Page 276 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 10.5. Temporal Error Concealment for Multiple-Reference 253

Table 10.6: Spatial-temporal recovery for QSIF TABLE TENNIS with M =10, QP =10, skip = 3, and
a macroblockerror rate of 30%
Spatial-components recovery

ZR AV BM MFI
Temporal- TR 19.62 19.57 20.06 20.46
component AV 19.54 19.57 20.02 20.40
recovery BM 19.68 19.87 20.09 20.58
MFI 19.55 19.74 20.00 20.40

candidate concealments, where each candidate concealment is provided using
a di,erent recovered motion vector. This is very similar to multihypothesis
motion compensation [106]. Thus, it is termed multihypothesis temporal er-
ror concealment.
In this subsection a multihypothesis temporal concealment technique to be
used with long-term memory motion-compensated prediction is presented. In
this case, the candidate concealments are taken from di,erent reference frames.
The details of this technique are as follows. The spatial components are $rst
recovered using MFI (as suggested in Section 10.5.2). However, instead of
recovering a single temporal component, all four neighboring temporal com-
ponents are utilized. Combined with the recovered spatial components, each
neighboring temporal component provides a candidate concealment from the
corresponding reference frame. The four candidate concealments are then av-
eraged and used to conceal the damaged blockin the current frame. In other
words, a damaged pel (x; y) in the current frame f c is concealed as follows:

1
4
ˆ
ˆ
ˆ
f (x; y)= 4 i=1 f r (x + d x (x; y);y + d y (x; y);d t i ); (10.13)
c
where f r (·; ·;d t ) refers to reference frame d t in the multiframe memory,
ˆ
ˆ
(d x (x; y); d y (x; y)) are the spatial components recovered at pel (x; y) using
, i =1;:::; 4 are the temporal components of the four neighboring
MFI, and d t i
vectors. In what follows, this approach is designated as MFI-MH.
Figures 10.17, 10.18, and 10.19 compare the performance of the MFI-MH
technique to that of MFI-BM (which is the best combination, as suggested in
Section 10.5.3) and also to that of ZR-ZR (which is the simplest and most
commonly used combination). The $gures con$rm the superior performance
of the suggested combination, MFI-BM, compared to the most commonly used
combination, ZR-ZR. In addition, the $gures show that further improvements
can be achieved using the multihypothesis MFI-MH technique.