[XviD-devel] Xvid v's H.264/AVC

[Daniel Larkin]

> I'm not the Qpel expert, but IIRC, MPEG-2 uses bilinear subpel
> interpolation and MPEG-4 does pure bilinear filtering only in non-Qpel
> mode. In QPel mode, it uses some 6-tap filtering process for the halfpel
> positions (and bilinear on top of that for the Qpels) or so.
> So, one explanation for MPEG-4s rather poor performance in Qpel
> mode would
> therefore be that MPEG-2 images are more smoothed/blurred level than
> MPEG-4 can directly reproduce with it's sharpness preserving filter. It
> therefore needs to spend additional texture bits to just simulate the
> smoothness of MPEG-2.
> Also, if the input doesn't contain real Qpel motion (e.g. because they
> are quantized away due to MPEG-2s halfpel vectors), the qpel won't help
> anything, and just costs extra bits because vectors become longer (in
> units).
> But all this is just speculation. To my knowledge, there are no real
> studies available on reencoding MPEG-2 to MPEG-4.
>
> chl
>

Adding to the speculation, if thats ok :-), this would be my understanding,
which I think is the same(??) as yours Christoph. Please correct me if I'm
incorrect....

Is the reason for this problem (i.e. the lack of improvement of QPel for
MPEG2 source) not simply because the source has undergone quantization and
thus there are more homogenous pixels in the MPEG-2 decoded source compared
to the YUV raw source and in this situation a follow up MPEG-4 encode the
Qpel interpolation wont as "beneficial" because you are generating "poor"
interpolated pixels (relative to the YUV source). These quarter position
pels aren't going to be used because the shorter motion vector to full/half
pels will be chosen??? Of course that assumption is based on a situation
where you've effectively quantized away the quarter pel motion. **The
mismatch in the two resolution ME doesn't contribute to the problem**

For example imagine for a moment a hypothetical codec where there wasn't
anyway transform or quantization, simply ME and entropy coding. Regardless
of whether this encoder used full, half, quarter or eighth pel ME you expect
the decoder to regenerate a perfect reconstruction (relative to the original
YUV sample) since after all the ME is merely providing a prediction to the
entropy encoder. The difference from the ME Pel resolution effecting only
the filesize produced. If you then re-encoded (again with this imaginary
ME+entropy encoder) what appeared at the decoder of this imaginary codec,
this time choosing a different pel resolution you'd still expect a perfect
reconstruction when decoded. In this situation a mix-match of different
Pel'ed motion estimation doesn't effect quality. Sooooooo doesn't this point
to the fact any lack of performance benefit from MPEG2->MPEG4 using QPel is
down to the lossy/Quantization process of MPEG-2 and there is *no
performance degrade* coming from the mismatched half pel resolution in
MPEG-2 versus Quarter Pel in MPEG-4 ASP.