Internet Engineering Task Force                    G. Liebl,
                                                   T.Stockhammer
Internet Draft                                       LNT, Munich Univ.
                                                        of Technology
Document: draft-ietf-avt-uxp-01.txt
               November 2001 draft-ietf-avt-uxp-02.txt
March 1, 2002                                      M. Wagner, J.Pandel,
                                                   W. Weng, G. Baese,
                                                   M. Nguyen, F. Burkert
Expires: May Sept. 1, 2002                               Siemens AG, Munich

 An RTP Payload Format for Erasure-Resilient Transmission of Progressive
                           Multimedia Streams

Status of this Memo

   This document is an Internet-Draft and is in full conformance with
      all provisions of Section 10 of RFC2026 [].

   Internet-Drafts are working documents of the Internet Engineering
   Task Force (IETF), its areas, and its working groups. Note that
   other groups may also distribute working documents as Internet-
   Drafts. Internet-Drafts are draft documents valid for a maximum of
   six months and may be updated, replaced, or obsoleted by other
   documents at any time. It is inappropriate to use Internet- Drafts
   as reference material or to cite them other than as "work in
   progress."
   The list of current Internet-Drafts can be accessed at
   http://www.ietf.org/ietf/1id-abstracts.txt
   The list of Internet-Draft Shadow Directories can be accessed at
   http://www.ietf.org/shadow.html.

1. Abstract

   This document specifies an efficient way to ensure erasure-resilient
   transmission of progressively encoded multimedia sources via RTP
   using Reed-Solomon codes. The level of erasure protection can be
   explicitly adapted to the importance of the respective parts in the
   source stream, thus allowing a graceful degradation of application
   quality with increasing packet loss rate on the network. Hence, this
   type of unequal erasure protection (UXP) schemes is intended to cope
   with the rapidly varying channel conditions on wireless access links
   to the Internet backbone. Nevertheless, backward compatibility to
   currently standardized non-progressive multimedia codecs is ensured,
   since equal erasure protection (EXP) represents a subset of generic
   UXP. By defining a comparably simple payload format, the proposed
   scheme can be easily integrated into the existing framework for RTP.

Liebl,Stockhammer,Wagner,Pandel,Weng,Baese,Nguyen,Burkert      [Page1]

2. Conventions used in this document

   The following terms are used throughout this document:

   1.) Message block: a higher layer transport unit (e.g. an IP
   packet), that enters/leaves the segmentation/reassembly stage at the
   interface to wireless data link layers.

   2.) Segment: denotes a link layer transport unit.

   3.) CRC: Cyclic Redundancy Check, usually added to transport units
   at the sender to detect the existence of erroneous bits in a
   transport unit at the receiver.

   4.) Segmentation/Reassembly Process: If the size of the transport
   units at the link layer is smaller than that at the upper layers,
   message blocks have to be split up into several parts, i.e.
   segments, which are then transmitted subsequently over the link. If
   nothing is lost, the original message block can be restored at the
   receiving entity (reassembly).

   5.) Quality-of-service: application-dependent criterion to define a
   certain desired operation point.

   6.) Codec: denotes a functional pair consisting of a source encoding
   unit at the sender and a corresponding source decoding unit at the
   receiver; usually standardized for different multimedia applications
   like audio or video.

   7.) Progressive source coding: results in successive blocks of
   (source-)encoded data (e.g. a single video or audio frame), each of
   which can be viewed as a bitstream of certain length, whose distinct
   elements are of different importance to the reconstruction process
   at the decoder. Elements are commonly ordered from highest to least
   importance, where the latter elements depend on the previous.

   8.) Reed-Solomon (RS) code: belongs to the class of linear nonbinary
   block codes, and is uniquely specified by the block length n, the
   number of parity symbols t, and the symbol alphabet.

   9.) n: is a variable, which denotes both the block length of a RS
   codeword, and the number of columns in a TB (see 16).

   10.) k: is a variable, which denotes the number of information
   symbols in a RS codeword.

   11.) t: is a variable, which denotes the number of parity symbols in
   a RS codeword.

   12.) Erasure: When a packet is lost during transmission, an erasure
   is said to have happened. Since the position of the erased packet in
   a sequence is usually known, a corresponding erasure marker can be
   set at the receiving entity.

   13.) Base layer: comprises the first and most important elements in
   a progressively encoded bitstream, without which all subsequent
   information is useless.

   14.) Enhancement layer: comprises one or more sets of the less
   important subsequent elements in a progressively encoded bitstream.
   A specific enhancement layer can be decoded, if and only if the base
   layer and all previous enhancement layer data (of higher importance)
   is available.

   15.) Info stream: denotes the final bitstream which has to be
   protected by the proposed UXP scheme. It usually consists of the
   (source-encoded) bitstream (progressive or not), which is already
   arranged according to a desired syntax (e.g. as specified in the
   respective RTP profile for the media codec in use).
   In any case, it is assumed that every info stream is already octet-
   aligned according to the standard procedures defined in the context
   of the used syntax specifications.

