order of magnitude as the number of bits that have otherwise been
spent for intra information. However, this mechanism does not add
any delay to the system.
Again, the complete parameter set establishment is performed through
control protocol means.
12.5. Video Telephony or Streaming with FUs and Forward Error
Correction
This scheme has been implemented and has been shown to provide good
performance, especially at higher packet loss rates [19].
The most efficient means to combat packet losses for scenarios where
retransmissions are not applicable is forward error correction (FEC).
Although application layer, end-to-end use of FEC is often less
efficient than an FEC-based protection of individual links
(especially when links of different characteristics are in the
transmission path), application layer, end-to-end FEC is unavoidable
in some scenarios. RFC 2733 [18] provides means to use generic,
application layer, end-to-end FEC in packet-loss environments. A
binary forward error correcting code is generated by applying the XOR
operation to the bits at the same bit position in different packets.
The binary code can be specified by the parameters (n,k) in which k
is the number of information packets used in the connection and n is
the total number of packets generated for k information packets;
i.e., n-k parity packets are generated for k information packets.
When a code is used with parameters (n,k) within the RFC 2733
framework, the following properties are well known:
a) If applied over one RTP packet, RFC 2733 provides only packet
repetition.
b) RFC 2733 is most bit rate efficient if XOR-connected packets have
equal length.
c) At the same packet loss probability p and for a fixed k, the
greater the value of n is, the smaller the residual error
probability becomes. For example, for a packet loss probability
of 10%, k=1, and n=2, the residual error probability is about 1%,
whereas for n=3, the residual error probability is about 0.1%.
d) At the same packet loss probability p and for a fixed code rate
k/n, the greater the value of n is, the smaller the residual error
probability becomes. For example, at a packet loss probability of
p=10%, k=1 and n=2, the residual error rate is about 1%, whereas
for an extended Golay code with k=12 and n=24, the residual error
rate is about 0.01%.
For applying RFC 2733 in combination with H.264 baseline coded video
without using FUs, several options might be considered:
1) The video encoder produces NAL units for which each video frame is
coded in a single slice. Applying FEC, one could use a simple
code; e.g., (n=2, k=1). That is, each NAL unit would basically
just be repeated. The disadvantage is obviously the bad code
performance according to d), above, and the low flexibility, as
only (n, k=1) codes can be used.
2) The video encoder produces NAL units for which each video frame is
encoded in one or more consecutive slices. Applying FEC, one
could use a better code, e.g., (n=24, k=12), over a sequence of
NAL units. Depending on the number of RTP packets per frame, a
loss may introduce a significant delay, which is reduced when more
RTP packets are used per frame. Packets of completely different
length might also be connected, which decreases bit rate
efficiency according to b), above. However, with some care and
for slices of 1kb or larger, similar length (100-200 bytes
difference) may be produced, which will not lower the bit
efficiency catastrophically.
3) The video encoder produces NAL units, for which a certain frame
contains k slices of possibly almost equal length. Then, applying
FEC, a better code, e.g., (n=24, k=12), can be used over the
sequence of NAL units for each frame. The delay compared to that
of 2), above, may be reduced, but several disadvantages are
obvious. First, the coding efficiency of the encoded video is
lowered significantly, as slice-structured coding reduces intra-
frame prediction and additional slice overhead is necessary.
Second, pre-encoded content or, when operating over a gateway, the
video is usually not appropriately coded with k slices such that
FEC can be applied. Finally, the encoding of video producing k
slices of equal length is not straightforward and might require
more than one encoding pass.
Many of the mentioned disadvantages can be avoided by applying FUs in
combination with FEC. Each NAL unit can be split into any number of
FUs of basically equal length; therefore, FEC with a reasonable k and
n can be applied, even if the encoder made no effort to produce
slices of equal length. For example, a coded slice NAL unit
containing an entire frame can be split to k FUs, and a parity check
code (n=k+1, k) can be applied. However, this has the disadvantage
that unless all created fragments can be recovered, the whole slice
will be lost. Thus a larger section is lost than would be if the
frame had been split into several slices.
The presented technique makes it possible to achieve good
transmission error tolerance, even if no additional source coding
layer redundancy (such as periodic intra frames) is present.
Consequently, the same coded video sequence can be used to achieve
the maximum compression efficiency and quality over error-free
transmission and for transmission over error-prone networks.
Furthermore, the technique allows the application of FEC to pre-
encoded sequences without adding delay. In this case, pre-encoded
sequences that are not encoded for error-prone networks can still be
transmitted almost reliably without adding extensive delays. In
addition, FUs of equal length result in a bit rate efficient use of
RFC 2733.
If the error probability depends on the length of the transmitted
packet (e.g., in case of mobile transmission [14]), the benefits of
applying FUs with FEC are even more obvious. Basically, the
flexibility of the size of FUs allows appropriate FEC to be applied
for each NAL unit and unequal error protection of NAL units.
