Scripts and filters for an Opera (sample videos included)

[sanlyn]

sanlyn - thanks I had no idea about the curtain. I don't have much experience with operas . With all the colored makeup, lighting, costumes (not to mention the crazy color cast ), I had no idea how it "supposed" to look like

I know whatcha mean. Like most stage work, parts of scenes are lit with amber, blue, green gels, etc. Not like movie or tv lighting at all, and it's mostly from high overhead, the sides, or balconies. I see a lot of artsy lighting in this show, much of it without purpose, but that's the way stage fans like it. Tough as hell to figure a uniform color balance.

[sanlyn]

Folks, I'm back to my early contentions about these colors. First, we, know that Blue gets gradually squished to the left. By show's end, Blue is decimated. YV12 isn't connecting Blue with luma correctly anyway. Histograms and pixel readers show Blue values changing numerically, but YUV or RGB filters can't move Blue much above the equivalent of RGB 128 or so. Try it yourself. A color filter tells you bright Blue numbers are changing, but the image shows little or no action above the high midtones.

Must be from forcing VHS YDbDr source into recorded YCbCr (stored as YV12). It just doesn't work. It's worse when converting this convoluted YCbCr/YV12 into RGB. You could bypass RGB altogether and go from a YV12 AVI directly to MPG, but it's still a mess.

Why? YUV is a general term for several color spaces. YUV means you have a Y luma channel with U and V color channels. A YUV encoder takes RGB input (each RGB pixel has data for brightness and 3 primary colors). The encoder keeps luma for each pixel in a Y channel, then maps chroma data according to one of several systems (YCbCr, YPbPr, YIQ, etc., YV12 being just one way of storing YCbCr color). You can also work the reverse: convert YUV to RGB using math that remaps luma and chroma back to RGB pixels in exactly the reverse manner in which the data first got taken out.

But each YUV scheme has a different way of mapping luma and chroma. According to one report, SECAM was created for two reasons. The first was apparently political (Ah, mon ami, ze YCrCb she is not French! Ah, but ze YDbDr! She is tres French! Les plus uniques!). Reason #2 was to rotate chroma differently (i.e, NTSC/PAL rotates chroma 30 degrees -- why, beats me) around an axis to avoid NTSC/PAL hue shifts. Look at a SECAM tv: no Hue control.

YdbDr and YCbCr map differently, so we have scaling and gamma problems out the wazoo. In YV12 or RGB you can't shift channels without eerie problems. The VOB's YV12 colors aren't mapped correctly to Y in the first place.

So how do we work YDbDr to YCbCr? I'm still looking at the MATLAB code (need some C++ and some MATLAB too). And that crazy formula: Y-luma in YDbDr is the same RGB source luma other YUV's use. But SECAM's U and V are related to other YUV's as Db = (+3.059 * U) and Dr = (-2.169 * V). I tried feeding reversals of that to AviSynth and everything went haywire.

I find no info about YDbDr->YCbCr conversions. Apparently it's YDbDr->RGB only, then from RGB to other YUV's. We've seen what happens, doing YV12->RGB with the wrong mapping.

Back to working with ColorYUV, etc., for now. It's improving. But it still reeks.

[poisondeathray]

Folks, I'm back to my early contentions about these colors. First, we, know that Blue gets gradually squished to the left. By show's end, Blue is decimated. YV12 isn't connecting Blue with luma correctly anyway. Histograms and pixel readers show Blue values changing numerically, but YUV or RGB filters can't move Blue much above the equivalent of RGB 128 or so. Try it yourself. A color filter tells you bright Blue numbers are changing, but the image shows little or no action above the high midtones.

Back to working with ColorYUV, etc., for now. It's improving. But it still reeks.

Dumb question, but are you applying the filters before color tools ?

Also isn't this supposed to be on the darker side (lowly lit performance) , so isn't most of the data supposed to be in the shadows? The stage lighting will register as a small peak on the high end, and I think that's all the bright data there is supposed to be ?

Now are you talking about the distribution in the blue channel (the "shape" in the histogram) ? The amplitude of the blue channel is high (a large peak) but occurs at low values of the histogram (seems like the whole channel is squished to a peak and shifted left as you said earlier). So most of the blue data is distributed in the dark shadows (and half of it is clipped if rec601 is used). The color tools in vdub shows the whole histogram, so you see the entire top of the peak, and the rest of the data appears diminished - this skews the visual representation of the other channels (you can see scale factor at the top).