   16.) Transmission block (TB): denotes a memory array of L rows and n
   columns. Each row of a TB represents a RS codeword, whereas each
   column, together with the respective UXP header (see 33) in front,
   forms the payload of a single RTP packet.
   Each TB consists of at least two distinct transmission sub blocks
   (TSB, see 17): The first L_s rows belong to the signaling TSB,
   whereas the last L_d=(L-L_s) rows belong to one or more data TSB.

   17.) Transmission sub block (TSB): denotes a memory array of 0<l<L
   rows and n columns, which is a horizontal slice of a TB. Depending
   on whether the info byte positions are filled with descriptors (see
   28) or media data, the TSB is of type signaling or data,
   respectively.

   18.) L: is a variable, which denotes both the number of rows in a TB
   and the payload length (without UXP header) of an RTP packet in
   bytes.

   19.) Unequal erasure protection (UXP): denotes a specific strategy
   which varies the level of erasure protection across a TB according
   to a given redundancy profile.

   20.) Equal erasure protection (EXP): is a subset of UXP, for which
   the level of erasure protection is kept constant across a TB.

   21.) Redundancy profile: describes the size of the different erasure
   protection classes in a TB, i.e. the number of rows (codewords) per
   class.

   22.) Erasure protection class: contains a set of rows (codewords) of
   the TB with same erasure correction capability.

   23.) i: is a variable, which denotes the number of parity bytes for
   each row in erasure protection class i.

   24.) CA_i: is a variable, which denotes the set of rows contained in
   erasure protection class i.

   25.) A_i: is a variable, which denotes the total number of rows
   contained in erasure protection class i, i.e. the cardinality of
   CA_i.

   26.) T: is a variable, which denotes the number of parity bytes for
   each row in the highest erasure protection class (with respect to
   application data) in a TB.

   27.) AV: denotes the erasure protection vector of length (T+1) used
   to describe a certain redundancy profile.

   28.) DP: descriptor used for in-band signaling of the erasure
   protection vector.

   29.) SI: stuffing indicator, which contains the number of media
   stuffing symbols at the end of a data TSB (see 31).

   30.) Descriptor Stuffing: insertion of otherwise unused descriptor
   values (i.e. 0x00) at the end of the signaling TSB. Descriptor
   stuffing is performed, if the final sequence of descriptors and
   stuffing indicators for a valid redundancy profile is shorter than
   the space initially reserved for it in the signaling TSB.

   31.) Media Stuffing: insertion of additional symbols at the end of a
   data TSB. Media stuffing is performed, if the info stream (see 15)
   is shorter than the space reserved for it in the data TSB for a
   desired redundancy profile. Since the number of stuffing symbols is
   signaled in the respective SI, any byte value may be used (e.g.
   0x00).

   32.) Interleaver: performs the spreading of a codeword, i.e. a row
   in the TB, over n successive packets, such that the probability of
   an erasure burst in a codeword is kept small.

   33.) UXP header: is the additional header information contained in
   each RTP packet after UXP has been applied. It is always present at
   the start of the payload section of an RTP packet.

   34.) X: denotes a currently not used extension field of 1 bit in the
   UXP header.

   35.) P: is a variable which denotes the number of parity symbols per
   row used to protect the inband signaling of the redundancy profile.

   36.) ceil(.): denotes the ceiling function, i.e. rounding up to the
   next integer.

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED",  "MAY", and "OPTIONAL" in
   this document are to be interpreted as described in RFC-2119 [].

3. Introduction

   Due to the increasing popularity of high-quality multimedia
   applications over the Internet and the high level of public
   acceptance of existing mobile communication systems, there is a
   strong demand for a future combination of these two techniques: One
   possible scenario consists of an integrated communication
   environment, where users can set up multimedia connections anytime
   and anywhere via radio access links to the Internet.
   For this reason, several packet-oriented transmission modes have
   been proposed for next generation wireless standards like EGPRS
   (Enhanced General Packet Radio Service) or UMTS (Universal Mobile
   Telecommunications System), which are mostly based on the same
   principle: Long message blocks, i.e. IP packets, that enter the
   wireless part of the network are split up into segments of desired
   length, which can be multiplexed onto link layer packets of fixed
   size. The latter are then transmitted sequentially over the wireless
   link, reassembled, and passed on to the next network element.

   However, compared to the rather benign channel characteristics on
   today's fixed networks, wireless links suffer from severe fading,
   noise, and interference conditions in general, thus resulting in a
   comparably high residual bit error rate after detection and
   decoding. By use of efficient CRC-mechanisms, these bit errors are
   usually detected with very high probability, and every corrupted
   segment, i.e. which contains at least one erroneous bit, is
   discarded to prevent error propagation through the network. But if
   only one single segment is missing at the reassembly stage, the
   upper layer IP packet cannot be reconstructed anymore. The result is
   a significant increase in packet loss rate at IP level.

