Game Development Reference
In-Depth Information
4.3.3 Multiple Hypothesis Prediction
Besides the conventional single hypothesis prediction mode in P frame, multi-
hypothesis techniques are added for more efficient prediction, including advanced
skip/direct mode, temporal multi-hypothesis mode, and spatial multi-hypothesis
mode (directional multi-hypothesis prediction, DMH) (Ling et al. 2013 ; Ling and
Yu 2014 ; Kim et al. 2013 ), called F picture in AVS2.
In F picture, an advanced skip/direct mode is defined using a competitive motion
derivation mechanism (Shao and Yu 2013 ). Two derivation methods, one of which
is temporal and the other is spatial, are used. For temporal derivation, the prediction
block is obtained by an average of the prediction blocks indicated by theMVP and the
scaled MV in a second reference. The second reference is specified by the reference
index transmitted in the bitstream. For spatial derivation, the prediction block may
be obtained one or two prediction blocks specified by the motion copied from its
spatial neighboring blocks. As illustrated in Fig. 4.11 , the neighboring blocks are
searched in a pre-defined order F, G, C, A, B, D, and the selected neighboring block
is signaled in the bitstream.
Temporal multi-hypothesis mode combines two predictors on the pre-defined
temporal direction, while spatial multi-hypothesis mode combines two predictors
along the pre-defined spatial direction. For temporal multi-hypothesis prediction,
asshowninFig. 4.12 , one predictor ref _ blk 1 is generated with the best motion
vector MV and reference frame ref 1 obtained from motion estimation. Then this
motion vector is linearly scaled to a second reference, specified by the reference
index transmitted in the bitstream, to generate another predictor ref _ blk 2.
In DMH mode, as specified in Fig. 4.13 , seed predictors are generated from one
initial MV. The number of seed predictors is restricted to 8 and finally one is selected
and coded into bitstream. If one seed predictor is selected for combined prediction,
for example “Mode 1”, then two MVs are generated from the initial MV, by adding
or subtracting the same delta MV indicated from the mode.
Fig. 4.10 Progressive motion vector resolution, a MVP at integer or half pixel position, b MVP at
1/4 pixel position
 
Search WWH ::




Custom Search