Did you try swapping out the "bad" blue channel ? The scale (if you use vdub color tools) will appear to be more balanced, because the amplitude of the blue peak and the distribution will be in line to those of the other channels. In the example below, if you use the color picker you will notice that the dress and clothes are closer to white now (assuming that's what they were supposed to be - and if it's not - maybe there was colored stage lighting - you can add you own color cast with other filters if you know what the stage lighting was supposed to be) . Moreover, this is almost automatic (you don't have to change scene-by-scene or keyframe changes) , because it swaps the RGB blue channel from luminace values of the corresponding YUV clip . Now it still looks green, just a different shade of pukey green - but the point was to "fix" the blue channel before you start with your final color corrections.

Curves (there's one called "gradation curves" in vdub) is the most powerful tool IMO. Non linear and almost infinite control. You are not just limited to 3 way shadows, midtones, and highlights. You have a lot more control in redistributing data in each channel than say, levels which only allows for linear control.

[sanlyn]

Dumb question, but are you applying the filters before color tools ?

No filters, just run ChubbyRain2, tweak with ColorYUV, then to RGB32 AVI. We are getting similar histograms, but I'm not having to move Blue as much as you seem to be. I think the lighting in this show is warmish anyway at the high end (theater lights usually run around 5500K, unless gelled). The cap you posted of the ending scene is too cyan. But I think those folks are under amber gels; the color balance has more high blue after the curtain closes and the lights come up for applause.

To answer the next question in your post: the last 1/3 of the show is fairly somber. It's preceded by a violent fire scene at night (!!).

Later I'll use my usual filters: gradation curves, RGB-EQ, ColorMill. I found a little trick with ColorYUV. Tried a contrast tweak (cont_u=-25 to -50) -- it pushes blue out a little. Going below -50 didn't help much. "Cont" spreads from the middle outward. Now near the point where really heavy post-filtering isn't needed, just tweaking humps and trailing ends. With poor mapping, the colors won't get that clean, so flesh tones are really trying.

Swapping channels gave worse results, even tho I spent most of today with it. The blue peak around RGB 64 and the sharp upper-end slope is due to poor mapping (and protection effects? ?), but swapping seemed to replace problems with new ones. Best to concentrate on one color; working with two is tough, three is impossible. ColorYUV is getting close, with fiddling. It's more helpful in later scenes. I have a pretty good handle on Red and Green; the only thing I'll likely have to struggle with is Blue, about once every half-hour of video. It seems to be mapped to some really high luma at the low end (i.e, oversaturated, even at low numeric values. Remember, there's "hue" and then there's "brightness"), and low luma at the high end. I still think that's deliberate; the other colors aren't acting that way.

The trick at the end is to compare the curtain call scene with the earlier one, and you'll see they use tricky lighting that throws you off. Get lots of blue in the death scene, then the curtain opens and everything's blue as hell. Also check fades to black, fade to white, white titles on black screens, etc. If they look clean, the rest should fall into place. That fade to white is a blessing. You don't see many of those!

Gotta go repair a PC now. Will try to get some reasonably short YV12 clips direct from the VOB and post them on 4shared for those who want to torture themselves. Believe me, the first 1/3 of this video will look easy.

[2Bdecided]

Anyway, however, according to GSpot data and TMPGEnc Editor, the VOBs are NTSC.

DVDs aren't really PAL, or NTSC, or SECAM, or whatever.

They're 25fps 576-line (most people call these "PAL") or ~30fps 480-line (most people call these "NTSC").

If you correctly digitise any of the following formats without framerate conversion, you will get a ~30fps 480-line ("NTSC") DVD:
M/NTSC
M/PAL
4.43 NTSC
4.43 M/PAL 60

If you correctly digitise any of the following formats without framerate conversion, you will get a 25fps 576-line ("PAL") DVD:
PAL (all variants except M/PAL and 4.43 M/PAL 60)
SECAM (all variants)

Hope this helps.

Cheers,
David.

[2Bdecided]

btw, I think it's quite dangerous to assume the R, G, and B curves should match - that depends entirely on the content. They'll only "match" perfectly for image content without any colour.

While U and V are normally centred, that again isn't necessarily the case - it depends entirely on the content. It's the presence of white or black (or grey!) somewhere in the image which creates a peak or spike at the middle of U and V - but if there's no white or black in the image, this won't happen.

