Deinterlacing in Premiere
 

Anyone knows how advanced is the de-interlacing at export in Premiere?

I use it all the time and find it good, but it just hit me: what if Premiere just does something basic, like blending two adjacent lines.

I know there are advanced filters for VirtualDub, that do motion detection and stuff, so I'm wondering: how much do I lose by being lazy and just check-marking "De-interlace" at export?

Thankful for any advice or thoughts.

Nikolaj

I'd be very interested in this as well. Best I can figure is that PP deinterlaces with blue smoke and black magic. Actually, I think it just throws away one field like Photoshop does. Someone might know a better way.

The Premiere deinterlace is not a "smart" deinterlace of the kind that the Vdub and Avisynth plugins can achieve. If you want to actually quantify how big a difference there is with the various methods, then I would recommend filming a resolution chart in interlaced mode, then examine frames captured from the footage after it has been deinterlaced by the methods you are considering - its really quite a quick test to do.

There are various resolution charts online ... easily found via Google Images.

Thank you both. Tripp, I don't think Premiere throws one field away. I base this on some observations I did comparing the image sent to external monitor (which is one field only in my setup) with the one deinterlaced/exported. Field blending is my guess. But I will conduct some experiments and return with my observations.

I have used de-interlacers in VirtualDub that are said to be the best free ones out there.
But they require tweaking and sometimes they give strange results.

I don't think it can use both fields in deinterlacing because of the temporal displacement between the two. I would think that using both fields would create artifacts commonly called jaggies or mice teeth. But I could be talking through my hat.

I'm also curious about "smart deinterlacing" versus the alternative. "Not smart"?

A better term is perhaps "motion compensated" deinterlacing, with motion vectors being calculated from more than just the immediate two fields. There's a short paragraph about it in the "Deinterlacing" entry on Wikipedia.

Quote:
"I don't think it can use both fields in deinterlacing because of the temporal displacement between the two. I would think that using both fields would create artifacts commonly called jaggies or mice teeth. But I could be talking through my hat."

I'm obviously no authority on deinterlacing, an activity I hope to abandon altogether in a year when stepping up to Panasonics P2 camera... but anyway: Blending two fields will replace the jaggies with a more homogeneous looking "motion blur" area, which is more representative of what one would expect if filming progressively with double the exposure time compared with filming in interlaced mode.

Did I express that well? Hmmm...

I understand what you're saying. But if a motion blur is applied to a jaggy don't you just get a fuzzy jaggy? On fast motion, it might mitigate the jaggy problem but they'll still be visible.

I'm probably poking about in the dark with a barge pole.

An easier way to figure out what Premiere is doing, would be to make an image in photoshop, and paint in horizantal lines, maybe black and white in one area, red green in another, and maybe a small checkerboard with 1x1 and 2x2 pixel patterns. Don't fill the whole image, just enough to have a consistent pattern, like 10 lines. Then import that still into PPro and interpret as interlaced. Export and analyse the result in Photoshop. This has picqued my curiousity and I may test it myself, if I have some time tomorrow.

Yes, and if I'm not mistaken that's simply called a "blend". Looks better than field doubling, but still nothing to write home about.

There are far more complex techniques, for example Xin Li's New Edge-Directed Interpolation, a technique that I don't believe is offered in the base package of any video editing or compositing application (Premiere, Vegas, After Effects, Combustion, Shake, etcetera) unless you start buying plugins.

If you explore the realm of open source software you can find freely available examples of the algorithms in question, one of which I've in fact just offered up to Nikolaj himself in this thread, over on the PC editing board. The thread focuses on ultimately upscaling SD footage, but part of the process involves a version of NEDI as part of the deinterlacing process, implemented as an Avisynth filter. Motion compensation, while not always used in conjunction with deinterlacing, is also part of the steps I've outlined there, and has been ably explained (or so I assume; I can barely handle algebra) in The Engineer's Guide to Motion Compensation by John Watkinson of Snell & Wilcox. More whitepapers and guides can be found in the company's Knowledge Center.

As for figuring out Premiere's method, until Adobe starts publishing technical details this intricate, Mike's suggestion sounds like the way to go.

Field blending
 

Tripp and Robert, I have to point out, just for the record, that field blending DOES make the jaggies disappear. It's not motion blur applied on jaggies.

Field blending is just blending two fields together and then using the result as a replacement for both fields. You get half the vertical resolution, but use info from both fields, not just one of them as in field doubling.

Correct me if I'm wrong.

Hey Robert... Still have to apply all the steps, it's quite complicated with my source footage being spread out on several disks...

By the way: field blend is the most stable de-interlacing method from my experience.
Even though in theory it sounds good to do various motion-detection stuff, in practice, if you're not prepared to tweak parameters in VDubs "Smart" filter (to take an example) for hours... field blending never gives unexpected artefacts and doesn't require tweaking...

Mikes suggestion is acually great. I might try it tonight.

No need to correct you, I just wasn't thinking; you're right, I think that is actually how field blending works, my mistake.

As for most stable, I think you're in for a pleasant surprise. The motion compensated NEDI used in the TGMC script is extremely stable, even with only the default settings, despite being a bobber. It has problems with shallow diagonal lines, like most deinterlacing does, but does an excellent job of avoiding the jumpiness that other bob software usually produces. Donald Graft's "Smart Deinterlacer" for VirtualDub is good, but I've never gotten results out of that as easily as I have TGMC, or that were as good looking. A blend is much faster, of course, so it's give and take between all these methods.

Well that's very interesting; I'll definitely try the NEDI.

Where are the deinterlace options in CS4? I usually use CS3 at work, but only have CS4 on my system at home. Unless I find those options on CS4, I will put off the test until I have a CS3 system available next week.

