Cvs 2002-02-26

[Koepi]

Ok, 18 MB oversized.
There's a serious bug in curve compression, set it to 0 for now.
I'll investigate further (the "bug" I've seen before is just a cosmetical one, a variable doesn't get updated...)

Regards,
Koepi

[sierrafoxtrot]

@koepi: just a thought ... is the bug due to

i) asymmetric curve correction or,
ii)curve correction in general?
iii) curve correction with Modulated Quant-type? (oh, just read rui's post again, he's using H263 and still getting an overshoot ... so i guess it must be due to the curve correction)

i might do a 2nd pass with lo/hi set at 25/25 with modulated quant-type, just to see what happens. let you know in the morning.


@nic: i share your frustration mate, am trying to write a goddamned PhD thesis, feels like heads and brick walls ATM. remind me to buy you a pint (or several) if we ever meet. :D

cheers.

[Koepi]

head, bricks, wall?

You should definatly get some sleep ;) Your PhD would else get very stuffed by macroblocks ;)

Good luck with that anyways - it's one thing to write a PhD - and another to throw the whole studies away with a simple diploma ;)

Regards,
Koepi

[sierrafoxtrot]

@koepi

LoL, unfortunately i (quite stupidly) got myself into a full-time job that leaves me too exhausted to do much until i finally perk up at about 11pm (nightowls unite!). and all that sleep deprivation is acting like tempsmoother(5) :D

[Koepi]

Humm, my vocabulary starts to suck, I didn't get a word, got some blurryness like tempsmoother(5beck's), after that spatial soften (5 pint of cider),...

But cheers anyways :)

ATM. increasing tempsmoother to 6, 7 or 8 (or even more) ;)

[-h]

This sounds bad.

I can't look into it for another 6 hours however..

-h

[wing1]

kudos to the developers :)) This is getting better by the builds. 2-25-2002 build yields extremely sharp capture result and fast :) 2-26-2002 even faster:)

At this rate, perhaps no need for deinterlacing at all.

[Foxer]

I apologize for my messed up asymetric curve compression formula. I was very tired when I thought it up :)

I've made it much much more accurate and later, if -h drops by, he'll commit the new source I e-mailed him.

In the mean time, putting the same value in both boxes for symetric curve compression works fine.

[Gaia]

I can't understand you guys.

Yesterday i tested Blade, used latest Nic build. Tried again this new 2 pass thing, it just made blocks more visible. Tried oldway, scaled curve with Gordian Knot. Filesize was supposed to be about 600 mb but it was almost 1.5 Gb!

So far i have got best results with about month old Nic build.

I don't know, maybe i am doing somenthing wrong.

I use:

h. 263 quantization
Motion Search Quatization 6
I-Frame min 2
I-Frame max 5 and smooth quantizer
Resolution 640 & ?

All these new things just seems to make picture more blocky, not better. Especially modulated quantization. I have trid lots of different settings, it just doesn't work for me. Put all the lights off then you check out the movie, you can't see most of the blocks in bright room.

Maybe i should test latest Koepi build but i don't think it will help.

You just hope these new things would make quality better but it doesn't. Just encode even short clip oldway and with these new things, i am sure quality is better without them. Just look real close.

[Nic]

Hi Gaia,

Could you tell me which build (the date of it) that you are using that gave better results....I hear this alot & wouldnt mind testing the theory over the weekend.

@SierraFoxtrot:
PhD.....Double Argh! :) (Im downloading the proper sources from projectmayo.com now....so ill give it another try tonight) (But I have also just got Command&Conquer Renegade (And im still playing medal of honour) Where does the time go? :)


-Nic

[rui]

Originally posted by sierrafoxtrot
@koepi: just a thought ... is the bug due to

i) asymmetric curve correction or,
ii)curve correction in general?
iii) curve correction with Modulated Quant-type? (oh, just read rui's post again, he's using H263 and still getting an overshoot ... so i guess it must be due to the curve correction)



Beware that i only used H.263 in the first pass. For second pass i used modulated quantization.

To Gaia:
I can't understand either. Believe me, the Platoon rip i did last night, using the settings i posted, was great. It was a 1 cd rip.
Can you see any diference by using motion search precision 6 instead of 5? How much time longer does the encoding takes by using 6?

If y remember correctly, Balde is a very dark movie, right? Maybe that is the reason you see more macroblocks. But then again, Paltoon is a very dark movie in a lot of scenes, and i assure you that my encode was very good. Sure it had some blocks here and there, but few of them, and for 1 cd rip that is very good.
Once i tried to do a rip from the movie "I know what you did last summer", and there was a scene on the beach at night. Both divx4 as SBC gave me lots of blocks, divx4 less because of the codec's blurier nature. But it was visible that SBC had more details. Now, with XviD, i can have those details back, with just a tinny bit of increment in blocks in comparison with divx4.
By the way, what is your monitors resolution?

Keep trying, and best of luck to you.

[sierrafoxtrot]

hey all. just to confirm that it's probably the asymmetric curve compression code. tried a couple of 2nd passes, one with hi/lo 25/25 and the other 25/10, and 25/10 was oversized by 14Mb, but 25/25 was spot on.

hope this helps ... :(

@nic games are the work of the devil!! LoL had to burn my copy of max payne, it was starting to possess me ... :devil: but having said that ... good luck and all ... ;)

[rui]

sierrafoxtrot, thanks for the tips.
But i believe that Foxer already sent to -h a corrected code.
So it's only a matter of time that a new build appears with this bug taken care.

[Franko30]

OK - good morning to everyone.

I'm done with Demolition Man for the second time (first with Nic's Feb25 build and symetric curve compression, payback 240frames; second with Koepi Feb26 build, asymetric curve compression low10/high25, payback 480 frames) etc.

And Voyager has been encoded for the second time, too, with the latter build, but with quant modulation on for the first try and off (h263 only) for the second try.

Asymetric curve compression looks really great, oversize and other settings have been discussed already.

But these Voyager encodings convinced me of one thing:

I'm not going to use quant modulation for one CD rips any more. Although it loks a tiny little bit sharper, the blockyness issue (dark parts, noise on static walls etc.) is s.th. that annoys me. Maybe it's the little bit of extra noise from my analogue source (although normal bicubic resize and, sometimes, the "little noise" filter should take care of that). I just like the smoother look of h263 for my one CD rips.

So, if the oversize bug has been taken care of, I'm a really happy encoder again, going strong for movie number 2000 - "only" about 800 to go :D :D

Thanks for all the great work so far!

Frank

[Gaia]

Nic: I think the build i am talking about is 30.1.02.

Some people say difference beetween motion search precision 5 and 6 is very little but what i have tested it's huge! It's slower ofcourse but 2 h movie, maybe about 30 minutes slower, not sure about that.

Blade is quite dark but i am talking about lighter scenes too, that new 2 pass thing just doesn't work for me.

[rui]

Gaia:

This just crossed my mind. It’s a little thing, but nevertheless, i remembered to ask you this.
The latest builds use for default the XviD fourCC, so when playing the avi’s one is using the XviD DS filter, and it doesn’t have post processing yet. The older builds used for default the Divx FourCC, and used the divx4 post processing.
Could this be the cause that you see macroblocks?

[Gaia]

No it's not the reason. Have to do some more testing with latest build.

[Actron]

Ok, heres my experience with the newest build :

Its UNDERSIZED (*huh*) (edit : about 50mb of 1200mb undersized...)

sets are as following :
bitrate about 1700 (could be the reason)
mpg quant.
frame limit 1-10
iframe limit 1-6

curve modulation:
high 25%
low 15%


greetings

act