   Since most multimedia applications can only recover from a very
   limited number of lost message blocks, it is vitally necessary to
   keep packet loss at IP level within a certain acceptable range
   depending on the individual quality-of-service requirements.
   However, due to the delay constraints typically imposed by most
   audio or video codecs, the use of ARQ-schemes is often prohibited
   both at link level and at transport level. In addition,
   retransmission strategies cannot be applied to any broadcast or
   multicast scenarios. Thus, forward erasure correction strategies
   have to be considered, which provide a simple means to reconstruct
   the content of lost packets at the receiver from the redundancy that
   has been spread out over a certain number of subsequent packets.

   There already exist some previous studies and proposals regarding
   erasure-resilient packet transmission, of whom the most important
   one with respect to RTP is described in [1]. Since most of them are
   based on the assumption that all parts in a message block are
   equally important to the receiver, i.e. the respective application
   cannot operate on partly complete blocks, they were optimized with
   respect to assigning equal erasure protection over the whole message
   block. However, recent developments both in audio and video coding
   have introduced the notion of progressively encoded source streams,
   for which unequal erasure protection strategies seem to be more
   promising, as it will be explained in more detail below. Although
   the scheme defined in [1] is in principle capable of supporting some
   kind of unequal erasure protection, possible implementations seem to
   be quite complex with respect to the gain in performance. Finally,
   in [1] it is assumed that subsequent RTP packets can have variable
   length, which would cause significant segmentation overhead at the
   link layer of almost all wireless systems.

   This document defines a payload format for RTP, such that different
   elements in a progressively encoded multimedia stream can be
   protected against packet erasures according to their respective
   quality-of-service requirement. The general principle, including the
   use of Reed-Solomon codes together with an appropriate interleaving
   scheme for adding redundancy, follows the ideas already presented in
   [2], but allows for finer granularity in the structure of the
   progressive source stream. The proposed scheme is generic in the way
   that it (1) is independent of the type of multimedia stream, be it
   audio or video, and (2) can be adapted to varying transmission
   quality very quickly by use of inband-signaling.

4. Reed-Solomon Codes

   Reed-Solomon (RS) codes are a special class of linear nonbinary
   block codes, which are known to offer maximum erasure correction
   capability with minimum amount of redundancy.

   An arbitrary t-erasure-correcting (n,k) RS code defined over Galois
   field GF(q) has the following parameters [3]:
   - Block length:                                      n=q-1
   - No. of information symbols in a codeword:          k
   - No. of parity-check symbols in a codeword:         n-k=t
   - Minimum distance:                                  d=t+1

   In what follows, only systematic RS codes over GF(2^8) shall be
   considered, i.e. the symbols of interest can be directly related to
   a tuple of eight bits, which is commonly called a byte in packet
   transmission. The principle structure of a codeword is shown in Fig.
   1.
   By shortening the initial (n=255,n-t) RS code, any desired (n',n'-t)
   RS code for a given erasure correction capability t may be obtained.

     block of n bytes
   <----------------->
   +-+-+-+-+-+-+-+-+-+
   |&|&|&|&|&|&|&|*|*|
   +-+-+-+-+-+-+-+-+-+
   <------------><--->
       k=n-t       t
     (&:info)     (*:parity)

   Fig. 1: Structure of a systematic RS codeword

5. Progressive Source Coding

   If the output of a multimedia codec, be it audio or video, is said
   to be progressive, the encoded bitstream must consist of several
   distinct elements, often organized in separate layers. The latter
   shall be defined via their relative importance with respect to the
   quality of the reconstruction process at the receiver. Hence, there
   exists at least one layer, often called base layer, without which
   reconstruction fails at all, whereas all the other layers, often
   called enhancement layers, just help to continually improve the
   quality. Consequently, the different layers are usually contained in
   the (source-)encoded bitstream in decreasing order of importance,
   i.e. the base layer data is followed by the various enhancement
   layers.
   An example can be found in the fine granular scalability modes which
   have been proposed to various standardization bodies like MPEG-4 [4]
   or ITU (H.26L) [5], where the resolution of the scaling process in
   the progressive source encoder is as low as one symbol in the
   enhancement layer.

   From the above definition, it is quite obvious that the most
   important base layer data must be protected as strongly as possible
   against packet loss during transmission. However, the protection of
   the enhancement layers could be continually lowered, since a loss at
   this stage has only minor consequences for the reconstruction
   process. Thus, by using a suitable unequal erasure protection
   strategy across a progressive source stream, the overhead due to
   redundancy spent per (channel-)encoded block is reduced.
   Furthermore, if channel conditions get worse during transmission,
   only more and more enhancement layers are lost, i.e. a graceful
   degradation in application quality at the receiver is achieved [6].

