Re: Improving H263 quantization

Rounding to the nearest gives you less error and more data to be 
encoded, so unless the benefits overweight the loss, it's not better at all.

Do some experiments and check if it's better (keep trellis off though).

In general, inter quantization rounding is almost irrelevant because 
trellis quantization, which should be active for majority of encodes, 
gives the perfect (most RDO efficient) set of coefficients anyway.

Adept wrote:
> Hi ALL!
> 
> Let's speak about H263 quantization. Why rounding to down?
> IMHO, rounding to the nearest is better. Because of 
> reducing
> of quantization error two times.
> 
> For example. Required to change 4 identical lines in 
> quant_h263_intra_c
> and quant_h263_inter_c from
> 
> acLevel = (acLevel * mult) >> SCALEBITS;
> 
> to
> 
> acLevel = (acLevel * mult + 32768) >> SCALEBITS;
> 
> Because 32768=2^(SCALEBITS-1).
> 
> ---
> Professional hosting for everyone - http://www.host.ru
> _______________________________________________
> XviD-devel mailing list
> [email protected]
> http://list.xvid.org/mailman/listinfo/xvid-devel
> 
>