Shooting progressive-scan versus de-interlacing (properly) in post

[Jack Kelly]

Hi there,

As we all know, there's lots of hype which goes something like "shooting progressive-scan looks more filmic than shooting interlaced". Yet there are plenty of great cameras out there which only shoot interlaced.

My question is this: how does de-interlaced footage compare to footage shot with a progressive-scan sensor. And by "de-interlaced" I don't mean "throw away one field" or "average and blend the two fields" like the de-interlace function on many NLEs. I mean de-interlacing done using motion-compensation like AviSynth's MVbob algorithm.

I'm a proud owner of a DVX100a and I love the look of the 25p video that it shoots. But I'm wondering if I should actually start shooting interlaced (both on my DVX and on various HD cameras that I hire). I'd edit in 50i. In terms of distribution... I'll leave the video as 50i for PAL DVDs. But for web and distribution I'd output an uncompressed 50i video from Premiere Pro and then use MVbob to convert that 50i to 50p on the basis that 50p will look great on computer screens.

Does that sound at all sensible?!?

Thanks,
Jack

[Jack Kelly]

I've been playing with some DV interlaced footage and PPro 2.0's "always de-interlace" option seems to definitely soften the image. AviSynth with MVbob produces a good image (and doubles the frame rate!) but takes AGES to render and isn't multi-threaded.

[Jack Kelly]

Hi again,

So - more tests with footage (both 50i and 25p) shot with my DVX100a. Some notes:

I tested five different scenarios: a) shooting 25p with the DVX b) shooting 50i with the DVX, c) de-interlacing the 50i using PPro's "always deinterlace" function, d) de-interlace the 50i using Cineform's HDlink, e) de-interlace using MVbob (with all the defaults).

Some observations:

1) Just to confuse you, PPro 2.0 always does some sort of de-interlacing when viewing footage on your computer screen. The only way to see the "true" output is to render out frames.

2) MVbob does indeed produce a 50p output which results in very nice slow mo

3) MVbob reduces the noise a little bit

4) MVbob is very slow!

5) MVbob doesn't like ingesting Cineform files (this might be fixable by using something other than DirectShowSource to load the video - not sure)

6) None of the tested de-interlace routines reduced the resolution by a noticeable amount (not to my eyes, anyway)

7) None of the de-interlacers produced motion-blur identical to shooting "true progressive". That doesn't necessarily mean that de-interlaced footage was aesthetically inferior to "true" progressive captured footage.

So, my conclusions:

1) All the tested de-interlacer routines produce results which are not significantly de-graded compared with shooting progressive scan.

2) The 50i-to-50p conversion that MVbob does is really handy for slow-mo effects. Combined with the 240-field-per-second feature on some new Sony HDV cameras, this could produce some fantastic slow-mo photography.

3) MVbob is probably too slow to use on rushes. But I like the "clean" signal that MVbob produces. I might tinker with using MVbob on finished edits.

4) These tests were knocked off fairly quickly. More testing is needed, especially on different display devices.

[Bryan Worsley]

Hi,

In case you are not aware, there is a new AVISynth bob-deinterlacer on the scene, MCBob, that gives superior results to MVBob. It uses a fully auto-adaptive motion mask that achieves more accurate discrimination of motion and (relatively) static elements. As a result it better preserves detail and is markedly less prone to vertical jitters and shimmering. Inevitably however, it is also slower, but if your prepared to wait, the results are truly impressive.

I posted the deinterlacing script I use and links for the sources over in the Pana3ccduser forum:

http://www.pana3ccduser.com/showpost...3&postcount=50

In practice, I use MVBob as my mainstay deinterlacer and reserve MCBob for more challenging scenes (detail moving over detail etc).

In the post before it I also included information on the method I use for converting the bobbed (50p) output to 25p with a motion-compensated blur.

http://www.pana3ccduser.com/showpost...6&postcount=49

If you find MVBob and MCBob too slow for your needs, there is another deinterlacing function in the MVBob.avs called SecureBob. It preserves a little less detail than MVBob and tends to produce smoother edges, but is much faster.