   Nevertheless, it should be mentioned that the specific structure of
   a (source-)encoded bitstream strongly depends on the actual media
   codec in use, and the desired syntax which is used for adapting the
   output of the codec to a suitable transport level format (see also
   7.3). In order to keep the description of the unequal erasure
   protection strategy in section 6 as general as possible, the final
   bitstream which has to be protected by the proposed UXP scheme will
   be called "info stream" in the following. Furthermore, it is assumed
   that every info stream is already octet-aligned according to the
   standard procedures defined in the context of the used syntax
   specifications.

6. General Structure of UXP schemes

   In this section, the principle features of the proposed UXP scheme
   are described with a special focus on the protection and
   reconstruction procedure which is applied to the info stream. In
   addition, the behavior of the sender and receiver is specified as
   far as it concerns the reconstruction of the info stream. However,
   the complete UXP payload structure, including the additional UXP
   header, is described in section 7.

   Fig. 1 already illustrated the structure of a systematic codeword,
   which shall be represented by a single row and n successive columns
   that contain the information and the parity bytes. This structure
   shall now be extended by forming a transmission block (TB)
   consisting of L codewords of length n bytes each, which amounts to a
   total of L rows and n columns [7]: Each column, together with the
   respective UXP header in front, shall represent the payload of an
   RTP packet, i.e. the whole data of a TB is transmitted via a
   sequence of n RTP packets all carrying a payload of length (L+2)
   bytes (UXP header included).

   The value of L should be chosen in such a way that the whole length
   of the resulting IP packet (i.e. RTP payload plus sum of RTP, UDP,
   and IP header) equals a multiple of the segment size on the wireless
   link to avoid stuffing at the data link layer.

   Each TB usually consists of two or more horizontal slices, the so-
   called transmission sub blocks (TSB), as can be seen in Fig. 2: The
   first L_s rows always belong to the signaling TSB, which is used to
   convey the actual redundancy profile in the data part to the
   receiver (see 7.3). The following L_d=(L-L_s) rows belong to one or
   more data TSBs, which contain the interleaved and RS encoded info
   stream, as will be described below.

   Transmission Block (TB)

                /\ +-+-+-+-+-+-+-+-+-+ /\
                |  |  signaling TSB  |  |  L_s bytes
                |  +-+-+-+-+-+-+-+-+-+ \/
                |  |                 | /\               /\
                |  +   data TSB #1   +  |  L_d(1) bytes  |
                |  |                 |  |                |
                |  +-+-+-+-+-+-+-+-+-+ \/                |
   L bytes      |  |                 | /\                |
   payload      |  +   data TSB #2   +  |  L_d(2) bytes  |
   per packet   |  +                 |  |                |  L_d bytes
                |  +-+-+-+-+-+-+-+-+-+ \/                |
                |  |        .        |  .                |
                |  +        .        +  .                |
                |  |        .        |  .                |
                |  +-+-+-+-+-+-+-+-+-+ /\                |
                |  |   data TSB #z   |  |  L_d(z) bytes  |
                \/ +-+-+-+-+-+-+-+-+-+ \/               \/
                   <----------------->
                         n packets

   Fig. 2: General structure of a TB

   Since the UXP procedure is mainly applied to the data TSBs, it will
   be described next, whereas the content and syntax of the signaling
   TSB will be defined in section 7.3.

   For means of simplification, only one single data TSB will be
   assumed throughout the following explanation of the encoding and
   decoding procedure. However, an extension to more than one data TSB
   per TB is straightforward, and will be shown in section 7.4.

   As depicted in Fig. 3, the rows of a transmission sub block shall be
   partitioned into T+1 different classes CA_i, where i=0...T, such
   that each class contains exactly A_i=|CA_i| consecutive rows of the
   matrix, where the A_i have to satisfy the following relationship:

   A_0+A_1+...+A_T=L_d
   Data Transmission Sub Block (data TSB)
                                 T
                             <------->
                /\ +-+-+-+-+-+-+-+-+-+ /\
                |  |&|&|&|&|&|*|*|*|*|  |
                |  +-+-+-+-+-+-+-+-+-+  |  A_T=3
                |  |&|&|&|&|&|*|*|*|*|  |
                |  +-+-+-+-+-+-+-+-+-+  |
   L_d bytes    |  |&|&|&|&|&|*|*|*|*| \/
   per packet   |  +-+-+-+-+-+-+-+-+-+ /\
                |  +%|%|%|%|%|%|*|*|*|  |%|%|%|%|%|%|*|*|*|  |  A_(T-1)=1
                |  +-+-+-+-+-+-+-+-+-+ \/
                |  |$|$|$|$|$|$|$|*|*|  .
                |  +-+-+-+-+-+-+-+-+-+  .
                |  |||||||||*|  |!|!|!|!|!|!|!|!|*|  .
                |  +-+-+-+-+-+-+-+-+-+ /\
                |  |#|#|#|#|#|#|#|#|#|  |  A_0=1
                \/ +-+-+-+-+-+-+-+-+-+ \/
                   <----------------->
                         n packets

                  &,%,$,,#

   &,%,$,!,# : info bytes belonging to a certain info stream in
               decreasing order of importance
   * :         parity bytes gained from Reed-Solomon coding

   Fig. 3: General structure for coding with unequal erasure protection

   Furthermore, all rows in a particular class CA_i shall contain
   exactly the same number of parity bytes, which is equal to the index
   i of the class. For each row in a certain class CA_i, the same (n,n-
   i) RS code shall be applied.