Cheers,
David.

[sanlyn]

btw, I think it's quite dangerous to assume the R, G, and B curves should match - that depends entirely on the content. They'll only "match" perfectly for image content without any colour.

While U and V are normally centred, that again isn't necessarily the case - it depends entirely on the content. It's the presence of white or black (or grey!) somewhere in the image which creates a peak or spike at the middle of U and V - but if there's no white or black in the image, this won't happen.

Cheers,
David.

Good points, 2B. I hope readers understand that RGB curves will never be exact matches. We're showing scenes with a fairly uniform range of brightness and colors, or scenes showing obvious imbalance where the curve/colors are grossly mismatched (like the fade-to-white scene). Curves will look pretty well-matched on plain white, gray, or black screens.

For those interested, tips on color balance were mentioned in several earlier posts.

[sanlyn]

For those who feel compelled to suffer for their art, here are links to 4 cuts from the original VOBs. Each is cut-and-copy in TMPGenc Editor, no "processing" except outputting video as .m2v and audio as .wav (PCM). Each ZIP has .m2v and .wav together.

Fin_PartA.zip (30.4 MB), first part of earlier "Fin.mpg", no fade-in.
http://dc179.4shared.com/download/J-zFP7n1/Fin_PartA.zip

Fin_PartB.zip (13.8 MB), last part of earlier "Fin.mpg", no fade-out.
http://dc179.4shared.com/download/9QFIcPEl/Fin_PartB.zip

Act2_End.zip (26.3 MB), not posted earlier. Night! Smoke! Fire! VHS noise!
http://dc179.4shared.com/download/rmcEVLTw/Act2_End.zip

Opera_C.zip (20.8 MB), same as OperaC_original.mpg, but .m2v with .wav audio.
http://dc179.4shared.com/download/zJdGig-C/Opera_C.zip

I use AViSYnth/DGindex to convert to huffyuv AVI. Attached is a homogenized copy of the .avs script. It cleans top and right borders and does some UV channel shifting. You need the Avisynth plugins called and DGIndex/DGDecode. poisondeathray's ChubbyRain2 included + comments.

[jairovital]

Good job, sanlyn. I could duplicate your/poisondeathray script here. I think it solves the borders and UV shifting.
I know you are in the middle of the work and have other fixes ahead, but could you point me some NV settings (or noise profiles) to help me test some parts of the show? Mainly that ones to remove the *mosquitos* noises and unwanted movements at the curtain and backgrounds? NV or VDub built-in temporal denoises?

thx

[sanlyn]

At the risk of being a real trouble maker:

Most a/v hobbyists should know (I hope) that correcting video in b&w or color requires some degree of monitor adjustment to certain standards. No self-respecting photog works with images on uncalibrated monitors, but video hobbyists like to poo-pah such things. "Heck, my LCD's digital! Any bozo knows digital is perfect!" Wish it were so, Sparky.

The typical PC monitor or TV out of the box at factory settings is darn near non-functional for accurate video, even if your prize is that $800 Apple or $6000 Pioneer Elite. One of the worst offenders is Samsung; they look like crap out of the box. You can usually eke out a decent calibration even from most budget monitors, but you need the right tools (i.e, you have to spend $$$). You can Google and find free monitor test patterns. Better than nothing. But you need a trained eye for good results and you're hampered with seriously limited user controls on most monitors.

You can spend $5K on tools, but most mere earthlings use something like these, sold at various prices (try B&H Photo or J&R):

EyeOne Display 2
http://www.amazon.com/X-Rite-EODIS2-Eye-One-Display-2/dp/B000JLO31M/ref=sr_1_5?ie=UTF8...2055811&sr=8-5

EyeOne Display 2 LT (actually the same hardware, but less extensive software)
http://www.amazon.com/X-Rite-EODLT-Eye-One-Display-LT/dp/B000CR78CE/ref=sr_1_3?ie=UTF8...2055811&sr=8-3

Here's a test site telling you why and what:
http://www.tftcentral.co.uk/reviews/eye_one_display2.htm
Wait til you see this site's report on that expensive Dell's atrocious default settings!

There's also the cheaper Spyder2 - not that accurate, but very popular and much better than yuckie test patterns. Don't forget, use these tools with Windows/Mac drivers for your specific display. The default generic drivers that load with Windows/Mac are too limited.