Even faster is an AVISynth port of the MPlayer deinterlacer - Yadif (literally 'Yet-Another-Deinterlace-Filter'):

http://forum.doom9.org/showthread.ph...ighlight=Yadif

Quality-wise, the results are comparable to SecureBob, but it provides the option of double-rate or same-rate out-put.

Another important consideration for processing DV with AVISynth/VirtualDub is the choice of DV decompressor, as a number of proprietary (vfw) DV codecs are known to be susceptible to chroma and luma upsampling errors, at least when used independantly of the editing/coding programs they were designed for.

http://forum.doom9.org/showthread.ph...erences+Update

http://www.pana3ccduser.com/showpost...55&postcount=3

For these reasons, I use and wholly recommend the open-source Cedocida DV codec (source cited in second link above) which can output RGB, YUY2 and YV12 and upsamples correctly in each. For processing of interlaced DV in YV12, the decoder should be set to 'forced YV12' output with 'MPEG-2 interlaced' sampling.

Hope this is useful.

[Jack Kelly]

Great stuff, thanks loads for the reply.

Is there a way to make AVIsynth multi-threaded to use multiple cores? I guess not.

[Bryan Worsley]

Originally Posted by Jack Kelly

Is there a way to make AVIsynth multi-threaded to use multiple cores? I guess not.

AFAIK, AVISynth (per se) does not yet support multithreading, although this is in development:

http://forum.doom9.org/showthread.ph...60#post1004960

There is an AVISynth filter (MT) that enables (some) other filters to run multi-threaded:

http://forum.doom9.org/showthread.php?t=94996

However, it appears that the motion-vector (MVTools) filter, upon which MVBob and MCBob depend, is not MT-compatible.

http://forum.doom9.org/showthread.ph...901#post958901

[Jack Kelly]

Ah, interesting - thanks for the links.

Are there any free, multi-threaded, high quality de-interlace apps out in the wild? Either Linux or Windows (preferably Windows as I don't have Linux installed on my quad-core box yet ;-)

Thanks,
Jack

[Jack Kelly]

Incidentally, I should explain one of the reasons why I'm interested in high-quality de-interlace and de-bobbing systems...

I'm working on a corporate video at the moment that's destined for display on large LCD screens and projectors, all fed from quite high-spec computers via DVI connections. I'm investigating the possibility of sending my client 1080/50p (50-frames-per-second-progressive-scan) renders. Of course, we'll do some testing first to make sure this is viable. I'm interested in 50p because it looks great - it has the smooth motion of 50i combined with the "solid" feel of 25p.

My scheme for making 1080/50p videos on a relatively low budget goes something like this:

1) Shoot 1080/50i on a Sony Z1
2) Edit in Premiere Pro CS3 as 1080/50i HDV
3) When the edit is finished, create a new 1080/50p Premiere Pro project and import the 1080/50i project. Then replace all the 1080/50i HDV rushes with aviSynth scripts which will convert the 1080/50i HDV rushes to 1080/50p.
4) Render as either H.264 or WMV9 as 1440x1080 50p

I could simplify this workflow a little by rendering out an uncompressed 1080/50i movie from Premiere Pro and then de-interlacing that movie but I'm worried that Premiere Pro will do some nasty things to the fields on the shots that I've slightly re-framed (e.g. some of the shots taken at the wide end of the lens on the Z1 include the shotgun mic's wind muffler in the top right of the frame so I've had to crop into the image a little!)

[Bryan Worsley]

Originally Posted by Jack Kelly

Are there any free, multi-threaded, high quality de-interlace apps out in the wild? Either Linux or Windows (preferably Windows as I don't have Linux installed on my quad-core box yet ;-)

I'm afraid my experience is almost entirely with AVISynth and since my CPU is single core I've never really looked into multi-threading.

