Is it now SAVE to use motion vector hintfiles ?

[ookzDVD]

SAVE =
FASTER (2nd pass) than without MV hintfiles enabled
BUT
SAME QUALITY without MV hintfiles enabled ?

[Koepi]

It will _never_ be the same quality.
But it won't be "better" or "worse" as it depends on your taste.

It's safe to use.

Some people like the way the video looks with MV hints more than the usual way ( I for one do so ;) ).

(I heard of no person who disliked the way the video looks when using MV hints. MV hints should give it a more natural quality.)

[rui]

If only we didn't have the modulated quantization+MV hints bug... :(

[Dali Lama]

ookzDVD, Koepi is right. Just to add another thing. MV Hints helps stop the ringing and "talking head" phenomina (background moves around) that are prevailent in anime. -h pointed this out when MV hints first came out. So, I would reccommend using it.

[saVe]

@dali:
i don't want to doubt you, but isn't that due to epzs? i read that pmvfast has the tendency to invent "wrong" vectors if they are more compressible than the original ones. since epzs doesn't show that behauviour that would be an explanation.

in some testing i found out that the newest builds of nic and koepi with epzs enabled didn't show any "dancing" backgrounds any more, even though i didn't use mv hints.

as always, correct me if i'm wrong ;)

p.s: got a little off topic, sorry for that!

[rui]

I agree with saVe. I even created a thread about this http://forum.doom9.org/showthread.php?s=&threadid=24065
In my experiments, the builds with epzs are always more sharper than the instant builds, which use pmvfast.

[Dali Lama]

I understand what you are saying rui and Save, but I still think that MV Hints helped out that situation and should be used, because:

1. -h said so. (he coded mv hints afterall)
2. I tested using the same EPSZ^2, setting (6), and found that noticable ringing in anime diminished or disapeared with mv hints. Some have pointed out that theoretically mv hints will decrease quality, but as Koepi has said, its the perceived quality that counts.
3. MV hints also saves time on the second pass.

Nice to read your inputs rui and save,

Dali

[unplugged]

Why this difference between 1st pass MV calculations (MV hints) and 2nd pass MV?

This MV calculations are preliminary involved by DRF decisions?

Because the 1st pass is performed with fixed quant 2... so MV detections are little different compared to 2nd pass MV hints (redone)?

[canadian_fbi]

i think where some of this talk of MV hints decreasing quality came from is that the use of the hints always (i think) increase quant values. where things get shady is that it seems that the slight increase in average quant gets offset by better motion vectors. and the quant increase is small anyway, so it's hard to really notice. correct me if i'm wrong, but i've gotten the idea that since the MV hints are done at quant 2, when they're used in the second pass, it's using the better (and bigger) motion vector hints from the first pass, so the quantizers have to be slightly bigger to compensate. giving bigger quants but better motion. so the effect on the overall quality sort of evens out.... maybe. :)

[MfA]

If you were to work under the assumption that the "most truethfull" motion vectors would give the least annoying artifacts then the first pass motion vectors shouldnt even be determined from the quant 2 encoding, but from the original frames.

It would be interesting to determine how motion vectors varied with increasing quantizers, if the variation is mostly noiselike (ie no greater correlation between say the vectors at quant 4 and 2 than say quant 4 and original sequence MV's) then calculating the motion vectors from the original sequence would be best regardless.

[-h]

It would be interesting to determine how motion vectors varied with increasing quantizers, if the variation is mostly noiselike (ie no greater correlation between say the vectors at quant 4 and 2 than say quant 4 and original sequence MV's) then calculating the motion vectors from the original sequence would be best regardless.

I can't say whether it's a bug/feature of the current ME implementation or not, but many of my TV captures exhibit *extremely* annoying shifting of static backgrounds when a foreground object is moving (visible but diminished with quant=2 vectors, extremely annoying at quant>=4), and pmvfast decides it'd be better to shift the background MVs to match the foreground. The last time I looked at ME, the current block's quant value was being used as a bias in the SAD decision, so I assume this is the cause (which makes sense, as higher quants will skew the cost of a MB from texture storage to MV storage, however slightly).

I initially thought it was only due to noisy sources, but it appears to happen on clean input as well. I'd like the ability to perform ME on original or reconstructed images at run time, however I have no time to try this.

-h

[ookzDVD]

@Koepi
thank you for the support. ;)

@rui
how about epsz^2(ultra high 6) compared with epsz (very high 5)?

[rui]

Good question! ;)
i will conduct a small test and compare the clips in Vdub, and then i will post.

[Nic]

Sorry to be OT:

But, http://nic.dnsalias.com for the latest build.

It should have removed that "grey block" artefact....

-Nic

[Koepi]

Well, I can be OT too:

Changelog:

XviD-02052002-1:
- EPSZ and EPSZ^2 activated.
- "Grey block" bug fixed
- Updated to include Nic's latest DirectShow playback filter

The overall quality should be better with this new release since some significant changes were made to the XviD core since my last release (XviD-17042002-1).

Thanks Nic for your DSF again :)

Best regards,
Koepi

[Belgabor]

Originally posted by Koepi

- "Grey block" bug fixed


YAY! :D

[rui]

Originally posted by ookzDVD
@rui
how about epsz^2(ultra high 6) compared with epsz (very high 5)?

Ok, test made.
EPSZ^2 is sharper. That's it. :D

I compared two clips from the trailer of the movie The Replacements in Vdub, using in both encodes the EXACT same settings, aside from motion search preciosion 5 on one and 6 on the other.

Then i pointed to the exact same frame in both, copied it to Paint Shop pro, and compared both. It was clear that the epsz^2 (6) was sharper (more details) than the epsz (5). But this was comparing still frames, in a moving video it would be dificult to notice the diference.

[MfA]

I guess -h will hate me for this ... but there have been quite a few papers which detect edges in blocks and adapt the block motion search to avoid the worst of the problems mentioned here.

Look for "edge oriented block" if you have access to IEEExplore (not a lot of stuff on the web).

[ookzDVD]

hi rui,

thank you for the test.
I've been using it for my 1 CD rip.