When FUs and FEC are used, the incurred overhead is substantial but
is in the same order of magnitude as the number of bits that have to
be spent for intra-coded macroblocks if no FEC is applied. In [19],
it was shown that the overall performance of the FEC-based approach
enhanced quality when using the same error rate and same overall bit
rate, including the overhead.
12.6. Low Bit-Rate Streaming
This scheme has been implemented with H.263 and non-standard RTP
packetization and has given good results [20]. There is no technical
reason why similarly good results could not be achievable with H.264.
In today’s Internet streaming, some of the offered bit rates are
relatively low in order to allow terminals with dial-up modems to
access the content. In wired IP networks, relatively large packets,
say 500 - 1500 bytes, are preferred to smaller and more frequently
occurring packets in order to reduce network congestion. Moreover,
use of large packets decreases the amount of RTP/UDP/IP header
overhead. For low bit-rate video, the use of large packets means
that sometimes up to few pictures should be encapsulated in one
packet.
However, loss of a packet including many coded pictures would have
drastic consequences for visual quality, as there is practically no
other way to conceal a loss of an entire picture than to repeat the
previous one. One way to construct relatively large packets and
maintain possibilities for successful loss concealment is to
construct MTAPs that contain interleaved slices from several
pictures. An MTAP should not contain spatially adjacent slices from
the same picture or spatially overlapping slices from any picture.
If a packet is lost, it is likely that a lost slice is surrounded by
spatially adjacent slices of the same picture and spatially
corresponding slices of the temporally previous and succeeding
pictures. Consequently, concealment of the lost slice is likely to
be relatively successful.
12.7. Robust Packet Scheduling in Video Streaming
Robust packet scheduling has been implemented with MPEG-4 Part 2 and
simulated in a wireless streaming environment [21]. There is no
technical reason why similar or better results could not be
achievable with H.264.
Streaming clients typically have a receiver buffer that is capable of
storing a relatively large amount of data. Initially, when a
streaming session is established, a client does not start playing the
stream back immediately. Rather, it typically buffers the incoming
data for a few seconds. This buffering helps maintain continuous
playback, as, in case of occasional increased transmission delays or
network throughput drops, the client can decode and play buffered
data. Otherwise, without initial buffering, the client has to freeze
the display, stop decoding, and wait for incoming data. The
buffering is also necessary for either automatic or selective
retransmission in any protocol level. If any part of a picture is
lost, a retransmission mechanism may be used to resend the lost data.
If the retransmitted data is received before its scheduled decoding
or playback time, the loss is recovered perfectly. Coded pictures
can be ranked according to their importance in the subjective quality
of the decoded sequence. For example, non-reference pictures, such
as conventional B pictures, are subjectively least important, as
their absence does not affect decoding of any other pictures. In
addition to non-reference pictures, the ITU-T H.264 | ISO/IEC
14496-10 standard includes a temporal scalability method called sub-
sequences [22]. Subjective ranking can also be made on coded slice
data partition or slice group basis. Coded slices and coded slice
data partitions that are subjectively the most important can be sent
earlier than their decoding order indicates, whereas coded slices and
coded slice data partitions that are subjectively the least important
can be sent later than their natural coding order indicates.
Consequently, any retransmitted parts of the most important slices
and coded slice data partitions are more likely to be received before
their scheduled decoding or playback time compared to the least
important slices and slice data partitions.
13. Informative Appendix: Rationale for Decoding Order Number
13.1. Introduction
The Decoding Order Number (DON) concept was introduced mainly to
enable efficient multi-picture slice interleaving (see section 12.6)
and robust packet scheduling (see section 12.7). In both of these
applications, NAL units are transmitted out of decoding order. DON
indicates the decoding order of NAL units and should be used in the
receiver to recover the decoding order. Example use cases for
efficient multi-picture slice interleaving and for robust packet
scheduling are given in sections 13.2 and 13.3, respectively.
Section 13.4 describes the benefits of the DON concept in error
resiliency achieved by redundant coded pictures. Section 13.5
summarizes considered alternatives to DON and justifies why DON was
chosen to this RTP payload specification.
13.2. Example of Multi-Picture Slice Interleaving
An example of multi-picture slice interleaving follows. A subset of
a coded video sequence is depicted below in output order. R denotes
a reference picture, N denotes a non-reference picture, and the
number indicates a relative output time.
... R1 N2 R3 N4 R5 ...
The decoding order of these pictures from left to right is as
follows:
... R1 R3 N2 R5 N4 ...
The NAL units of pictures R1, R3, N2, R5, and N4 are marked with a
DON equal to 1, 2, 3, 4, and 5, respectively.
