From my unpretentious opinion alt-CC needs motion influence

[unplugged]

I see tons of posts regarding new compression curve method and some positive comparison, happier results.

I have tried old-CC: good results.
I have tried new-CC: some parts good, some parts perfect, some parts ugly.
I have tried no CC (VBR): a litte confused... seem identical as first.
1'50'' medium action movie at 704x432.

But what magical results :rolleyes: do you expect from redistribute VBR quantizers having as input data only the *same* list of quantizers?

About the new CC, correct me if I'm wrong, but doesn't it collect *only* the *weight* of each frame and thus apply 2nd pass VBR dynamized or normalized with CC *arbitrary* rules?

For me this has not much sense as happen with ABR, even if ABR has his reasons...
In this case I have a name: AVBR
Arbitrary Variable Bit Rate.

Motion and/or something regarding the per-frame complexity should be considered...
???


P.S.: sure, I'm for pure VBR, based on *detected* quantization rules.

[-h]

I'm not sure what you were getting at, but all of XviD's 2-pass modes are based on taking bits from large (high-motion) frames and giving them to small (low-motion) frames. The curve tweaking is just a variation on this theme.

The 2-pass modes could be altered to take into account the frame's motion, but a much better solution IMHO would be to integrate some kind of motion masking system into the core itself.

-h

[unplugged]

Thank you much -h for your explanation and *presence*.

Does all large frames necessarily correspond to high motion frames?

As statistics...


There could be some eye-susceptive large frames having high motion, and thus getting badly a bitrate cut...
(truly high or at least *detected* as high)

[-h]

Does all large frames necessarily correspond to high motion frames?

Yep, unless the encoder detects "fake" motion, i.e. fades or superimpositions. I-frames are also artificially large, but they're treated differently anyway.

-h

[Doom9]

just to throw in something from my early nandub experience. When we were still playing around with quantizers I spent a lot of time checking nandub output and check the debug log for parts I didn't like.. and in many case adjusting quantizers was useless becuase in the end I had a good looking video that was severely oversized because some parts that looked bad did not have a high motion value so I had almost the whole movie encoded using low quantizers which completely screwed the size predictability. So.. motion can't be it.. there has to be some hvs adapted quality measure on a frame level but this is rather hard to achieve.

[MfA]

I think calling them high motion frames is slighting misleading. Complex motion describes it better, what causes large frames is motion which cannot be accurately captured by motion compensation. Bitrate is an ideal indicator for that.

The trouble is that motion which is complex for a block based translational motion model might not be hard to follow for our eyes. Good example's are rotation and lots of moving particles over a static background. Or some nice shimmering water.

BTW someone from m$ came up with a very nice method of pre-processing which uses the localized high frequency content in the motion compensated residual to determine which parts of the frame to filter (he also takes into account the high frequency content in the original image, so by comparing the two he can detect camera noise which should also be filtered). This is much better than filtering (or quantizing more heavily) everything that moves (fast) because if objects can be easily tracked by block motion compensation there is no reason not to retain all the detail. It was in a paper called "Motion sensitive pre-processing for video". Some papers describing a method which tries to achieve some of the same goals, but by working on the residual directly is here (http://ivpl.ece.northwestern.edu/Research/Current/PrePost/PrePost.html).

Marco

[OUTPinged_]

-h, is there any way to check if codec detect's "fake" motion at user end?

[unplugged]

Originally posted by MfA
This is much better than filtering (or quantizing more heavily) everything that moves (fast) because if objects can be easily tracked by block motion compensation there is no reason not to retain all the detail. It was in a paper called "Motion sensitive pre-processing for video".

Great MfA :D, very great
this tecnique is often the sensation that give me the latest codec from DivXNetworks, especially with Q-Pel activated.
Analyzing and comparing some slow-motion encoded frames I have discovered that it's tendence, now very polular, to add *random* detail in some frames, it's not always casual as seem...
...almost if you don't remain to analyze only "that" criticized frame.

Starting by taking in exam a frame from a slow motion session (original content).
If you "look around", I mean the previous and the next frames, you can see in motion areas there are some moving pixels that aren't switched on costantly with same level of luma, generally are involved by classical luma fluctuations (thus they aren't always strongly present in all those frames).

Looking with same accuracy the same content encoded to DivX5 has given me an incredible surprise when I have seen moving pixels as "normalized" from frame to frame!!!!!!!!!!!!! :eek:

I have seen pixels little attenuated in alternate frames getting more luma and thus clearly and linearly visibile in all output frames!!!

Pixels seem to get more *strongly* "joint" and levelled in time axis... (frame1, frame2, frame3...)

A kind of smart detail moving detection, or better, as MfA says "Motion sensitive pre-processing for video".

:eek: simply stunning :eek::D

[-h]

The trouble is that motion which is complex for a block based translational motion model might not be hard to follow for our eyes. Good example's are rotation and lots of moving particles over a static background. Or some nice shimmering water.

True, but we work with what we have..

BTW someone from m$ came up with a very nice method of pre-processing which uses the localized high frequency content in the motion compensated residual to determine which parts of the frame to filter (he also takes into account the high frequency content in the original image, so by comparing the two he can detect camera noise which should also be filtered).

This is very very cool. Why only mention it now when I'm leaving XviD development for a while :)

-h

[Nic]

How long is a while ??!? :)

-Nic

[-h]

How long is a while ??!? :)

Leaving May 7th, won't be able to develop (or actively view forums) for 3 or maybe 4 months. Moving to the USA, doncha know :)

-h

[Hanty]

Gah! First DGraft and now -h...

Somebody preserve his sperm before he disappears.

[kvs]

Yeah, this board will feel kinda empty without -h.
better clone some developers while we still can!
Who's up for some CTXD*?
:D




*=Capture The XviD Developer

[gldblade]

Reminds me of what happened with Vvnn... :( So, what did happen? I remember something about him moving, but nothing after that.

[ammer]

Cool. ol' USA(100s years old hehe) Hey where you going to? Whats going on there? Heard a lot bad people who drink and have wild partys over there...oh was I talking about college kids there. hehe

[-h]

Cool. ol' USA(100s years old hehe) Hey where you going to? Whats going on there? Heard a lot bad people who drink and have wild partys over there...oh was I talking about college kids there. hehe

Starting in WV, going to Seattle, NY, then settling down in SC. Or so the plan goes.. not actually sure what's happening yet, I'm depending on what my cousins have planned for me.

Should be a fun few months.

-h

[Teegedeck]

Boy, whatta chance! Enjoy it, -h, I really envy you. Right now I wish I could fly to the US - a close friend of my 'spouse' and me is throwing a really big wedding. Even if we'll miss you terribly :p I'd still say enjoy, enjoy, enjoy!

[ammer]

Touring the states ey. Sounds fun. But be warned, Seattle and some of the northern states have pretty cold weather so you might be the wiser to dress warmly. Then again it's not winter now so it probably won't be that bad. Anyway, enjoy yourself and have fun. ;)

[OUTPinged_]

btw -h, was that macroblock problem (aka shit) solved or it will be in for all that time?