Inverse quantization is sometimes denoted IQ, not to be confused with IQ colour difference components.

When a decoder reconstructs a P-picture, it is displayed; additionally, the picture is written into a reference frame so as to be available for subsequent predictions.

Each B-picture contains elements that are bipredicted from one or both reference frames. The encoder computes, compresses, and transmits residuals. The decoder reconstructs a B-picture, displays it, then discards it: No B-picture is used for prediction.

Each reference picture is associated with a full frame of storage. When a decoder reconstructs a reference field (an I-field or a P-field), half the lines of the reference framestore are written; the other half retains the contents of the previous reference field. After the first field of a field pair has been reconstructed, it is available as a predictor for the second field. (The first field of the previous reference frame is no longer available.)

Prediction

In Figure 16.1, on page 152, I sketched a naïve interpicture coding scheme. For any scene element that moves more than few pixels from one video frame to the next, the naïve scheme is liable to produce large interpicture difference values. Motion can be more effectively coded by having the encoder form motioncompensated predictions. The encoder also produces motion vectors; these are used to displace a region of a reference picture to improve the prediction of the current picture relative to an undisplaced prediction. The residuals are then compressed using DCT, quantized, and VLE-encoded.

At a decoder, predictions are formed from the reference picture(s), based upon the transmitted motion vectors and prediction modes. Residuals are recovered from the bitstream by VLE decoding, inverse quantization, and inverse DCT. Finally, the decoded residual is added to the prediction to form the reconstructed picture. If the decoder is reconstructing an I-picture or a P-picture, the reconstructed picture is written to the appropriate portion (or the entirety) of a reference frame.

The obvious way for an encoder to form forward interpicture differences is to subtract the current source picture from the reference picture. (The reference

A prediction region in a reference frame is rarely aligned to

a 16-luma-sample macroblock grid; it is not properly referred to as a macroblock. Some authors fail to make the distinction between macroblocks and prediction regions; other authors use the term prediction macroblocks for prediction regions.

In a closed GoP, no B-picture is permitted to use forward prediction to the I-picture that starts the next GoP. See the caption to Figure 16.5, on page 155.

picture would have been subject to motion-compen- sated interpolation, according to the encoder’s motion estimate.) Starting from an intra coded picture, the decoder would then accumulate interpicture differences. However, MPEG involves lossy compression: Both the I-picture starting point of a GoP and each set of decoded interpicture differences are subject to reconstruction errors. With the naïve scheme of computing interpicture differences, reconstruction errors would accumulate at the decoder. To alleviate this potential source of decoder error, the encoder incorporates a decoder. The interpicture difference is formed by subtracting the current source picture from the previous reference picture as a decoder will reconstruct it. Reconstruction errors are thereby brought “inside the loop,” and are prevented from accumulating.

The prediction model used by MPEG-2 is blockwise translation of 16× 16 blocks of luma samples (along with the associated chroma samples): A macroblock of the current picture is predicted from a like-sized region of a reconstructed reference picture. The choice of

16× 16 region size was a compromise between the desire for a large region (to effectively exploit spatial coherence, and to amortize motion vector overhead across a fairly large number of samples), and a small region (to efficiently code small scene elements in motion).

Macroblocks in a P-picture are typically forwardpredicted. However, an encoder can decide that a particular macroblock is best intracoded (that is, not predicted at all). Macroblocks in a B-picture are typically predicted as averages of motion-compensated past and future reference pictures – that is, they are ordinarily bidirectionally predicted. However, an encoder can decide that a particular macroblock in a B-picture is best intracoded, or unidirectionally predicted using either forward or backward prediction. Table 47.7 at the top of the facing page indicates the four macroblock types. The macroblock types allowed in any picture are restricted by the declared picture type, as indicated in Table 47.8.

Each nonintra macroblock in an interlaced sequence can be predicted either by frame prediction (typically chosen by the encoder when there is little motion