Best MPEG-2 encoder for difficult footage?

[Trixter]

I've got some fairly gnarly footage that is going onto a DVD and it has so much motion and interference that I can't get a decent encode out of it. I've tried CCE and TMPGEnc and haven't been happy with the result from either. Procoder 2 does a slightly better job, but I wanted to ask the people who bang on MPEG-2 encoders all day long: What is the best software MPEG-2 encoder out there?

If you'd like to read the background on this, and see the "impossible" footage I'm working with, see this thread: http://forum.doom9.org/showthread.php?t=110690

[Revgen]

The footage is TFF PAL video. I found out using AssumeTFF and SeparateFields() and the video scrolled smoothly in Vdub. Using AssumeBFF and SeparateFields() gave jerky motions.

I've encoded 3 different versions with Quenc .70

All encodes are 8500 KBPS, Open GOP, GOP=15, No Scene Detection, No b-frames, No trellis, VBR ON, HQ ON, 2 pass ENABLED, Max Bitrate 9500KBPS, TFF ON, Interlaced ON, DC Precision=10.

Except for the following:

1) Quenc QLB OFF. This encode uses No QLB. Not as blurry as QLB but more blocky.

2) Quenc QLB ON. Not as blocky, but kinda blurry.

3) Quenc W/CG_Matrix from the Xvid Matrice collection. Images seem sharper than both QLB On and OFF while blockiness is not as bad as QLB OFF but not as good as QLB ON. This appears to be the better of the 3.

IMHO they all look the same as long as I play them full-speed. I can only see the visual differences when I look at them in still frames.

You can download the clips here. (http://rapidshare.de/files/20076770/SOTA.7z.html)

[Trixter]

All encodes are 8500 KBPS, Open GOP, GOP=15, No Scene Detection, No b-frames, No trellis, VBR ON, HQ ON, 2 pass ENABLED, Max Bitrate 9500KBPS, TFF ON, Interlaced ON, DC Precision=10.


I'll check these out, but unfortunately due to the high number of audio tracks in the DVD the max bitrate can't go beyond 8700kbps. But thanks very much for trying various things, I'll definitely check out your efforts.

Procoder 2 on mastering quality produces the best results, but unfortunately the beginning of the video has areas where only chroma is changing and not luma... and procoder mangles those areas! Ironically, CCE does the best with those areas.

Maybe I should encode using multiple encoders and just cut/join on I-frames, using the best sections from each? What is the best cutter/joiner out there right now?

[Revgen]

I use MPEG Tools that come with TMPGEnc. I don't know of any free ones though.

[Mug Funky]

you can drop DC precision to 8-bit, but that probably wont buy you much at all :)

i'm afraid you're not too likely to get the sharpness you desire out of it - it's a tradeoff, and you'll either get sharp + blocks + crap chroma or soft + less blocks + still crap chroma.

i'd try HC (the latest) on best profile, 1 or no b-frames. probably 1 because the ME advantage offered by reducing b-frames will not really pay off with a clip such as that - motion is unpredictable enough as it is. though try both to see what happens :)

you could try a flat matrix (all 16s for intra and possibly 32sfor inter, though if ME is proving useless perhaps smaller numbers can help here to tip it in favour of correcting mismatches), and you could even try a very slight filter on your source.

i really don't think you should expect much out of this video though... just be glad it's a test signal and not a piece of natural video that's bollocking up :)

[dragongodz]

unfortunately the beginning of the video has areas where only chroma is changing and not luma...
you may also want to try QuEnc 0.70 with extreme setting enabled(but not trellis) and see how that handles that.

EDIT:
oh and asking whats best is against the forum rules. link to them is at the top of the page.

[Revgen]

you may also want to try QuEnc 0.70 with extreme setting enabled(but not trellis) and see how that handles that.


Hmm... I just tried the Extreme option and lowered the MaxBitrate to 8700KBPS (per Trixter's suggestion). The Extreme option seems to give slightly better yet noticible results. I never use this option because the benefits are usually never apparent in normal video. In this case, it actually helps.

[Revgen]

you can drop DC precision to 8-bit, but that probably wont buy you much at all :)

i'm afraid you're not too likely to get the sharpness you desire out of it - it's a tradeoff, and you'll either get sharp + blocks + crap chroma or soft + less blocks + still crap chroma.

i'd try HC (the latest) on best profile, 1 or no b-frames. probably 1 because the ME advantage offered by reducing b-frames will not really pay off with a clip such as that - motion is unpredictable enough as it is. though try both to see what happens :)

you could try a flat matrix (all 16s for intra and possibly 32sfor inter, though if ME is proving useless perhaps smaller numbers can help here to tip it in favour of correcting mismatches), and you could even try a very slight filter on your source.

i really don't think you should expect much out of this video though... just be glad it's a test signal and not a piece of natural video that's bollocking up :)

I tried HC and it didn't look as good as Quenc for this video. Even Quenc without the Extreme or CG Matrix settings appeared sharper than HC while blocking remained about the same. The CG matrix video looked sharp and with less blocking. I used the same parameters in HC that I used in Quenc. Unfortunately HCEnc doesn't support Xvid ".txt" CQMs, so I can't test with the CG matrix.

