVPEklWPteAN7GJ2vgGN7B2KA

[manono]

Yeah, it's a mess alright. I think I'd keep it at 29.97fps. It's got way too many blends to be able to unblend it successfully. And it plays more smoothly at 29.97fps than at anything else.

Yadif(Mode=1,Order=1)#or QTGMC
Srestore(Frate=29.97)

[provato]

Ok, manono thanks once again for sharing your experienced opinion on this video.
I guess I'll just leave it to 59,94 progressive (double rate) then, which has the smoothest and most unblended playback (with srestore(omode))

[ndjamena]

I still have this spreadsheet on my HDD so I thought I'd post it here in case anyone is interested.

These are the relevant lines of code to determine speed:

srad  = isfloat(speed) && abs(speed)>=1 ? sqrt(abs(speed))*4 : 12

det	= det.pointresize(srad==4 ? det.width : int(det.width/2/srad+4)*4, srad==4 ? det.height : int(det.height/2/srad+4)*4).trim(2, 0)

This is what the readme says about speed:

speed
With the speed parameter you can set a speed-vs-detail-detection value. If you set speed=25, only one of 25
Pixels (5*5) is used for the detection. Higher values result in a lower possibility of low-detail detection.
Set a negative speed value for high quality and hd sources. This can improve the detection of small details.
Default - 9 (recommed range: 1-25, same for the negative values)

Yet according to my spreadsheet setting speed to 25 results in 1 out of 65 pixels being used for detection, not 1 of 25.

To get the results the readme describes you'd need the second line to read like this:

det	= det.pointresize(int(det.width/srad)*4, int(det.height/srad)*4).trim(2, 0)

You can see how several speed settings give the same result, even 1080p at speeds of 24 or 25.

Ignore the numbers for speed=1, that's just what it would be without the conditional operator that overrides it: (srad==4) ?

The difference in the code definitely looks deliberate, and I guess it's not actually important. But there you go.

speed		srad	bsize	720	576	414720		1920	1080	2073600	
-25	20		32	88	72	6336	1 in 65	208	124	25792	1 in 80
-24	19.59591794	32	88	72	6336	65	208	124	25792	80
-23	19.18332609	32	88	76	6688	62	216	128	27648	75
-22	18.76166304	32	92	76	6992	59	220	128	28160	73
-21	18.33030278	32	92	76	6992	59	224	132	29568	70
-20	17.88854382	32	96	80	7680	54	228	136	31008	66
-19	17.43559577	32	96	80	7680	54	236	136	32096	64
-18	16.97056275	32	100	80	8000	51	240	140	33600	61
-17	16.4924225	32	100	84	8400	49	248	144	35712	58
-16	16		32	104	88	9152	45	256	148	37888	54
-15	15.49193338	32	108	88	9504	43	260	152	39520	52
-14	14.96662955	32	112	92	10304	40	272	160	43520	47
-13	14.4222051	32	112	92	10304	40	280	164	45920	45
-12	13.85640646	32	116	96	11136	37	292	168	49056	42
-11	13.26649916	32	124	100	12400	33	304	176	53504	38
-10	12.64911064	32	128	104	13312	31	316	184	58144	35
-9	12		32	136	112	15232	27	336	196	65856	31
-8	11.3137085	32	140	116	16240	25	352	204	71808	28
-7	10.58300524	32	152	124	18848	22	376	220	82720	25
-6	9.797958971	32	160	132	21120	19	404	236	95344	21
-5	8.94427191	32	176	144	25344	16	444	256	113664	18
-4	8		32	196	160	31360	13	496	284	140864	14
-3	6.92820323	32	220	180	39600	10	568	324	184032	11
-2	5.656854249	32	268	216	57888	7	692	396	274032	7
-1	4		32	376	304	114304	3	976	556	542656	3
0	12		32	136	112	15232	27	336	196	65856	31
1	4		16	376	304	114304	3	976	556	542656	3
2	5.656854249	16	268	216	57888	7	692	396	274032	7
3	6.92820323	16	220	180	39600	10	568	324	184032	11
4	8		16	196	160	31360	13	496	284	140864	14
5	8.94427191	16	176	144	25344	16	444	256	113664	18
6	9.797958971	16	160	132	21120	19	404	236	95344	21
7	10.58300524	16	152	124	18848	22	376	220	82720	25
8	11.3137085	16	140	116	16240	25	352	204	71808	28
9	12		16	136	112	15232	27	336	196	65856	31
10	12.64911064	16	128	104	13312	31	316	184	58144	35
11	13.26649916	16	124	100	12400	33	304	176	53504	38
12	13.85640646	16	116	96	11136	37	292	168	49056	42
13	14.4222051	16	112	92	10304	40	280	164	45920	45
14	14.96662955	16	112	92	10304	40	272	160	43520	47
15	15.49193338	16	108	88	9504	43	260	152	39520	52
16	16		16	104	88	9152	45	256	148	37888	54
17	16.4924225	16	100	84	8400	49	248	144	35712	58
18	16.97056275	16	100	80	8000	51	240	140	33600	61
19	17.43559577	16	96	80	7680	54	236	136	32096	64
20	17.88854382	16	96	80	7680	54	228	136	31008	66
21	18.33030278	16	92	76	6992	59	224	132	29568	70
22	18.76166304	16	92	76	6992	59	220	128	28160	73
23	19.18332609	16	88	76	6688	62	216	128	27648	75
24	19.59591794	16	88	72	6336	65	208	124	25792	80
25	20		16	88	72	6336	65	208	124	25792	80

[Adamcarter]

Sorry to bring up an old topic
I have a similar situation.

I have a dvd by disney which i am 100% sure it was a PAL source 25fps then blended to NTSC 29.97.

If i do qtgmc bob brings my 29.97 to 59.9x
Clearly it needs to run at either 25 fps or 50 fps.
Do i set the frame rate to 25? Or 50?
Do i need to decimate? It drops it down by 20% which doesn’t seem to meet any standard!