New Unofficial DSCALER5 with 3:2 Pulldown for 1080i

You will find here links for a modification of Dscaler5 that offers a first level of Film Detection and 3:2 Pulldown. This version has been optimized for 1080i film content and should not be used for anything else.

You are welcome to give me your feedbacks, there is still a big margin for optimizations.

Summary
-------

Film detection is not an easy process and its efficiency depends on many factors such as the decoder, video card, quality of the compression. If a decoder can not properly detect if a content is a film, it will assume it is a video and try to deinterlace it, which means most of the time "blend" the two fields or discard half the fields and basically reduce the resolution from 1080 lines to 540 lines.

As far as I can tell most of the mpeg2 decoders available (except some hardware based decoders) do not do a very good job at identifying films. There are rumours the 81.84 nvidia Forceware drivers are supposed to offer much better 3:2 Pulldown.

If the new option "inverse telecine" is enabled in Dscaler and you are watching telecined content you have got now quite a high chance to watch it in its true high definition that is at last in 1080p and no longer in crappy 540p. The decoder will try to detect as well if some video content is detected between two movies portions (for instance commercials), the decoder will try to ignore it and then resync after.

It has quite a low CPU overhead and since Dscaler is software based decoder it can be used with ffdshow.

Link
-----
Access Source + Binaries + Guide through this page:

http://www.geocities.com/kzeuhkzeuh/index.htm

Credits
-------
Many thanks to the authors of Dscaler Michel Lespinasse, Aaron Holtzman, Gabest and John Adcock that have created this awesome decoder with great picture quality.

Thanks to FMalibu for his "blind IVTC" modification which gave me the inspiration for the full IVTC algorithm.

Please note that this version is not official and is not supported by the original Dscaler authors. I hope they won't mind me releasing it.

Enjoy it!


Kzeuh


Edit: Fixed Link

If no film content is detected, it will assume it is video and Dscaler will apply the the deinterlace process that is chosen in the Dscaler settings (weave or bob). There is no way at the moment except using ffdshow to specify a specific deinterlacer. I think I will leave this part to John

Anyway, I do not recommend to enable the inverse telecine enabled if you plan to watch video, parts of the optimisations rely on the assumption that the content is mostly film and it is likely to see sometime film content where there is in fact only video. Moreover, as long as it hasn't found any film content it uses a slower film detection algorithm which means that dscaler may consume more CPU than usual to watch video.

There is an algorithm that is supposed to detect breaks in 3:2 cadences due to glitches or to TV commercials. At this point it will switch back to the film detection mode. Therefore, during all the duration of the glitch plus a couple of extra frames required to do a proper film detection the weave effects will be back.

Well, before even my modification purely flagged content although they turn out to be 23.997fps were sent to reclock at 29.97fps. I will look at what I can do, but can't promess anything in the shortterm. I know FMalibu had been working on that..so may he can help.

Otherwise, can anybody give me some feedback on this modified Dscaler version regarding the PQ, smoothness or simply the usefulness so that I know it is worth working on it? Thanks.

This film detection algorithm as mentioned in my first top has been optimised for 1080i film content, in fact some internal thresholds are based on this assumption and it had to be simplified in order to leave enough CPU to handle HD. I must say I haven't even tried it on anything else but 1080i.

It gives some good results on 1080i that respects a 3:2 cadence either with flags or without flags. This corresponds to 95% of the HD movies I have tested. It checks only cadence breaks to be able to skip short video content like ads that have been inserted. Movies that have lots of cadence breaks per minute are not consequently handled.

Its goal is to enable people who enjoy the PQ of Dscaler to watch most of their HD movies. Since it is assumes the content is mostly film and not video you are likely to get a better PQ because most decoders usually switch back to video mode (and reduce the resolution) as soon as they have a slight doubt about the content. This could maybe explain your last observation.

You will find attached a new version of Dscaler with IVTC support, it is bit experimental so I will first ensure first it works correctly before I release the new source code.

What's new:
- more tolerant Film Detection that is when some film has been found a quite obvious video content will be necessary to switch back to video mode. The upside is that glitches and compression artfifacts are less likely to cause temporarly an unwanted return of video and its weave effects. The downside is that resync after a lost sync may take a bit more time

- new option to switch to 24fps, hopefull this is what expected BangoO...

- internal optimisations that should make the code faster. At least movies seems a bit smoother to me, but given the stuff I have done it might have the opposite effect on your system.

Edit: Updated the attachment, to make the new option 'Switch to 24fps' available

 

DscalerIVTC041105.zip 174.7431640625k . file

I will have a look at that, but it will be mostly for reason 1). Once a Film has been detected the CPU overhead is very low because there is as switch to a less precise film detection algorithm which is just there to detect potential return of video content.

However, improper film detection could be due simply to the fact that there isn't any 3:2 cadence. I have noticed that if remember correctly on CBS and FOX. I have read somewhere that some network in order to have more time for ads drop sometime a field and consequently break the cadence...

miketech,

The dll have been optimized for Pentium IV during the compilation. This could explain poor performances on non Intel processors. I might remove this compilation option in the future so that more people can use this Dscaler.

Otherwise just to know if my optimisations were efficient, would you say this version is slower or faster that the first one? thanks