Page 231 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

208 Chapter 9. Error-Resilience Video Coding Techniques

and can be handled relatively easily. However, such errors can sometimes be
catastrophic, as, for example, in the case of errors in vital header information
(e.g., frame size, and quantizer step size).

9.4.2 Spatial Error Propagation

This is mainly due to two mechanisms:
1. Errors in VLC Coded Data: If an error converts a VLC codeword into
an invalid codeword or into a valid codeword of a di erent length, then
this causes loss of bitstream synchronization. This can occur in two
forms [177]:

(a) Loss of Codeword Synchronization: In this case an error causes the
decoder to decode a codeword of the wrong length. As a result, the
next codeword will be decoded in the wrong position and all fol-
lowing codewords may be a ected. This e ect is usually temporary,
and the decoder eventually regains codeword synchronization [178].
(b) Loss of Coe$cient Synchronization: The second form of loss of
synchronization is the loss of coe$cient synchronization. Even
when codeword synchronization is regained, the decoder will be
decoding coe$cients that have no meaning without the previous,
lost coe$cients. For example, in run-length encoding, if an incor-
rect run-length has been decoded, then all the following data will
be misplaced even if it is decoded correctly. Since this form of loss
of synchronization usually causes data to be misplaced, it is also
referred to as loss of positional synchronization.
2. Errors in Predictively Coded Data: The second mechanism that causes
spatial error propagation is the loss of predictively coded data. For
example, a motion vector is usually predictively coded with reference
to one or more previous motion vectors. If those previous vectors are
in error, then the prediction will be wrong and the errors will propagate
to the current motion vector, and so on.

9.4.3 Temporal Error Propagation
This is due mainly to the use of motion compensated prediction (or any other
form of predictive coding in the temporal dimension). As already described,
in motion-compensated prediction, parts of the current frame are copied (or
motion compensated) from a reference frame. If the copied reference parts
already contain errors, then those errors will also occur in (i.e., propagate to)
the current frame.