Mapping GOPS in an Improved DVC to H.264 Video Transcoder - Advanced Concepts for Intelligent Vision Systems - page 25

studies about this topic but there are mainly two major approaches: hash-based
motion estimation and Motion Compensated Temporal Interpolation (MCTI). In
particular, VISNET-II codec employs the latter one. In Figure 3 the first step of SI
generation is shown, which consists in matching each forward frame MarcoBlock
(MB) with a backward frame MB inside the search area. This matching takes all the
possibilities into account and chooses the lowest residual MB. Notice that DVC works
with 16x16 partitions to generate the SI (subpartitions are not used) and the search
area is defined by a window 32 pixel length. Through this process a MV is obtained
for each MB which quantifies the displacement between both MBs, and the middle of
this MV represents the displacement for the MB interpolated. The complete SI
estimation procedure is detailed in [9].
Fig. 3. First step of SI generation process
The present approach proposes to reuse these MVs calculated by WZ algorithm to
improve the transcoding process for every WZ GOP to H.264 GOP I11P. Figure 4
represents the transcoding from a WZ GOP 2 to a H.264 GOP I11P. The first K frame
is passed to an I-frame without any conversion, as was shown in Figure 2. On the
other hand, for every WZ frame a SI is estimated and one MV for each MB. This is
shown in the top row where V 0-2 represents the MVs calculated between K 0 and K 2 to
estimate SI 1 for WZ 1 and so on. In other words, V 0-2 in Figure 4 corresponds with MV
in Figure 3. Each MV is divided into two halves and it is applied in H.264 encoding
process to accelerate it in the way described in section 3.2. This part is shown in the
second row.
Fig. 4. Mapping from DVC GOP 2 to H.264 GOP I11P