Weird blocks when vhq enabled

[Zarxrax]

I have tried enabling vhq mode 1 and mode 4, and this feature always causes blocks very similar to the shit blocks that were in divx3.11a
I KNOW that I have read something about this back when vhq was first released, but I've searched, and nothing comes up. Also since everyone is saying vhq shouldnt negatively effect quality anymore, im not quite sure whats wrong. I'm using koepi's 14052003 build.

[JohnMK]

In dark places? Encode Star Wars: Attack of the Clones, 1 hour+43/44 minutes in, you should see some pretty horrible blocking. This was with VHQ4 enabled. :)

[Zarxrax]

No, not in dark places, but in spots where there is lots of motion. Its like luma inverted blocks or something.

[Defiler]

Me too:
http://forum.doom9.org/showthread.php?s=&threadid=41616

[Zarxrax]

Ah, I didn't check that thread because the title made it sound like a different problem.
Mods delete this thread if you want.

[CruNcher]

@ all
this is a Trellis Quantization problem if you don't want it don't use it i got exactly the same problems a week ago and reported it in the channel

http://www.mufflastig.com/CruNcher/XviD/trellisquantbug/

[Defiler]

You've verified that simply turning off Trellis is enough to prevent the problem?

[CruNcher]

@ Defiler
jep exactly but only for this encode and only for this special scene wich is a Strobo light effect Scene where Lights go off and on very quickly if i read now that also this problems can occour with VHQ i belive it could have something todo with the R-D cause VHQ also uses R-D now maybe there is some problems with Motion Estimation in such scenes i can't say that for sure but for this encode rurning off trellis helped and VHQ was still on Wide Search so its hard to tell if people now say it can also happen with only VHQ on what the problem is i would tip then in a logical manner on the R-D part both useing it but maybe Syskin can bring light into that :)

[Defiler]

I experienced the same thing. Turning off Trellis, but leaving VHQ=4 solved my problem.
Has anyone seen them with Trellis enabled, VHQ off?

[Calculon]

Trellis caused this problem for me as well. I just turned it off and encoded again(a captured episode of Sliders) and it was fine. I had VHQ 4 on both times.

[angelyote]

another me too post.
Noticed some ugly blocking in some of the dark solid areas, like peoples hair. Turned off Trellis and kept VHQ 4 and beauty once more. That's why it's in the debug section though.

Dave

[Animaniac]

I just encoded a 2 hour movie with VHQ 4 and R-D, and no weird blocking. I can safely say it's my best encode yet.

[jrmillerUT]

what do you mean by Trellis and R-D...

[mf]

Originally posted by jrmillerUT
what do you mean by Trellis and R-D...
Trellis: Trellis Quantization
R-D: Rate Distortion

Just look it up in google...

[soujir0u]

If you're using ffdshow try using Xvid.dll to decode the video. I got this strange problem with one of my recent encodes...

[BiaTch 5.0]

VHQ4 goes about 3x slower on my PC than VHQ1 so a movie that takes 2 hour to encode would take me 6 with VHQ4, is it worth the time?

[Gaia]

Yes ofcourse but you have to try yourself. It's up to you. I don't care about speed.

[stax76]

VHQ4 goes about 3x slower on my PC than VHQ1 so a movie that takes 2 hour to encode would take me 6 with VHQ4, is it worth the time?


I made a couple of compressibility tests using DVX which uses a complete different method than Gordian Knot. From VHQ 0 to VHQ 3 there was a constant increase of the compressibility but VHQ 3 and VHQ 4 had the same compressibility but VHQ 4 was a lot slower than VHQ 3 so my conclusion is VHQ 4 is a overkill

edit: I saw a note by a XviD developer that stated comp tests aren't the full truth

[CruNcher]

@Dolemite

Vhq in the latest available koepi build is b0rked it has been fixed in the newest internal build and also in the cvs. Its in testing stage @ the moment, because it also contains syskins new ipb decission code which is not in cvs and needs to be tested first.

[stax76]

@CruNcher

thanks, I'm a XviD n00b using it since two weeks, currently reading all the guides and topics, quite a lot things to read :(

[mellon]

I have coded a litte tool which helps to spot errors in an encoded movie. Originally designed for Divx 3.11 to find shit, it now works with any kind of video material as long as you have a vfw decompressor installed which can output to RGB32.
http://www.hu-berlin.de/~h00542gu/MOV_ANALYZER.zip
It works by calculating the standard deviation between two frames, the original and the encoded one. The greater the deviation the more likely it is to have a frame containing shit. Every frame with a deviation above a given threshold will be treated as bad and optionally saved as bitmap. But you have to decide by yourself if the frame contains errors. Some frames with a very high complexity may have a deviation above the threshold. A higher quantizer usually means a higher deviation. So you can play around with this threshold.
I was able to find errors in frames with only 3 bad blocks (8x8 pixel).


mellon

[unmei]

good idea :D
but the trellis bug showed up very rarely for me, maybe every 3-5 hours one block, always only one block per frame, so detecting 3 or more wrong blocks wouldn't help :(