As I understand it MT is possible with VirtualDub for which there are a number of deinterlace filters....some now quite old. From my limited testing in the past, I would say that none of them comes close to the quality of MVBob and MCBob. Best of the bunch is probably Smooth Deinterlace, but IMHO the 'faster' smart-bobbers available for AVISynth (SecureBob, TDeint/EEDI2 in particular) will give as good, if not superior, results.

Its also worth bearing in mind that VDub filters only work in RGB color space. Processing in YUV will give you sharper results generally.

Nonetheless, if you are interested, here is a link to a (modestly priced) GUI for 'enhanced' processing with VDub filters that claims superior multi-threading efficiency:

http://forum.doom9.org/showthread.ph...354#post892354

Cant say I've tried it myself.

[Jack Kelly]

Brian, thanks loads for all your help. You've opened my eyes to AviSynth and I'm sure it'll come in handy for a project in the future.

I've spent the day testing various de-interlacing workflows with HDV footage and I have to say that I'm confused.

The main objective of my testing was to compare the results from de-interlacing using PPro 2.0's de-interlacing, yadif and MCbob. The source footage was 1080/50i shot on a Sony Z1e and a Sony HC7e.

First, I shot a "resolution test" using my HC7e. The test consists of an A4 sheet of paper with ten thin lines on it. The lines fan out from a single point. Each line is separated by 1 degree. I shot this test using my HC7e as 1080/50i. The test gives an objective measurement of the resolving ability of the imaging system. There will be a specific point at which the lines blur into each other. The high the resolving power of the system, the nearer to the point of convergence that the lines should be distinguishable.

I imported this test into PPro and compared the results of:

1) interlaced
2) de-interlaced using PPro's internal de-interlace routines
3) de-interlaced using yadif
4) de-interlaced using MCbob

This test does show that PPro's internal de-interlace routine does indeed reduce the vertical resolution, although not by quite as much as half. yadif and MCbob both produce results which look as sharp as the interlaced footage field-blended.

So - to summarise - PPro's de-interlace routine does indeed reduce the vertical res.

However, things get more complicated with "real" footage (e.g. people moving, people static, panning etc). The bottom line is that, to my eye, PPro's de-interlacing doesn't produce results which look noticeably "softer" than MCbob or yadif. Even if I zoom in by 1600%, the differences are still only subtle. And, importantly, using aviSynth with PPro does seem a little unstable and is certainly slower than using native HDV footage.

The lack of improvement of using MCbob versus PPro's internal de-interlace routines confused me a lot so I did a further test. I exported still frames from PPro and took them into photoshop. I then reduced the frame by 50% vertically and then enlarged it again by 200% (i.e. I'm trying to halve the vertical resolution). To my surprise, a comparison of the original image and the "half vertical res" image revealed very little difference. My hunch is that neither then Z1 nor the HC7 resolve enough vertical detail to mean than a halving of vertical res damages the image significantly (i.e. the images aren't particularly sharp to start with so you're not loosing much information by squishing-and-then-unsquishing the image).

My conclusion at this point is that the costs of using aviSynth for de-interlacing 1080/50i HDV outweigh the benefits and so I'm going to use PPro's de-interlace routine to produce a 1080/25p finished video.

As a side note - I discovered something I didn't know about PPro 2.0... if you create a 50p project in premiere and import a 50i video into the project then premiere pro bobs the 50i footage to create a 50p video.

[Bryan Worsley]

Glad you've reached a satifactory conclusion. I'm afraid, until your last post, I was labouring under the misconception that you were processing DV. I have no experience deinterlacing HD other than playing around with a few reference test clips.

Having recently snapped up a Pana NV-GS400 (before they become a collectors item), my personal little dilemma is whether now to shoot interlaced or in 'pseudo-progressive' 25p (FrameMode). Having spent many, many hours deinterlacing DV, the convenience of progressive (even if not true progressive scan) video has considerable appeal. However, I'm still trying to decide whether I'm sufficiently happy with the somewhat fluttery motion cadence of frame-mode 25p to abandon bob-deinterlaced 50p, and the superb fluidity it acheives, despite its imperfections.