[Trixter]

you may also want to try QuEnc 0.70 with extreme setting enabled(but not trellis) and see how that handles that.

EDIT:
oh and asking whats best is against the forum rules. link to them is at the top of the page.

I'll give that a shot, thanks.

BTW, I was asking for what was best for a particular set of footage, so I didn't think that was against the rules. The "don't ask what is best" rule IIRC was to prevent advocacy flamewars :-)

[siddharthagandhi]

What happened to the rule about no "best" questions?

[Revgen]

He asked what was "the best for his footage", which is basically a complex hi-tech demo. That's different than asking what was "the best" in general. Doom9 typically hates questions that ask "what's the best" without any criteria to establish what the tool is best at.

[siddharthagandhi]

I asked a similar question, when I first registered at the forum, and I was busted. I asked what was best for HD and high bitrate encoding...

[Trixter]

Right, because that's too generic. If you had provided a specific video clip, that's different. I had a very VERY specific piece of footage, which I provided, that NO encoder could get right so that's why my question was okay.

IIRC, the rule was created to prevent advocacy flamewars like "QENC is the best! "No, HC018 is the best!" "Screw you!" "Screw yer mom!" etc.

[Guest]

Best MPEG-2 encoder for difficult footage? Please read and follow forum rules, specifically, rule 12: do not ask what's best. Thank you.

http://forum.doom9.org/forum-rules.htm

[Guest]

What happened to the rule about no "best" questions? Please read and follow forum rules, specifically, rule 3: keep the focus. You should have reported this thread to a mod, not complained publicly like this. Thank you.

[Trixter]

Please read and follow forum rules, specifically, rule 12: do not ask what's best. Thank you.


I thought it was pretty clear from the thread that I was asking which encoder could encode my difficult test footage the sharpest, something which is asked on this forum all the time. In the future, I guess I'll avoid the word "best".

[onesoul]

Maybe the thing to ask should be "what could be better" but then again asking "what could be the best" should be the same, because reasonable people would only answer what is best with the current knowledge they have and every other reasonable person would know that and respect it.
But who knows what other interpretation someone would have, that's why the rule exists.

So to achieve perfection the best is... I am kidding ;)

I will try the sample footage you provide. But for what I tried before with interlaced sources, I would go for HC or maybe Quenc.

[siddharthagandhi]

"Please read and follow forum rules, specifically, rule 3: keep the focus. You should have reported this thread to a mod, not complained publicly like this. Thank you."

The reason why I didn't report it is because four users were giving help, so apparently the rule didn't apply to this.

[Guest]

The reason why I didn't report it is because four users were giving help, so apparently the rule didn't apply to this. And since none of those four was a moderator, then it's quite possible that no moderator took notice of it. Anyway, this is totally off-topic discussion. You need to keep this to PMs, else you risk being struck.

[onesoul]

I was going to try each encoder but I couldn't pass convertion successfully from the rgb32 source file.

Original
http://img92.imageshack.us/img92/8614/avsc30000572wb.th.jpg (http://img92.imageshack.us/my.php?image=avsc30000572wb.jpg)

converttoyv12(interlaced=true) ruins the picture so unless there is something wrong with my setup, HC and Quenc cease to be an option, (they only take yv12 as input).
http://img108.imageshack.us/img108/6899/avsc20000574be.th.jpg (http://img108.imageshack.us/my.php?image=avsc20000574be.jpg)

converttoyuy2() seems fine except there appears to be some kind of pixel shifting to the left, so this could explain some of the blurring you were having when encoding to mpeg2 with CCE.
http://img108.imageshack.us/img108/439/avsc10000570kr.th.jpg (http://img108.imageshack.us/my.php?image=avsc10000570kr.jpg)

I tried to correct the shifting but my poor avisynth programming didn't allow me to do this, I aslo tried Chromashift plugin from SiWalters without success, so if there is anyone who would help I too would be interested.
And also is it possible to only shift luma?

[dragongodz]

converttoyv12(interlaced=true) ruins the picture so unless there is something wrong with my setup, HC and Quenc cease to be an option, (they only take yv12 as input).
how about if you seperated the fields, did the conversion and then weaved the fields back ?
and of course it doesnt matter what the encoder takes as input it will be converted to yv12 anyway since thats what mpeg2(for dvd etc) use. so those that accept other colouyr spaces must just do the conversion internally.

[onesoul]

how about if you seperated the fields, did the conversion and then weaved the fields back ? That didn't work.

and of course it doesnt matter what the encoder takes as input it will be converted to yv12 anyway since thats what mpeg2(for dvd etc) use. so those that accept other colouyr spaces must just do the conversion internally. You are right, so the solution would be to deliver a corrected yv12 stream through avisynth. Is it possible to do that?

CCE can take rgb32 as input and the encoded file presents same problem as converttoyv12(interlaced=true) of avisynth. It seems to blur every 2 interlaced lines.