Page 77 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

54 Chapter 3. Video Coding: Standards

3.4.4 Decoding Process
3.4.4.1 Motion Vector Decoding
For each TRANSMITTED INTER MB, the decoder calculates the same
motion vector predictor MVP used at the encoder and adds it to the
decoded motion vector di1erence MVD to obtain the decoded motion vec-
tor MV. The motion vector of a SKIPPED INTER MB is set to 0.

3.4.4.2 Motion Compensation

The decoded motion vector is used to compensate the four Y blocks in the

MB. Motion vectors for both C and C blocks are derived by dividing the
B
R
component values of the decoded motion vector by 2. The resulting quarter-
pel resolution components are modi/ed toward the nearest half-pel resolution
(both 0.25 and 0.75 are rounded to 0.5). If motion compensation requires
accessing half-pel positions, then bilinear interpolation is used to calculate the
pel values at those positions.
3.4.4.3 Inverse Quantization
As already discussed, quantization is achieved by dividing the transform coef-
/cient by a quantization step size and rounding the result (refer to Equations
4
(2.8) and (2.10)). Inverse quantization is the process of reconstructing an ap-
proximation of the original coe,cient by multiplying the quantized coe,cient
by the same step size (refer to Equations (2.9) and (2.11)). The reconstructed
coe,cients are then clipped to the range [−2048; +2047] and inverse zigzag
scanned to put them in an 8 × 8 block.

3.4.4.4 Inverse Transform

The reconstructed block of coe,cients is processed by a separable 2-D 8 × 8
inverse DCT. The arithmetic procedures for computing the inverse DCT are
not de/ned by the standard, but should meet a de/ned error tolerance.

3.4.4.5 Reconstruction of Blocks

For INTRA blocks, the reconstructed block is equal to the result of the inverse
DCT. For INTER blocks, the reconstructed block is formed by summing the
motion-compensated prediction and the result of the inverse DCT. The recon-
structed values are clipped to the range [0; 255].

4 It should be emphasised that the term inverse here does not mean that quantization is a
reversible process. Quantization is irreversible since rounding leads to loss of information.