   As can be observed from Fig. 3, class CA_T contains the largest
   number of parity bytes per row, i.e. offers the highest erasure
   protection capability in the block. Consequently, the most important
   element in the info stream must be assigned to class CA_T, where the
   value of T should be chosen according to the desired outage
   threshold of the application given a certain packet erasure rate on
   the link.
   All other classes CA_(T-1)...CA_0 shall be sequentially filled with
   the remaining elements of the info stream in decreasing order of
   importance, where the optimal choice for the size of each class (0
   or more rows), i.e. the structure of the redundancy profile, should
   depend on the quality-of-service requirements for the various
   (progressively-encoded) layers.

   The following set of rules contains a compact description of all the
   operations that must be performed for each transmission block:

   1.) The total number of columns n of the TB shall be chosen
   according to the actual delay constraints of the application.

   2.) Next, the expected number of rows reserved for the signaling TSB
   has to selected, which limits the data TSB to L_d=(L-L_s) rows.

   3.) The maximum erasure correction capability T in the data TSB
   should be chosen according to the desired outage threshold of the
   application given the actual packet erasure rate on the link.

   4.) The redundancy profile for the rest of the data TSB should
   depend on the size and number of the various layers in the info
   stream, as well as the desired probability of successful decoding
   for each of them (quality-of-service requirement).

   5.) Any suitable optimization algorithm may be used for deriving an
   adequate redundancy profile. However, the result has to satisfy the
   following constraints:
   a) All available info byte positions in the data TSB have to be
   completely filled. If the info stream is too short for a desired
   profile, media stuffing may be applied to the empty info byte
   positions at the end of the data TSB by appending a sufficient
   number of bytes (with arbitrary value, e.g. 0x00). The actual number
   of stuffing symbols per data TSB is then signaled via the respective
   stuffing indicator (see 7.3). However, before resorting to any
   stuffing, it should be checked whether it is possible to strengthen
   the protection of certain rows instead, thus improving the overall
   robustness of the decoding process.
   b) The info stream should be fully contained within the data TSB
   (unless cutting it off at a specific point is explicitly allowed by
   the properties of the used media codec).
   c) The number of required descriptors and stuffing indicators (see
   section 7.3) to signal the profile shall not exceed the space
   initially reserved for them in the signaling TSB.
   Constraints a) and b) should be already incorporated in the
   optimization algorithm. However, if constraint c) is not met, the
   data TSB has to be reduced by one row in favor of the signaling TSB
   to accomodate more space for the descriptors and stuffing
   indicators, i.e. steps 2-5 have to be repeated until a valid
   redundancy profile has been obtained.

   6.) For each nonempty class CA_i, i=T...0, in the data TSB, the
   following steps have to be performed:
   a) All rows of this specific class shall be filled from left to
   right and top to bottom with data bytes of the info stream in
   decreasing order of importance (i.e. starting with the most
   important element).
   b) For each row in the class, the required i parity-check bytes are
   computed from the same set of codewords of an (n,n-i) RS code, and
   filled in the empty positions at the end of each row. Thus, every
   row in the class constitutes a valid codeword of the chosen RS code.

   7.) After having filled the whole data TSB with information and
   parity bytes, the redundancy profile is mapped to the signaling TSB
   as described in section 7.3.

   8.) Each column of the resulting TB is now read out byte-wise from
   top to bottom and, together with the respective UXP header (see
   section 7.2) in front, is mapped onto the payload section of one and
   only one RTP packet.

   9.) The n resulting RTP packets shall be transmitted subsequently to
   the remote host, starting with the leftmost one.

   10.) At the corresponding protocol entity at the remote host, the
   payload (without the UXP header) of all successfully received RTP
   packets belonging to the same sending TB shall be filled into a
   similar receiving TB column-wise from top to bottom and left to
   right.

   11.) For every erased packet of a received TB, the respective column
   in the TB shall be filled with a suitable erasure marker.

   12.) Before any other operations can be performed, the redundancy
   profile has to be restored from the signaling TSB according to the
   procedure defined in section 7.3. If the attempt fails because of
   too many lost packets, the whole TB shall be discarded and the
   receiving entity should wait for the next incoming TB (the source
   decoder may be informed about the missing info stream, if required).

   13.) If the attempt to recover the redundancy profile has been
   successful, a decoding operation shall be performed for each row of
   the data TSB by applying any suitable algorithm for erasure
   decoding.

   14.) For all rows of the data TSB for which the decoding operation
   has been successful, the reconstructed data bytes are read out from
   left to right and top to bottom, and appended to the reconstructed
   version of the info stream.

