[x264-devel] Re: CABAC performance

[Alex Izvorski]

On Thu, 2007-05-31 at 11:41 +0200, Son Minh Tran wrote:
> Dear All,
> Thank you very much for your reaction on the topic I launched up, 
> especially for the documentation from Peter List.
> I would like to go further a little bit on modifying the state of the 
> CABAC encoder. So in the the code of  x264, I  change the static table 
> x264_cabac_range_lps and x264_cabac_transition with the dynamic ones 
> (calculated in correspondance with the state number and quantized Range 
> given as options of x264). Then I made some proper changes concerning 
> the new dynamic tables like
> -in x264_cabac_context_init: I use the init state ~ 0,5 probablitity for 
> all contexts
> -in x264_cabac_encode_decision,  I deployed the new tables for 
> transition and range calculation
> My question is do I miss some steps in changing the state number and 
> quantized Range for CABAC. That is because with greater number of state 
> and finer quantization of Range (so the CABAC model is much closer to 
> the theory arithmetic coder) I don,t have a solid increase in 
> compression efficiency
> Thank you for your hints
> Best Regards
> Son

Son,

I'm not sure I understand where you're headed ;)  Are you trying to make
a better cabac which is still standards-compliant (i.e. can be decoded
by a compliant decoder), or a modified cabac which requires a modified
decoder? 

If the first, how?  Perhaps you could post a patch?  I'm sure people
will be much better able to help get it working if they could look at
the code.  This sounds like something that can improve the efficiency of
x264 at very low bitrates or with certain types of material, so there
would be a fair amount of interest.

If the second, perhaps you may be interested in range encoding instead?
(free code here: http://www.compressconsult.com/rangecoder/ )

Regards,
--Alex

-- 
This is the x264-devel mailing-list
To unsubscribe, go to: http://developers.videolan.org/lists.html