Ideas about deinterlace

Some of my findings:
Deinterlace video mostly exist on “old” DVD sources. After industry moved to better bandwidth/TVs the video became progressive.
DVD interlaced sourced are of different frame rates:
29.97 fps
And nearly all of them have mixed content. I mean some scenes are progressive and some are fully interlaced. By fully interlaced I mean that each field represent a frame.
Mpeg source can contain flags that show where progressive frames and where interlaced scenes but as I’v found in most cases there are no such flags.

Here is the problem of current Topaz Labs deinterlacer. It is trained to deinterlace only content where each field is a different frame (there is a movement between fields).
If there is no movement and fields can be combed (bottom field + top field = full frame without movement) - deinterlacer fails (artifacts, strange movement to another frame).

Like for example if I deinterlace 25 fps movie that has mixed scenes using Robust version. Then progressive scenes are ok, but fully interlaced have weird movement.
If I deinterlace using not robust version, then progressive frames loose details (because different half of same frame are used to reconstruct same frame 2 times resulting in same frame with lost detail). But fully interlaced scenes are deinterlaced very fine (except that they should be only deinterlaced without any other processing - denoise/sharpening etc).

So process of properly deinterlacing such mixed content is quite a hard task.
Here are what needs to be done.

  1. You need to find out which scene is progressive and which is full interlace.
    Progressive scene can be recovered by correctly combing fields.
    Progressive scenes in 29.97 source also need pulldown to remove duplicate fields.
    From content to content which field is duplicated for pullup is different.
  2. Full interlaced scenes need deinterlacer like Topaz Labs one that interpolates (guesses) dropped fields.
    But in result you get 2x fps (better to retain double fps and not to drop half of them)

At last we need to have movie with different frame rate scenes. Like scene 1 with frame rate 24000/1001 (progressive combed scene with pulldown), scene 2 (30000/1001 * 2 - fully deinterlaced scene). Mixed frame rate is fine for mkv files with timecodes.
This will retain the movie as it should be viewed.

And then there are some other complications may be.
Sometimes source contain bad fields (not frames). For example there can be a field that looks like two fields were blended together. Its impossible to do something good with this field. It must be dropped. And it can be 2 or 3 such fields one after another.
In film source such frames can be replaced using frame interpolation (Apollo/Chronos). But in animation Apollo & Chronos will not help much.

As you see its nearly impossible to create automatic solution. Maybe its better to program some user interface where user can get help from models and can tweak what to do on each scene.

And also one idea for deinterlace topic.
There are sources of improperly deinterlaced video where instead of pull down - interlaced frames (because of pull up field duplication) were blended. So video in result contain progressive frames mixed with blended frames.
We need model that drops blended frames and optionally interpolates them with chronos/apollo.
Optionally because chronos/apollo can’t interpolate animation normally and for animation its better to just drop them.

I completely agree. It’s something that happens with old DVDs and Betacam files. This thread is super important for those of us who need to do thorough and professional work.

The issue of deinterlacing is very old - and its implementation inside TVAI has been discussed countelss times before. Every now and then the wish for a better de-interlacer pops up again.

Bottom line: With deinterlacing, issues like telecining, frame rate conversion, etc… are also present. Content derived from a chemical film is different than content straight shot from a TV camera, is different from stuff originally filmed on film, postproccessed on video tape and then got some digital FX added. Norm conversion NTSC/PAL… All these are examples which somehow are stored as fields, ont frames and have visible interlace artefacts - but the “why and how” the interlace lines came to life in the footage are totally different.

Thats why its very difficult to have one universal deinterlacer - in some cases, deinterlacing even is not the proper way to deal with the footage, rather some inverse telecinig is needed.

The open source world has tackled the issue for decades now and some pretty robust solutions for most cases exist - but some still need manual tweaking and chaining different “filters” in order to sort out the interlacing.

duplicate frames are actually not related to interlacing per se, but to telecining or other things going on - which also have the “side effect” of intelace lines - but its two isses at once.

we have not even started to talk about field order etc…

And - after all - interlacing (as in “recorded interlaced”) is not a defect, its just the way the stuff was recorded… If one digs deeper into the issue, one finds that there is no perfect solution to this - its not “reverting a defect to the original state” but altering something thats present in the time domain to the spacial domain with a compromise…

at the moment, very traditional deinteralcing is done inside TVAI - and after that, an regular inference is thrown at the footage… So the deinterlacing by itself has nothing to do with “AI”…

so while interpolating frames might seem like the way to go and in some cases will look fine, in most PAL/NTSC scenarios, they are actually not what should be done. Field-Re-ordering, IVTC etc… are what needs to hapen…

And because Topaz aims for a “simple user interface”, this is not an easy task… I personally only know very well trained and competent humans who actually can sort out what exactly was the original process for EVERY case of something that is delieverd on a DVD and obviously has interlace lines (I remember years of discussions about many 90s Sci-Fi Series with these issues…)…


I think most everyone is under the impression that the deinterlacing is done by AI.
Is it safe to say that the models under the interlaced progressive category do try to remove scan lines with AI?