[XviD-devel] MBTransQuant2

Christoph Lampert xvid-devel@xvid.org
Wed, 4 Sep 2002 10:34:01 +0200 (CEST)


On Wed, 4 Sep 2002, daniel smith wrote:
> it still needs field-based motion estimation:
> 
> for each field of a macroblock, search both top and bottom fields of reference frame.  if the result is good enough, send 2 mv's instead of 1 for each field, and indiciate for each of the mv's which field it actually references.
> 
> this changes the way mv's are predicted (if reference block is field-predicted, the predictor mv for that block is the rounded average of the two mv's), the way the mv is stored (half-resolution y component, but written to the bitstream differently from memory), u/v mv calculation (i believe it's just halved with rounding from the top/bottom vector), compensation (we need field-aware 16x8 and 8x4 transfer functions) and of course a 16x8 ME search.  also interpolation is done vertically between field lines instead of frame, but that's just a stride issue.  field-based ME for b-frames seems more complex.

This sounds like quite a lot of work... :-(

gruel