[sanlyn]

Good job, sanlyn. I could duplicate your/poisondeathray script here. I think it solves the borders and UV shifting.
I know you are in the middle of the work and have other fixes ahead, but could you point me some NV settings (or noise profiles) to help me test some parts of the show? Mainly that ones to remove the *mosquitos* noises and unwanted movements at the curtain and backgrounds? NV or VDub built-in temporal denoises?

thx

So far, I know of no Vdub/Avisynth denoisers that can handle the kind of junk that NV does. I'm now in the process of making new NV test patterns for the new color corrections. Biggest problem: the colors as well as the noise character and levels change about every 30 minutes. So there's no one-size fits-all solution. That scene at the end of Act 2 is one example. Even NV will see the smoke as "noise", so you end up with no detail at all. That scene is going to take some heavy experimentation, and we'll have to tolerate a certain amount of leftover noise to preserve what little detail we have to begin with.

I'm also using several versions of the script's ColorYUV settings. Remember, the color isn't uniform from start to finish.

Rest assured, I'll pass these details along as soon as I get them tuned myself. In fact I started again from Step 1 last weekend. That's how extensive the changes are since the first tests!

[poisondeathray]

Hey guys, to clarify, I'm not the author of chubbyrain2 - I have very limited coding experience! I just use the filters

The author is "mug funky" for the original chubbyrain and ncahammer for chubbyrain2; here is the original post with script
http://forum.doom9.org/showthread.php?t=143438

And there are avisynth filters that handle denoising similar to Neat Video. I think where Neat video excels is heavy noise. IMO NV tends to erode details too much, even with careful sample and profile creation. Just like NV, to get the most out of avisynth denoising filters, considerable tweaking is required. You will find an avisynth filter specific for about any kind of noise.

[sanlyn]

If you correctly digitise any of the following formats without framerate conversion, you will get a 25fps 576-line ("PAL") DVD:
PAL (all variants except M/PAL and 4.43 M/PAL 60)
SECAM (all variants)

Hope this helps.

Cheers,
David.

Only Info I have is: GSpot sez the VOB source is MPEG-2/CVD (not DVD), 29.97 fps. TMPGenc says it's MPEG-2/29.97 fps in YV12. The imdb site has no record of a publicly available video of this live performance, which they say was broadcast in 1996 (over SECAM TV, naturally). An audio CD by the same cast and crew exists, but no video of this specific showing. The VHS was maybe a private edition through some sort of "society" membership.

I'm watching it on tv now (3rd time) and still see black levels cut out above RGB 32. Symptoms of a player using zero-black based at IRE-15, not IRE-0? On the PC You have to do two level-down conversions to get an RGB 16 "black". Or maybe it's <gasp> YDbDr!

I found this opera must be on multiple VHS cassettes, not one. The first cassette changes at the end of Act I, when the recorder seems paused and a new cassette is inserted with an abrupt, quantum leap in noise. Act 2 opens after a really long lead-in of black (er, green) and has the same Red stain but a new blue stain at left. If anybody wants to download 72-MB of this sequence, I can maybe mount it later. Maybe. Getting tired of finding problems with this idiotic transfer. Hope the owner didn't record and then throw the tape away. I need to find that tape, hopefully before I find and kill whoever recorded it (the actual sequence doesn't matter. Either will dp).

Info I can find on SECAM VHS sez many players in Brazil are SECAM-enabled. Many can play NTSC or PAL/M-PAL, etc., and convert or don't convert. But they play SECAM only as SECAM. I find no evidence of SECAM players outputting M-PAL. To watch SECAM tape properly you need a SECAM display. To record SECAM into anything, I don't know what the hell you need. Nobody ever heard of SECAM home recorders.

[sanlyn]

Hey guys, to clarify, I'm not the author of chubbyrain2 - I have very limited coding experience! I just use the filters

The author is "mug funky" for the original chubbyrain and ncahammer for chubbyrain2; here is the original post with script
http://forum.doom9.org/showthread.php?t=143438

Indeed, I found that site a while back myself. I'll clean up the script a bit, then, poison, and give the originators more notice. But you did recognize the problem, tweaked the original code, tried it to ensure it was worth the effort. So, we'd like to give credit for that.

And there are avisynth filters that handle denoising similar to Neat Video. I think where Neat video excels is heavy noise. IMO NV tends to erode details too much, even with careful sample and profile creation. Just like NV, to get the most out of avisynth denoising filters, considerable tweaking is required. You will find an avisynth filter specific for about any kind of noise.