   15.) For all rows of the data TSB for which the decoding operation
   has failed, a sufficient number of suitable dummy symbols may be
   added to the reconstructed info stream to inform the source decoder
   about the missing symbols.

   One can easily realize that the above rules describe an interleaver,
   i.e. at the sender a single codeword of a TB is spread out over n
   successive packets. Thus, each codeword of a transmitted TB
   experiences the same number of erasures at exactly the same
   positions.
   Two important conclusions can be drawn from this:
   a) Since the same RS code is applied to all rows contained in a
   specific class, either all of them can be correctly decoded or not.
   Hence, there exist no partly decodable classes at the receiver.
   b) If decoding is successful for a certain class CA_i, all the
   classes CA_(i+1)...CA_T can also be decoded, since they are
   protected by at least one more parity byte per row. Together with
   rule 6, it is therefore always ensured, that in case a decodable
   enhancement layer exists, all other layers it depends on can also be
   reconstructed!

   Given the maximum erasure protection value T, the redundancy profile
   for a data TSB of size (L_d x n) shall be denoted by a so-called
   erasure protection vector AV of length (T+1), where

   AV:=(A_0,A_1,...,A_(T-1),A_T)

   From the above definition, it is easy to realize that the trivial
   cases of no erasure protection and EXP are a subset of UXP:
   a) no erasure protection at all: all application data is mapped onto
      class CA_0, i.e. AV=(L_d,0,0,...,0).
   b) EXP: all application data is mapped onto class CA_T, i.e.
      AV=(0,0,...,0,A_T=L_d).

   Hence, backward compatibility to currently standardized non-
   progressive multimedia codecs is definitely achieved.

7. RTP payload structure

   For every packet whose payload is formed by reading out a column of
   the TB, the RTP header must be followed by an UXP header.

7.1. Specific settings in the RTP header

   The timestamp of each RTP packet resulting from reading out a TB is
   set to the time instant when the first byte of the progressive
   source data stream has been written into the TB. This results in the
   TS value being the same for all RTP packets belonging to a specific
   TB.

   The payload type is of dynamic type, and obtained through out-of-
   band signaling similar to [1]. The signaling protocol must establish
   a payload length to be associated with the payload type value. End
   systems, which cannot recognize a payload type, must discard it.

   The marker bit is set to 1 for every last packet in a TB. Otherwise,
   its value is 0.

   All other fields in the RTP header are set to those values proposed
   for regular multimedia transmission using the same source codecs,
   but no erasure protection scheme enabled.

   The RTP payload shall consist of the UXP header followed by one
   column of the TB.

7.2. Structure of the UXP header

   The UXP header shall consist of 2 octets, and is shown in Fig. 4:

    0                   1 1 1 1 1 1
    0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |X|  block PT   | block length n|
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

   Fig. 4: Proposed UXP header

   The fields in the header shall be defined as follows:
   - X (bit 0): extension bit, reserved for future enhancements,
                currently not in use -> default value: 0

   - block PT (bits 1-7): regular RTP payload type to indicate the
                          media type contained in the info stream

   - block length n (bits 8-15): indicates total number of RTP packets
                                 resulting from one TB (which equals
                                 the number of columns of the TB)

   The syntax of the info stream which is protected by UXP is specified
   by the RTP payload type field contained in the UXP header. For
   example, payload type H.263 means that the info stream conforms to
   the specifications of the RTP profile for H.263, but does not
   represent the "raw" H.263 stream produced by a H.263 encoder.
   However, UXP can also be applied to the raw output of the media
   codec (in case it is already octet-aligned), if this can be signaled
   to the receiver via other means, e.g. by use of H.245 or SDP.

   Based on the RTP sequence number, the marker bit, and the repetition
   of the block length n in each UXP header, the receiving entity is
   able to recognize both TB boundaries and the actual position of lost
   packets in the TB. Furthermore, the specific choice of equal TS
   values for all RTP packets belonging to a TB allows for overcoming
   possible sequence number overflow.

7.3. In-band signaling of the structure of the redundancy profile

   To enable a dynamic adaptation to varying link conditions, the
   actual redundancy profile used in the data TSB must be signaled to
   the receiving entity. Since out-of-band signaling either results in
   excessive additional control traffic, or prevents quick changes of
   the profile between successive TBs, an in-band signaling procedure
   is desired.

   As without knowledge of the correct redundancy profile, the decoding
   process cannot be applied to any of the erasure protection classes,
   it has to be protected at least as strongly as the most important
   element in the info stream against packet loss. Therefore, an
   additional class CA_P is used in the signaling TSB, where the number
   of parity symbols is by default set to the following value:

   P=ceil(n/2)

   Hence, up to 50% of the RTP packets can be lost, before the
   redundancy profile cannot be recovered anymore. This seems to be a
   reasonable value for the lowest point of operation over a lossy
   link. Alternatively, p may be explicitly signaled during session
   setup by means of SDP or H.245 protocol.

   Consequently, since all other classes must have equal or less
   erasure protection capability, the maximum allowable value for class
   CA_T in the data TSB is now limited to T<=P.

