Survey of Dirac: A Wavelet Based Video Codec for Multiparty Video Conferencing and Broadcasting - Intelligent Video Event Analysis and Understanding - page 229

Splitting level 0: no split, a single MV per reference frame for the MB;
Splitting level 1: split into four sub-macroblocks (sub-MBs), each a 2x2 array of
blocks, one MV per reference frame per sub-MB;
Splitting level 2: split into the 16 constituent blocks, each have one MB.
Fig. 16 MacroBlock Splitting Modes
There are four prediction modes available:
INTRA: intra coded, predicted by DC value;
REF1_ONLY: only predict from the first reference;
REF2_ONLY: only predict from the second reference (if one exists);
REF1AND2: bi-directional prediction, predicted from both first and second
reference frame.
In motion estimation, an overall cost for each MB is computed, and compared for
each legal combination of these parameters. This is a tricky operation, and has a
very significant effect on performance. The decisions interact very heavily with
those made in coding the wavelet coefficients of the resulting residuals, and the
best results probably depend on picture material, bit rate, the block size and its
relationship to the size of the video frames, and the degree of perceptual weighting
used in selecting quantisers for wavelet coefficients.
5.3.3 Overlapped Block-Based Motion Compensation
Motion compensation in Dirac uses Overlapped Block-Based Motion
Compensation (OBMC) to avoid block-edge artefacts which would be expensive
to code using wavelets. Pretty much any size blocks can be used, with any degree
of overlap selected: this is configurable at the encoder and transmitted to the
decoder. One issue is that there should be an exact number of macro blocks
horizontally and vertically, where a macro block is a 4x4 set of blocks. This is
achieved by padding the data. Further padding may also be needed because after
motion compensation the wavelet transform is applied, which has its own
requirements for divisibility.
Dirac's OBMC scheme is based on a separable linear ramp mask. This acts as a
weight function on the predicting block. Given a pixel p=p(x,y,t) in frame t, p may
fall within only one block or in up to four if it lies at the corner of a block (see the
figure below).