Um, well...Heavy noise, okay. Biggest prob with AviSynth plugins: horrible to non-existent documentation (some exceptions), dead links everywhere to filters, comments/instructions/examples in Russian, Polish, French, Chinese, no translations, often just a .dll hanging in midair. Obviously, some of them work (your example certainly did).

I wouldn't dare use NV on every minute of this opera, at least not the same one used on the real grunge. I could be running it for months. And if you saw that smoke and fire scene, NV would turn that to mush, can't use it there. Yes, tweaking required, with NV and with everything else.

[manono]

And there are avisynth filters that handle denoising similar to Neat Video. I think where Neat video excels is heavy noise. IMO NV tends to erode details too much, even with careful sample and profile creation. Just like NV, to get the most out of avisynth denoising filters, considerable tweaking is required. You will find an avisynth filter specific for about any kind of noise.

Yes, I agree. I was going to contradict him myself but you beat me to it. I've never used it myself (and have no intention to do so), but I've seen plenty of videos that have come out of Neat Video, and they've all been effectively ruined by all the detail being wiped out. Some of AviSynth's denoisers not only can handle the kind of junk that NV does but can do it better, without nearly as much detail loss. I've been doing a lot of work recently cleaning up lousy VCDs (themselves sourced from lousy VHS tapes) for conversion to DVD and have become very fond of the FFT3DFilter, but there are others. And isn't Neat Video quite slow in doing its work?

[manono]

Biggest prob with AviSynth plugins: horrible to non-existent documentation (some exceptions), dead links everywhere to filters, comments/instructions/examples in Russian, Polish, French, Chinese, no translations, often just a .dll hanging in midair.

Now you sound like lordsmurf. While it may not come easy, and while you might have to do some work to get optimum results, the payoff is worth it. And the one I mentioned in my previous post is very well documented with examples and all:

http://avisynth.org.ru/fft3dfilter/fft3dfilter.html

Download it at the very bottom of the page. And there are others. A lot of people like DFTTest, others MCTemporalDenoise:

http://avisynth.org/mediawiki/External_filters#Denoisers

[poisondeathray]

And isn't Neat Video quite slow in doing its work?

It is, but when you get those complicated stacks for avisynth, they become very slow too , even slower than neat

But no question neat video is easier to use, it has a nice gui and you sample noise by selecting an area.

I agree with sanlyn that the organization and trying to find obscure dll's in avisynth could be handled /organized better. No arguments there. But it's well worth the trip

But some of the real avisynth gurus have done amazing stuff, that neat video can't even touch. Neat video to me is like a tank - even when you do very fine custom profile and noise sampling - it tends to overblur even on low settings . But for very heavy noise, it seems to be better than most avisynth filters , even stacked chains

[sanlyn]

Poison, AviSynth on many an occasion has one big advantage over NV or VirtualDub, for which there's no substitute: working in YV12 and YUY2.

Game time! Below are caps of frame 13591. <- "A" at left, "B" at right ->. One is from the original after going thru ChubbyRain and into AVI/RGB32, but before NV. The other was run today with NeatVideo. Which is the original, A or B? This is the scene with bad motion flicker in the trees from "Opera_C", but these caps are from today and the new color work Poison and I, et al, are driving ourselves nuts with. This is some of the worst "detail" I've seen in VHS in a spell, but whaddya gonna do?

Clues. Look at how soft-edge areas blend into areas of a different color. VHS and low bitrates usually have noisy stuff there. Try the gray beard and the edges blending into skin tones: any noise in that blend? How are shadows defined on the left of his face, are they losing shape, blocking up? More clues: soft blobs of darker green on tree trunks, dark blobs and subtler blends; do they disappear, or is some of that more subtle stuff just "noise"? Even better: look at the two women, lower left (Ah, Mais oui!). They're in lower-density light, not the best for digital video. NV is a decent hand at removing compression artifacts ("wiggles") around edges. Look at the front girl's face and ts edges blending darker or similar colors; which has more distinct edges and transitions? There's some squiggles between the outer edges of her hat an the figure beside her and scenery behind. top of her into the darker shadows under the hat -- Which pic has more noise there? Which appears to have less definition in the softer clothing on the figure behind her?



