What is motion compensated prediction, and why is it a pain?
A. MCP in the decoder can be thought of as having four stages: 1. Motion vector computation 2. Prediction retrieval various predictions are 16×16, 16×8, 8×4, 8×8 plus any half-pel overhead (e.g. 17×16, 17×17, etc). 3. Filtering 3.1 Forming half-pel predictions through bi-linear interpolation. 3.2 Averaging two predictions together (B macroblocks, Dual Prime) 4. Combination and ordering 4.1 combining 1 or 2 predictions from stage three into upper and lower halves (16 x 8, field in frame) 4.2 interleaving or grouping together odd and even lines in frame picture predictions. The final, combined prediction is always a 16×16 block of luminance and 8×8 block of chrominance, just like we experience in MPEG-1. A single motion vector can be associated with each source, hence a macroblock can have as many as 4 motion vectors.
