Graphics Reference
In-Depth Information
same RPS information is repeated for the pictures that share the same position in
the GOP structure.
To exploit the redundancy and reduce the overall bit cost, some of the RPS syntax
can be sent in the SPS and referred to from the slice header. The short-term picture
part of an RPS is coded relative to the POC value of the current picture. This makes
it possible to store multiple short-term RPS parts in a list in the SPS and only signal a
list index in the slice header. The list may contain one RPS for each picture position
in the GOP structure. The number of RPSs in the SPS would then be equal to the
GOP length, and each slice would only need to send a list index in order to signal
its short-term picture part.
Sending the RPS information for each GOP position may require relatively many
bits in the SPS, there are examples of several hundred bits for some GOP structures.
In order to save bits in the SPS, there is the option in HEVC to use RPS prediction.
This requires the RPSs to be sent in GOP decoding order and exploits the fact that
each RPS is similar to the previous one. Picture references may be removed from
one RPS to the next in decoding order, but no more than one new picture is possible
to add relative to the previous RPS. The RPS prediction mechanism in HEVC is
utilizing this property, and reductions of up to 50 % of the RPS bit count in the SPS
can be achieved relative to explicit signaling.
2.4.4
Reference Picture Lists
There are two types of sample prediction in HEVC, intra prediction and inter
prediction. Intra prediction does not include prediction from any reference picture,
only sample prediction using reconstructed samples of the same picture is allowed.
Inter prediction uses reference pictures where picture identifiers, called reference
indices, and motion vectors are used to specify what part of which reference picture
to use for prediction.
There are three slice types in HEVC. The intra (I) slice for which only intra
prediction is allowed. The predictive (P) slice which in addition to intra prediction
also allows inter prediction from one reference picture per block using one motion
vector and one reference index. This is called uni-prediction. Finally, there is the
bi-predictive (B) slice which in addition to intra and uni-prediction also allows
inter prediction using two motion vectors and two reference indices. This results
in two prediction blocks that are combined to form a final prediction block. Using
bi-prediction is generally more compression efficient than using uni-prediction but
the computational complexity is higher.
Note that the use of uni-prediction and bi-prediction is completely decoupled
from both the output order as well as from the number of reference pictures used.
Different blocks within the same P slice can reference different reference pictures,
and a block that uses bi-prediction can have both its motion vectors pointing to the
same picture.