Each reference picture consists of three slice groups that are
scattered as follows (a number denotes the slice group number for
each macroblock in a QCIF frame):
0 1 2 0 1 2 0 1 2 0 1
2 0 1 2 0 1 2 0 1 2 0
1 2 0 1 2 0 1 2 0 1 2
0 1 2 0 1 2 0 1 2 0 1
2 0 1 2 0 1 2 0 1 2 0
1 2 0 1 2 0 1 2 0 1 2
0 1 2 0 1 2 0 1 2 0 1
2 0 1 2 0 1 2 0 1 2 0
1 2 0 1 2 0 1 2 0 1 2
For the sake of simplicity, we assume that all the macroblocks of a
slice group are included in one slice. Three MTAPs are constructed
from three consecutive reference pictures so that each MTAP contains
three aggregation units, each of which contains all the macroblocks
from one slice group. The first MTAP contains slice group 0 of
picture R1, slice group 1 of picture R3, and slice group 2 of
picture R5. The second MTAP contains slice group 1 of picture R1,
slice group 2 of picture R3, and slice group 0 of picture R5. The
third MTAP contains slice group 2 of picture R1, slice group 0 of
picture R3, and slice group 1 of picture R5. Each non-reference
picture is encapsulated into an STAP-B.
Consequently, the transmission order of NAL units is the following:
R1, slice group 0, DON 1, carried in MTAP, RTP SN: N
R3, slice group 1, DON 2, carried in MTAP, RTP SN: N
R5, slice group 2, DON 4, carried in MTAP, RTP SN: N
R1, slice group 1, DON 1, carried in MTAP, RTP SN: N+1
R3, slice group 2, DON 2, carried in MTAP, RTP SN: N+1
R5, slice group 0, DON 4, carried in MTAP, RTP SN: N+1
R1, slice group 2, DON 1, carried in MTAP, RTP SN: N+2
R3, slice group 1, DON 2, carried in MTAP, RTP SN: N+2
R5, slice group 0, DON 4, carried in MTAP, RTP SN: N+2
N2, DON 3, carried in STAP-B, RTP SN: N+3
N4, DON 5, carried in STAP-B, RTP SN: N+4
The receiver is able to organize the NAL units back in decoding order
based on the value of DON associated with each NAL unit.
If one of the MTAPs is lost, the spatially adjacent and temporally
co-located macroblocks are received and can be used to conceal the
loss efficiently. If one of the STAPs is lost, the effect of the
loss does not propagate temporally.
13.3. Example of Robust Packet Scheduling
An example of robust packet scheduling follows. The communication
system used in the example consists of the following components in
the order that the video is processed from source to sink:
o camera and capturing
o pre-encoding buffer
o encoder
o encoded picture buffer
o transmitter
o transmission channel
o receiver
o receiver buffer
o decoder
o decoded picture buffer
o display
The video communication system used in the example operates as
follows. Note that processing of the video stream happens gradually
and at the same time in all components of the system. The source
video sequence is shot and captured to a pre-encoding buffer. The
pre-encoding buffer can be used to order pictures from sampling order
to encoding order or to analyze multiple uncompressed frames for bit
rate control purposes, for example. In some cases, the pre-encoding
buffer may not exist; instead, the sampled pictures are encoded right
away. The encoder encodes pictures from the pre-encoding buffer and
stores the output; i.e., coded pictures, to the encoded picture
buffer. The transmitter encapsulates the coded pictures from the
encoded picture buffer to transmission packets and sends them to a
receiver through a transmission channel. The receiver stores the
received packets to the receiver buffer. The receiver buffering
process typically includes buffering for transmission delay jitter.
The receiver buffer can also be used to recover correct decoding
order of coded data. The decoder reads coded data from the receiver
buffer and produces decoded pictures as output into the decoded
picture buffer. The decoded picture buffer is used to recover the
output (or display) order of pictures. Finally, pictures are
displayed.
In the following example figures, I denotes an IDR picture, R denotes
a reference picture, N denotes a non-reference picture, and the
number after I, R, or N indicates the sampling time relative to the
previous IDR picture in decoding order. Values below the sequence of
pictures indicate scaled system clock timestamps. The system clock
is initialized arbitrarily in this example, and time runs from left
to right. Each I, R, and N picture is mapped into the same timeline
compared to the previous processing step, if any, assuming that
encoding, transmission, and decoding take no time. Thus, events
happening at the same time are located in the same column throughout
all example figures.
A subset of a sequence of coded pictures is depicted below in
sampling order.
... N58 N59 I00 N01 N02 R03 N04 N05 R06 ... N58 N59 I00 N01 ...
... --|---|---|---|---|---|---|---|---|- ... -|---|---|---|- ...
... 58 59 60 61 62 63 64 65 66 ... 128 129 130 131 ...
