Real-Time Analysis of Human Body Parts and Gesture-Activity Recognition in 3D - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 166

Figure 26. Balanced partition.
Figure 27. Hardware efficient partition.
Another factor that needs to be considered is communication cost. The amount
of data that needs to be transferred between the video camera interface and
region, and between the region stage and contour, is significantly larger than the
data size exchanged among another stages. Therefore, we would prefer to
allocate a set consisting of video interface, region stage, and contour stage into
one processor. In the above partitioning, we comply with this rule. The above
discussion is limited to inter-stage parallelism. In the following, we will show how
the inter-frame parallelism can be taken into consideration. At the first, we
duplicate every processing stage. For example, if we want to process two frames
in parallel, we will have two copies of each processing stage. After this, we can
perform scheduling and get the corresponding architecture. Figure 28 shows a
five-processor partition with two consecutive frames processed in parallel.
Figure 29 shows the corresponding architecture for the partition.