   The signaling of the erasure protection vector is accomplished by
   means of descriptors. For each class CA_i with A_i>0, there is a
   descriptor DP_i providing information about the size of class CA_i
   (i.e. the value of A_i) and establishing a relationship between the
   erasure protection of class CA_i and that of the first preceding
   class CA_(i+j) with A_(i+j)>0, where j>0. A descriptor DP_i is
   mapped onto one byte, which is sub-divided into two half-bytes (i.e.
   the higher and the lower four bits). The first half-byte is of type
   unsigned and contains the 4-bit representation of the decimal value
   A_i. The second half-byte is of type signed and contains the
   difference in erasure protection between class CA_i and class
   CA_(i+j), i.e. the signed 4-bit representation of the decimal value
   (-j) (where the MSB denotes the sign, and the lower three bits the
   absolute value). Note that the erasure protection p of class CA_p is
   fixed, whereas the size A_p may vary.

   Thus, the data to be filled into class CA_p shall consist of a
   sequence of descriptors separated by stuffing indicators (see
   below), where the number of descriptors is primarily given by the
   number of protection classes CA_i, 0<=i<=T, in the data TSB with
   A_i>0.
   Without a-priori knowledge, the initial value for the size of the
   signaling TSB should be set to one (row). When the number of
   necessary descriptors and stuffing indicators exceeds the (n-p)
   information positions, one or more additional rows have to be
   reserved. This is usually done by increasing the value for L_s to
   A_p>1, i.e. the data TSB is reduced to (L-A_p) rows. Hence, in order
   to indicate the actual size of the signaling TSB, an additional
   descriptor is inserted at the very beginning, which takes on the
   value 0xq0, where q denotes the (octal) four bit representation of
   the decimal value A_p.

   Furthermore, the end of each data TSB is signaled by the otherwise
   unused descriptor value 0x00, followed by exactly one stuffing
   indicator (SI). The latter is mapped onto a byte, which is of type
   unsigned and contains the 8-bit representation of the decimal value
   of the number of media stuffing symbols used at the end of the
   respective data TSB.

   The (extended) sequence of descriptors and stuffing indicators is
   then mapped to the info byte positions in the A_p rows of the
   signaling TSB from left to right and top to bottom. Each row is then
   encoded with the same (n,n-p) RS code.

   If the number of descriptors and stuffing indicators is less than
   the available info byte positions, however, empty positions in class
   CA_p may be filled up with the otherwise unused descriptor 0x00.

   At the receiving entity, the sequence of descriptors shall be
   recovered by performing erasure decoding on the first row of the TB
   (which definitely belongs to the signaling TSB) using the same
   algorithm as later for the data TSB. If successful, the very first
   descriptor now indicates the number of rows of the signaling TSB,
   and the next (A_p-1) rows are decoded to reconstruct the redundancy
   profile for the data TSB(s), together with the number of media
   stuffing symbols denoted by the respective SI(s).

   The complete structure of the TB is now depicted in Fig. 5.

   Transmission Block (TB)
                                P
                           <--------->
                /\ +-+-+-+-+-+-+-+-+-+ /\
                |  |?|?|?|?|*|*|*|*|*|  |  A_P=1
                |  +-+-+-+-+-+-+-+-+-+ \/
                |  |&|&|&|&|&|*|*|*|*| /\
                |  +-+-+-+-+-+-+-+-+-+  |  A_T=3
                |  |&|&|&|&|&|*|*|*|*|  |
                |  +-+-+-+-+-+-+-+-+-+  |
   L bytes      |  |&|&|&|&|&|*|*|*|*| \/
   payload      |  +-+-+-+-+-+-+-+-+-+ /\
   per packet   |  +%|%|%|%|%|%|*|*|*|  |  A_(T-1)=1
                |  +-+-+-+-+-+-+-+-+-+ \/
                |  |$|$|$|$|$|$|$|*|*|  .
                |  +-+-+-+-+-+-+-+-+-+  .
                |  |||||||||*|  |!|!|!|!|!|!|!|!|*|  .
                |  +-+-+-+-+-+-+-+-+-+ /\
                |  |#|#|#|#|#|#|#|#|#|  |  A_0=1
                \/ +-+-+-+-+-+-+-+-+-+ \/
                   <----------------->
                         n packets

   ? :          descriptors and stuffing indicators for in-band
                signaling of the redundancy profile

                  &,%,$,,#

   &,%,$,!,# :  info bytes belonging to a certain element of the
                info stream in decreasing order of importance

   * :          parity bytes gained from Reed-Solomon coding

   Fig. 5: General structure for UXP with in-band signaling of the
   redundancy profile

