Random real life sample, with entropy coding: averaged 6 bits per non-zero mv component in hpel, 7 bits in qpel.
Quote:
Originally Posted by Stux
2) With DCT based codecs the increased motion estimation accuracy means the DCT noise which varies between different DCT implementations becomes a significant part of the compensation. This results in qpel based artifacts like waterfalls/blurring on image decoding when using different "compliant" dcts.
|
Increased motion precision should reduce the number of nonzero dct coefficients, which should reduce the error from idct mismatch. (The same reason quant2 may look better than quant1 when decoded with the wrong idct.)
Otoh, what might exacerbate idct mismatch is that qpel uses an 8tap sharpening filter (and thus sharpens/spreads the error in addition to the content) while hpel is bilinear. But that's just mpeg4's specific versions of qpel and hpel; there's no reason you couldn't have a bilinear qpel or a 8tap hpel in some other format.