Not fair, those JPGs are only 93% PQ. Attached is a ZIP with 3 Photoshop .psd images. They are the original A and B caps. You need Photoshop-enabled software because they have their original layers and layer names. The layer names identify "13591 orig" or "13591 NV". The 3rd .psd in the attachment is titled "F_13591_Duo". It's the same two images with the same layers, but one layer is "in front". Turn the front layer's "visible" switch on and off quickly and make the lower layer visible. Switch quickly and compare layers. One of the layers has more perceived detail, cleaner colors, looks "sharper", seems less clouded and has more depth.

I'm thinking, maybe STMedian can help with that fire-and-smoke scene. That's one scene NeatVideo can't help.

[sanlyn]

[I've never used it myself (and have no intention to do so), but I've seen plenty of videos that have come out of Neat Video, and they've all been effectively ruined by all the detail being wiped out.

Since you've never used it, don't understand how it works or doesn;t work, never diagnosed its misuses, don't know anything about it, don't know how others use or misuse it, don't know that pros have used it for years but don't know how they got better results, don't know and don't wanna know, and ..... wait, repeat that??

The original unfiltered versions and early NeatVideo versions of 3 videos were posted earlier. They're free downloads, no wait. Take a look. We're working with a really crummy transfer and looking for your ideas. Some have offered theirs and we've learned much and used some of it. I used AviSynth for 7 years and was disappointed, but I still have a ton of their plugins, have been trying them again for several days, used Poisondeathray's and others' excellent AviSynth suggestions and tips, and I'll try more. I'm not averse to stuff that works.

Download OperaC_original.mpg. It's never been processed, it's straight off a VOB. Show us how to stop the ugly flashing on camera pans, we'll try it. I used NV and it cleaned it up. I don't see any detail loss (what little detail there is, which ain't much). fft3D couldn't even see noise and 3D convolution just made the old man's nose disappear. Tell me how I've erred. and show us your experiments.

Just behind this post is another one with two images posted. So we're having a little contest (well, we have to play sometime!). One of those images is from unprocessed AVI, the other is the same frame after NeatVideo finished removing the flashing noise. It should be easy for you to tell mus which image is one that's ruined.

[manono]

We're working with a really crummy transfer and looking for your ideas.

You want ideas? Delete it off of your hard drive as quickly as you can. I've seen a lot of garbage like this and that's what I do. It's not worth beating your head against a wall. You've been working with this thing for - what - nearly a month and a half now, and only God knows how many hours? It still looks like crap and it will always look like crap. Maybe some day it'll look like properly colored crap. One can't talk about preserving detail when there's none there to preserve. When I was talking about AviSynth filters doing a better job I was, of course, assuming there was something with which to work. That's not the case here. I initially walked in to help rebut your know-nothing attitude towards AviSynth and its filters. Now I'll walk back out again and leave you to your misery. There are plenty of projects which interest me and to which I contribute as best I can. This doesn't happen to be one of them.

[jairovital]

Hey guys, take it easy...

We are enjoying to change experiences and opinions about a specific subject. I think we are enjoying it. Nobody here is waisting time or hitting head against a wall. We all doing that for fun. The good news is that fun brings knowledge and more experience.

Some people like apples, other loves pineapples. Some prefers summer, other goes to spring. Here is the same. A big team uses Avisynth scripts and another one get a ticket for VDub filters. Is not a good practice to judge who is the best or who is the St. Guru. I think the title of this thread clarify that enough: Scripts and Filters for an Opera.

Crap videos is the best place to work on. Who can achieves the best result? How did he could that?
Better than work on a HD video where there is nothing to improve. Crap videos has millions points to looks better.

Lets go on with our toys, exchanging ideas and tryouts. Everybody wins in this way.

[jairovital]

sanlyn

Side-by-side is hard to see the differences between A and B. Photoshop layers shows better and we can see real improvement NV could do.

I'm not sure if is the point, but maybe we could try to apply some FFT3DFillter, DFTTest and MCTemporalDenoise to compare too. All that tools are valid when they bring and show good job.

[sanlyn]

sanlyn

Side-by-side is hard to see the differences between A and B. Photoshop layers shows better and we can see real improvement NV could do.

I'm not sure if is the point, but maybe we could try to apply some FFT3DFillter, DFTTest and MCTemporalDenoise to compare too. All that tools are valid when they bring and show good job.

I tried the first named filter on several scenes already (I know that one from 'way back) and will be trying more, later. I have a PC repair job 40 miles away this AM, so I guess the opera gets a rest today.