Shame Panasonic didnt bring down more of the pro features down to the consumer level, as they once teased. The DVX100 and HVX200 are way beyond my budget as a home hobbyist......I can only dream and drool.

[Jack Kelly]

Originally Posted by Bryan Worsley

I'm afraid, until your last post, I was labouring under the misconception that you were processing DV. I have no experience deinterlacing HD other than playing around with a few reference test clips.

Your advice was extremely helpful, none the less ;-)

[Marcus van Bavel]

You might also look at DVFilm Maker (http://dvfilm.com/maker) which can intelligently deinterlace HD, HDV and DV, as well convert HDV to DV, etc. Together with an intermediate codec like Raylight for example (http://dvfilm.com/raylight), you can process 50i HD m2t stream clips directly into 1440 x 1080 25P Raylight AVI, with a built-in low res proxy for real time editing. The proxy can be switched out for the full res version with the push of one button. It works with Premiere, Vegas, After Effects and many other programs.

See http://dvfilm.com/raylight/raylightTutorial7.htm
for an example of how Raylight and Maker work together.

[Jack Kelly]

Hi Marcus,

Thanks for the links - I'll take a look.

For this project, I think I've found a workflow that I'm happy with:

1) Shoot 1080/50i
2) Edit as 1080/50i
3) Render out the completed, colour-corrected edit as 1080/50i uncompressed
4) Use aviSynth to de-interlace the uncompressed footage (either as 1080/25p, 1080/50p or 720/50p depending on the client's needs)
5) Render to WMV9 using by importing the aviSynth script into PPro and rendering it

Tests with my DIY resolution chart show that this workflow produces no noticable drop in resolution. My previous concern that PPro will screw up clips that have been re-sized appears to be unfounded - PPro seems to deal very nicely with re-sized clips in the final render (although the preview looks horrible).

My only slight concern is that aviSynth has to convert the colour model. I can only get this workflow to work by exporting from PPro as "Microsoft AVI", compressor=none. This produces an RGB24 file which needs to be converted using ConvertToYUY2. I can't detect any noticeable damage to the image though.

Thanks,
Jack

[Marcus van Bavel]

Originally Posted by Jack Kelly

Hi Marcus,

Thanks for the links - I'll take a look.

For this project, I think I've found a workflow that I'm happy with:

1) Shoot 1080/50i
2) Edit as 1080/50i
3) Render out the completed, colour-corrected edit as 1080/50i uncompressed
4) Use aviSynth to de-interlace the uncompressed footage (either as 1080/25p, 1080/50p or 720/50p depending on the client's needs)
5) Render to WMV9 using by importing the aviSynth script into PPro and rendering it

Tests with my DIY resolution chart show that this workflow produces no noticable drop in resolution. My previous concern that PPro will screw up clips that have been re-sized appears to be unfounded - PPro seems to deal very nicely with re-sized clips in the final render (although the preview looks horrible).

My only slight concern is that aviSynth has to convert the colour model. I can only get this workflow to work by exporting from PPro as "Microsoft AVI", compressor=none. This produces an RGB24 file which needs to be converted using ConvertToYUY2. I can't detect any noticeable damage to the image though.

Thanks,
Jack

I recommend that you pre-process the clips to 25P before editing. This can be particularly important if you have any transitions, motion or time effects, or superimposed motion graphics. Even on a cuts-only edit the deinterlacer can be confused by processing edited material.

WMV9 tends to smear out and then undersample smooth areas to reduce bandwidth, I would not use it as an intermediate.

Raylight AVI (DVCPROHD) 1440 x 1080 ( http://dvfilm.com/raylight ) is a nice intermediate codec, it does not filter out detail and has only about .5-1% RMS added noise per generation.