[XviD-devel] 2pass curve compression

[Dirk Knop]

Hi,

yetserday I talked some time with Foxer about "his" alt CC curve 
compression algorithms.
In th end, the descriptions can be made a little more understandable in 
a less-technically language:

High+low distance and min. relative quality affect the amplitude of the 
scaled curve, while curve aggression low, medium, high affect the 
reaction speed for the overflow treatment.

High distance should have a name that reflects that it represents the 
upper bound of the curve. Something like "Max framesize (% of average 
framesize)" and for low distance accordingly "Min. framesize (% of 
average framesize)". But this isn't all the truth. What happens out of 
those borders is that this Min. relative quality stuff get's applied. 
There you have too many possibilities what to do as well.

I would like to see new default values implemented into XviD which gave 
me good results (I tested on matrix and rush hour2 again) and someone 
over at doom9 tested it with fight club, which now finally looks like he 
likes it ;)

Ok, those defaults should be:
iframe boost 0
below keyframe distance 10
bitrate reduction 20

payback delay 250 (this has still influence on the curve, even when that 
alt CC stuff looks like it shouldn't!)

using alt cc as default (!)
curve aggression medium
high distance 500
low distance 90
min rel quality activated, automatic mode, strength set to 30%

These settings result in a nearly linear downscaled curve, while 
applying a mechanism which prevents from already small frames trying to 
get over-quantized.

A little weird to setup all this, but the viewing pleasure of the 
encodings is remarkable :)

Long-term I really see the need to replace all that CC stuff with 
something easier.

I'd like to have a linear scaled curve with a high and lowpass, some 
overflow treatment which works like it now works (scene based), and some 
mechanism to prevent too small frames to get quantized more ;) Maybe a 
reaction delay for the oveflow treatment as well.

But please, nothing so complicated, who would ever guessed that these 
values can help the encoding?

Ok, just wanted to share thse results.

Best regards,
Koepi