OK Robert, now I'm confused. After reading some of "New Edge-Directed Interpolation" I can only conclude that this is a technique for image interpolation. That's great for what I was asking for in another thread (upscaling from SD to HD, for example) but it has not got to do directly with deinterlacing.

BTW, you don't happen to know some free implementation of Xin Li's method?

Regards,
Nikolaj

I understand that sometimes one may want to turn 60i to 60p, in sports for example, and then one has to use bob filters and interpolate from one field and the like. But for deinterlacing from 50i to 25p, as in my case, image interpolation is not relevant. Not directly, at least.

I suppose I could have done a better job of explaining this, let me try it a different way.

The "NNEDI" I mentioned as part of the enlarging process in that other thread (Neural New Edge-Directed Interpolation, I believe) is, in fact, a deinterlacer. It uses NEDI techniques as part of what it does; throwing away half of the fields produces frames with missing information, that are then interpolated (via NEDI) back to full height. The theory itself may not seem to apply directly to deinterlacing, but when applied in this fashion proves quite able.

This software, mind you, is NOT applied by itself in the TempGaussMC script. One of the fundamental aspects of Avisynth is the scripting language. Scripts can get very complicated, involving the creation of your own functions, and that's exactly what's happened in TGMC: the output of the NNEDI plugin is being manipulated to produce better results. The NNEDI plugin is useful in its own right when run as the only filter on an interlaced input file, but in this case is only one of many, many things being done by TempGaussMC.

Follow me back to the other thread and I can show you how to use that script just for deinterlacing footage.

Robert, let my quote you:

"Calling the deinterlacer is as simple as you see here, with the one change that we use NNEDI instead of the default EEDI2. I get the best results with that, though if you want to try the other options feel free.

TGMC is a bob deinterlacer, however, which means it turns fields into frames. You'll have 50 frames per second after it's finished, and since your project is 25p, the SelectEven() filter will grab every other frame and give you that. There's also a SelectOdd(), but since frames in Avisynth are numbered starting from zero, SelectEven will give you every other frame starting from the very first one."

Now, I'm really glad that you shared your workflow for upscaling with me. But: after reading more closely, it hit me that, according to the above, the method you describe is basically throwing one field away instead of using it for information. Now, what is my problem with that?
Well, my argument is this: except for the purpose of doubling the frame rate, one should always try to combine the two fields. Why? Because with 25 images per second (in my case) motion may become more "stroboscopic" if just using one field due to the shorter shutter time. Combining the two fields (which blending does) gives some "motion blur" automatically. This removes the "stroboscopic" effect.

I realize that tastes may vary. But the idea is that ideally, if a movie/film/moving image
is presented at n images per second, then the shutter speed for every image at capture should be 1/n of a second, so as to give a full temporal representation of the motif.typo

For what it's worth, the SelectEven() filter is called after deinterlacing is complete. It throws away frames that are the result of interpolation, and has no effect on the calculation of the progressive frames in the first place. You can see exactly how all of this is done by opening TempGaussMC_beta1.avs in Notepad and reading the various functions. From the sound of it you'd get more out of the technical info laid out there than I ever could.

You may be correct; I'm really not sure TGMC does anything at all about motion blur (though I could very easily be wrong). You seem to know more about this than I do, I just wanted to share my particular workflow. Keep in mind, TempGaussMC is only something I use to prepare clips for resizing. The Lanczos4Resize is what does all the work, and is still useful even if you don't like the way I do things.

To resize without deinterlacing, try deleting everything in the script I gave you after the AVISource line and pasting in this instead:

AssumeBFF()
SeparateFields()
Lanczos4Resize(1440,540,0,36,720,216)
Weave()

You even get to avoid the YV12 conversion. This script will separate the fields, resize them by themselves, and then re-interlace them so that you can deinterlace the result yourself using whatever method you prefer.

Or, of course, you can still deinterlace before the scale, using any number of third party options: http://avisynth.org/mediawiki/Extern...#Deinterlacing

Wow, the number of options is certainly overwhelming! AviSynth seems to be really fun, I'll definitely explore it and its plugins more from now on. Thanks Robert for all your input, I have more than enough now to play and tweak with until I find a workflow that suits me.

And then again, it is really tempting to just save some money and invest in a Panasonic P2 camera with 1080/25p recording mode, combined with Prospect HD for post... but where's the fun without all the tweaks?

You're very welcome! Avisynth is a lot of fun to play with, and I think you're right:

If it weren't for all the deinterlacing, color conversions and scaling, I'd be bored out of my mind. Footage that comes out of a camera the way you want it? Where's the fun in that?

Reality wins over perceived logic every time. I appreciate this discussion. There's lots to learn here.

So after some simple tests, I have determined that Premiere CS3's deinterlace export option throws away the second field, and fills in the missing data by blending the lines above and below at 50% each. Be advised that using this method throws away 50% of the image data right off the top, before compression.

It seems you are right. I did some tests on my Premiere 2.0 and they confirm your tests.

I have done each clips by right click-field option-always deinterlace...
When I export my videos to h.264 to post on vimeo and smugmug, I have to deinterlace everything (even though I shot everything in 30p) because when the clip has slow mo OR text animation from AE has those annoying lines...
I know there is "progressive" export in some preset for certain video types, but how about quicktime or h.264 output? How are we supposed to deinterlace those if the sequence contained more than 100 different clips???
I don't understand why Adobe take away a one simple option on Media Encoder to make it messy like that.

Does this review of twixtor and fieldskit help

Marco Solorio reviews: RE:Vision Effects, ReelSmart Twixtor with FieldsKit Deinterlacer : Adobe After Effects