[x264-devel] [PATCH]adaptive lowpass

[Eric Viscito]

Hi Gabriel,

On Thu, June 28, 2007 3:38 am, Gabriel Bouvigne wrote:
> But it's still usually a good thing to reduce your frequency range in
> order to increase precision of encoded freqs, once you are into the
> visible/audible degradation area. (of course, this is not RD optimal if
> you consider the whole spectrum)
> That's why every audio encoder is increasing its lowpass when you
> increase the compression ratio. Lowpass is better than a metallic
> wma-like sound. I think that this trade-off is also applicable in video.

I think your audio/video analogy is reasonable.  But isn't that audio
lowpass filter applied to the raw samples themselves prior to any
compression step, and not to some residual of a prediction step?  I think
you would get more perceptual improvement by applying a filter before
prediction and driven by analysis with a perceptual model.

>
>
>>> So the *adaptive* LPF provides a means to trade off the two types of
>>> distortions.  If the adaptive filter control is sound, this can be a
>>> benefit.
>>
>> It can adapt the amount of bits spent in high vs low freqs. But the CQM
>> is
>> still constant, so the LPF vs quantization noise is still not adaptive.
>
> Yes, you can not really decide locally about such trade-off. This is
> something that initially surprised me when going from audio to video.
> Ideally, a video coding scheme would allow you to specify the index of
> the CQM used for the current MB. With classical vlc coding, that would
> probably kill efficiency, but with cabac that would not impact much
> entropy coding.

Would you vary the lowpass filter in audio on a very local time scale?  Or
is it more that you filter certain signal types whenever they appear
because either the information is irrelevant or cannot be adequately
handled by the compressor?

You can still do different things with the quantizer of each macroblock
without adapting the frequency-dependent quantization weighting matrix.  I
think picking a reasonable CQM matrix and sticking with it and then
applying RD methods of grooming quantizer values is more promising.

BTW, there have been studies - at least on H.263, if I recall correctly -
showing that weighting matrices did not really help at all in that
context.

Regards,
Eric

_______________________________________________
x264-devel mailing list
x264-devel at videolan.org
http://mailman.videolan.org/listinfo/x264-devel