Page 271 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

248 Chapter 10. Error Concealment Using Motion Field Interpolation

Then a set C = {d 1 ;:::; d 4 } of four candidates is formed from the recovered
ˆ ˆ
spatial components ( d x ; d y ) and the four temporal components of the neigh-
ˆ ˆ
ˆ ˆ
ˆ ˆ
boring blocks. In other words: d 1 =(d x ; d y ;l t ), d 2 =(d x ; d y ;r t ), d 3 =(d x ; d y ;t t ),
ˆ ˆ
and d 4 =(d x ; d y ;b t ). The BM technique is then used to recover the temporal
component by choosing from this set of candidates. Thus
ˆ
d = arg min SMD(d i ): (10.12)
d i ∈C
A multiple-reference rate-constrained H.263-like codec was used to generate
the results of this section. This codec uses full-pel full-search blockmatching
with macroblocks of 16 × 16 pels, a maximum allowed spatial displacement
of ± 15 pels, SAD as the distortion measure, restricted motion vectors, and
reconstructed reference frames. Motion vectors are coded using the median
predictor and the VLC table of the H.263 standard. The frame signal (in case
of INTRA) and the DFD signal (in case of INTER) are transform encoded
according to the H.263 standard. The codec uses rate-constrained motion esti-
mation and mode decision as de$ned in the high-complexity mode of TMN10.
The codec employs a sliding-window control to maintain a long-term memory
of size M =10 frames. Only the $rst frame is INTRA coded, and no INTRA
refresh is employed. A $xed quantization parameter of QP =10 is used. Errors
were introduced randomly on a macroblocklevel. Thus, an error rate of 20%
means that 20% of the macroblocks are damaged per frame. It is assumed that
the decoder uses an ideal error detection mechanism. All quoted results refer
to the luma components of sequences.

10.5.1 Temporal-Component Recovery
This set of experiments investigate the best technique for recovering the tem-
poral component d t of a damaged long-term motion vector. In this case, the
spatial recovery technique S, in the combination S-T, was kept constant at ZR,
whereas the temporal recovery technique T was varied over ZR, AV, BM, and
MFI. In other words, four S-T combinations were considered: ZR-ZR, ZR-AV,
ZR-BM, and ZR-MFI.
Figures 10.11, 10.12, and 10.13 show the results for the QSIF sequences
AKIYO,FOREMAN, and TABLE TENNIS, respectively. Part (a) of each $gure shows
the performance with a frame skip of 3 over a range of macroblock error rates,
whereas part (b) shows the performance with a macroblockerror rate of 20%
over a range of frame skips.
In general, the best temporal-component recovery is achieved by ZR and
BM (i.e., ZR-ZR and ZR-BM). The good performance of ZR is due to the
zero-biased distribution of the temporal components (Property 6:3:1:2). In other
words, the temporal component d t = 0 has the highest frequency of occurrence