[2Bdecided]

Neat Video's spatial denoising is useful if used correctly - especially for "patterned" noise which it can learn and remove. In AVIsynth, fft3dfilt is closest to that, and I struggle to make it work well for those cases where Neat Video is at its best.

BUT Neat Video's temporal denoising is junk - AVIsynth has far better motion compensated temporal denoisers available. Where there's no pattern to the noise, this is often a better way to go.

Then there are specific AVIsynth tools for deblocking etc - something Neat Video is poorly suited to.


Now, I agree that Neat Video's default settings are beyond overkill. I also agree that there's so little detail in this source that denoising may not make it look better. If anything, I'd consider adding subtle fine grained noise at the end.


As for AVIsynth being difficult - if you start with the mediawiki, then read about advanced tools on doom9, then simply google for any missing plug-ins, it's complex but fine. I've never come across (never mind needed) pages in Chinese or any other non-English language.


I'll tell you what you do need (and I can't remember the name for it) : the tool that replaces bad frames. On several clips you've shared, there are moments (sometimes rather long moments!) where the entire picture is useless. I think there's an AVIsynth filter that lets you specify a frame or range of frames, which then replaces these with new frames made up from the surrounding frames. It's not perfect, but it's better than snow/static.

Cheers,
David.

[sanlyn]

Thanks, 2B. Your input would be helpful. I'll try it. Will look up some of that myself when I get back today.

Sometimes I just let dead frames pass and didn't hear of what you mention, but it sounds like it's worth a try.

[manono]

I'll tell you what you do need (and I can't remember the name for it) : the tool that replaces bad frames

It's called, surprisingly enough, BadFrames. You specify a frame or series of frames and it uses good frames on either side to morph/blend from the last good frame to the next. As 2Bd says, it's not perfect, but it's way better than just noise or blank frames. I use it quite a bit:


http://avisynth.org/warpenterprises/files/badframes_20060813.zip

And sanlyn, I was sure you were going to come back and hammer me, but yours was a reasoned response. I apologize for the harshness of my tone.

[sanlyn]

I'll tell you what you do need (and I can't remember the name for it) : the tool that replaces bad frames

It's called, surprisingly enough, BadFrames. You specify a frame or series of frames and it uses good frames on either side to morph/blend from the last good frame to the next. As 2Bd says, it's not perfect, but it's way better than just noise or blank frames. I use it quite a bit:


http://avisynth.org/warpenterprises/files/badframes_20060813.zip

And sanlyn, I was sure you were going to come back and hammer me, but yours was a reasoned response. I apologize for the harshness of my tone.

Fully understood, manono, I mighta started first by having what gals refer to as a bad hair day. Thanks for the link, I was looking for such on my way to bedtime (9 hours of repairing PC's wears...me...out!) .

[sanlyn]

It might just be me, but looks to me as if a second casette was loaded at the end of Act One (1hr 5 min into the opera). Looks like tape #1 started feeding trailer tape, recorder was paused, new tape loaded and recording resumed, new tape has noisy lead-in and the noise and color look worse. The links below show it all!

Link to 4shared's download page (the crummy little viewer probably won't show because audio is AC3):
http://www.4shared.com/video/s2NAwoyG/VOB15_NewTape.html

Direct link to the mpg (17.7 MB) right outta the VOB, no processing. Left-click, it loads into your PC's media player (takes a while, though). Right-click to download and see the damage.
http://dc244.4shared.com/download/s2NAwoyG/VOB15_NewTape.mpg

[jairovital]

I could try some scripts and settings our master sanlyn sent to me and the result is here:

http://www.4shared.com/file/rRsAuLLp/stackH_huff.html (it is only a 9.5 Mb file)

I have to admit that I am very happy with this. Not only with still frame, but mainly in movement we can see a very nice improvement of the video quality. It was well worth the effort and we do have a real better look, even with the poor quality of the source.

Ed.: Keeping all the fine details, destroying nothing!

Thank you sanlyn. Very good job!

[sanlyn]

Nice goin', jairo. I managed to finish the first 10 minutes. I appreciate the accolades, and don't forget others who contributed, but at this rate I'm far from being a "master". 3 PC repairs in 2 days slowed me down. But it's paying for all these video toys, so I'm not complaining (well, I am complaining, but...). Using some Avisynth plugins, too. Got a better look at that 2nd cassette recording, after Act One. oh, brother! Now the headaches really start.