Figure 16. Sequence of pictures in sampling order
The sampled pictures are buffered in the pre-encoding buffer to
arrange them in encoding order. In this example, we assume that the
non-reference pictures are predicted from both the previous and the
next reference picture in output order, except for the non-reference
pictures immediately preceding an IDR picture, which are predicted
only from the previous reference picture in output order. Thus, the
pre-encoding buffer has to contain at least two pictures, and the
buffering causes a delay of two picture intervals. The output of the
pre-encoding buffering process and the encoding (and decoding) order
of the pictures are as follows:
... N58 N59 I00 R03 N01 N02 R06 N04 N05 ...
... -|---|---|---|---|---|---|---|---|- ...
... 60 61 62 63 64 65 66 67 68 ...
Figure 17. Re-ordered pictures in the pre-encoding buffer
The encoder or the transmitter can set the value of DON for each
picture to a value of DON for the previous picture in decoding order
plus one.
For the sake of simplicity, let us assume that:
o the frame rate of the sequence is constant,
o each picture consists of only one slice,
o each slice is encapsulated in a single NAL unit packet,
o there is no transmission delay, and
o pictures are transmitted at constant intervals (that is, 1 / frame
rate).
When pictures are transmitted in decoding order, they are received as
follows:
... N58 N59 I00 R03 N01 N02 R06 N04 N05 ...
... -|---|---|---|---|---|---|---|---|- ...
... 60 61 62 63 64 65 66 67 68 ...
Figure 18. Received pictures in decoding order
The OPTIONAL sprop-interleaving-depth MIME type parameter is set to
0, as the transmission (or reception) order is identical to the
decoding order.
The decoder has to buffer for one picture interval initially in its
decoded picture buffer to organize pictures from decoding order to
output order as depicted below:
... N58 N59 I00 N01 N02 R03 N04 N05 R06 ...
... -|---|---|---|---|---|---|---|---|- ...
... 61 62 63 64 65 66 67 68 69 ...
Figure 19. Output order
The amount of required initial buffering in the decoded picture
buffer can be signaled in the buffering period SEI message or with
the num_reorder_frames syntax element of H.264 video usability
information. num_reorder_frames indicates the maximum number of
frames, complementary field pairs, or non-paired fields that precede
any frame, complementary field pair, or non-paired field in the
sequence in decoding order and that follow it in output order. For
the sake of simplicity, we assume that num_reorder_frames is used to
indicate the initial buffer in the decoded picture buffer. In this
example, num_reorder_frames is equal to 1.
It can be observed that if the IDR picture I00 is lost during
transmission and a retransmission request is issued when the value of
the system clock is 62, there is one picture interval of time (until
the system clock reaches timestamp 63) to receive the retransmitted
IDR picture I00.
Let us then assume that IDR pictures are transmitted two frame
intervals earlier than their decoding position; i.e., the pictures
are transmitted as follows:
... I00 N58 N59 R03 N01 N02 R06 N04 N05 ...
... --|---|---|---|---|---|---|---|---|- ...
... 62 63 64 65 66 67 68 69 70 ...
Figure 20. Interleaving: Early IDR pictures in sending order
The OPTIONAL sprop-interleaving-depth MIME type parameter is set
equal to 1 according to its definition. (The value of sprop-
interleaving-depth in this example can be derived as follows:
Picture I00 is the only picture preceding picture N58 or N59 in
transmission order and following it in decoding order. Except for
pictures I00, N58, and N59, the transmission order is the same as the
decoding order of pictures. As a coded picture is encapsulated into
exactly one NAL unit, the value of sprop-interleaving-depth is equal
to the maximum number of pictures preceding any picture in
transmission order and following the picture in decoding order.)
The receiver buffering process contains two pictures at a time
according to the value of the sprop-interleaving-depth parameter and
orders pictures from the reception order to the correct decoding
order based on the value of DON associated with each picture. The
output of the receiver buffering process is as follows:
... N58 N59 I00 R03 N01 N02 R06 N04 N05 ...
... -|---|---|---|---|---|---|---|---|- ...
... 63 64 65 66 67 68 69 70 71 ...
Figure 21. Interleaving: Receiver buffer
Again, an initial buffering delay of one picture interval is needed
to organize pictures from decoding order to output order, as depicted
below:
... N58 N59 I00 N01 N02 R03 N04 N05 ...
... -|---|---|---|---|---|---|---|- ...
... 64 65 66 67 68 69 70 71 ...
Figure 22. Interleaving: Receiver buffer after reordering
Note that the maximum delay that IDR pictures can undergo during
transmission, including possible application, transport, or link
layer retransmission, is equal to three picture intervals. Thus, the
loss resiliency of IDR pictures is improved in systems supporting
retransmission compared to the case in which pictures were
transmitted in their decoding order.
13.4. Robust Transmission Scheduling of Redundant Coded Slices
A redundant coded picture is a coded representation of a picture or a
part of a picture that is not used in the decoding process if the