   The following simple example is meant to illustrate the idea behind
   using descriptors: Let an erasure protection vector of length T+1=7
   be given as follows:
   AV=(A_0,A_1,...,A_5,A_6)=(7,0,2,2,0,3,10)
   Hence, the length L of the TB (including one row for the signaling
   TSB) is equal to 7+2+2+3+10+1=25 (rows/bytes). If the width is
   assumed to be equal to 20 (columns/packets), then the erasure
   protection of the descriptors is p=10.
   The corresponding sequence of descriptors can be written as
   DP=(DP_6,DP_5,DP_3,DP_2,DP_0)=(0xAC,0x39,0x2A,0x29,0x7A),
   where the values of the descriptors are given in hexadecimal
   notation. Next, the descriptor indicating the length of the
   signaling TSB has to be inserted, the end of the data TSB has to be
   marked by 0x00, and the SI has to be appended. If the number of
   media stuffing symbols is assumed to be 3, the 10 info bytes in the
   signaling TSB take on the following values (descriptor stuffing
   included):

   (0x10,0xAC,0x39,0x2A,0x29,0x7A,0x00,0x03,0x00,0x00)

   7.4 Optional Concatenation of Transmission Sub Blocks:

   The following procedure may be applied if a single info stream would
   be too short to achieve an efficient mapping to a transmission block
   with respect to the fixed payload length L and the desired number of
   packets n. For example, intra-coded video frames (I-frames) are
   usually much larger than the following predicted ones (P-frames). In
   this case, a certain number z of successive small info streams
   should be each mapped to a transmission sub block with length L_d(y)
   and width n, such that L_d(1)+L_d(2)+?+L_d(z)=L_d.
   The resulting transmission sub blocks can then be easily
   concatenated to form a TB of size L x n having one common signaling
   TSB: Since the second half-byte of the descriptors is of type
   signed, we are able to incorporate both decreasing and increasing
   erasure protection profiles within one single signaling TSB.
   Note that once the lengths L_d(y) of the individual blocks have been
   fixed, the respective redundancy profiles can be determined
   independently of each other. However, the space initially reserved
   for the signaling TSB should be already large enough to avoid
   profile recalculation for each of the data TSBs in case the sequence
   of descriptors gets too long!

   Again, we will give a simple example to illustrate this idea: Let
   the erasure protection vectors for two concatenated data TSBs be
   given as follows:

   AV1=(A1_0,A1_1,...,A1_5,A1_6)=(0,0,2,2,0,3,10),
   AV2=(A2_0,A2_1,...,A2_5,A2_6)=(0,0,2,2,0,3,10).

   Hence, two single identical data TSBs will be concatenated to form a
   TB of length L=2*(2+2+3+10)+2=36 (rows/bytes). If the width is again
   assumed to be equal to 20 (columns/packets), then the erasure
   protection of the descriptors is p=10, and therefore a total of two
   rows for the signaling TSB have been reserved this time. The
   corresponding sequence of descriptors can now be written as
   DP=(0xAC,0x39,0x2A,0x29,0xA4,0x39,0x2A,0x29), where the values of
   the descriptors are given in hexadecimal notation. If the number of
   media stuffing symbols is assumed to be 3 for each data TSB, the 20
   info byte positions in the signaling TSB are filled with the
   following values (descriptor stuffing included):

   (0x20,0xAC,0x39,0x2A,0x29,0x00,0x03,0xA4,0x39,0x2A,0x29,0x00,0x03,
   0x00,0x00,0x00,0x00,0x00,0x00,0x00)

8. Security Considerations

   The payload of the RTP-packets consists of an interleaved multimedia
   and parity stream. Therefore, it is reasonable to encrypt the
   resulting stream with one key rather than using different keys for
   multimedia and parity data. It should also be noted that encryption
   of the multimedia data without encryption of the parity data could
   enable known-plaintext attacks.

   The overall proportion between parity bytes and info bytes should be
   chosen carefully if the packet loss is due to network congestion. If
   the proportion of parity bytes per TB is increased in this case, it
   could lead to increasing network congestion. Therefore, the
   proportion between parity bytes and info bytes per TB MUST NOT be
   increased as packet loss increases due to network congestion.

   The overall ratio between parity and info bytes MUST NOT be higher
   than 1:1, i.e. the absolute bitrate spent for redundancy must not be
   larger than the bitrate required for transmission of multimedia data
   itself.

9. Application Statement

   There are currently two different schemes proposed for unequal error
   protection in the IETF-AVT: Unequal Level Protection (ULP) and
   Unequal Erasure Protection (UXP).
   Although both methods seem to address the same problem, the proposed
   solutions differ in many respects. This section tries to describe
   possible application scenarios and to show the strength and
   weaknesses of both approaches.

   The main difference between both approaches is that while ULP
   preserves the structure of the packets which have to protected and
   provides the redundancy in extra packets, UXP interleaves the info
   stream which has to be protected, inserts the redundancy information,
   and thus creates a totally new packet structure.

   Another difference concerns multicast compatibility: It cannot be
   assumed that all future terminals will be able to apply UXP/ULP.
   Therefore, backward compatibility could be an issue in some cases.
   Since ULP does not change the original packet structure, but only
   adds some extra packets, it is possible for terminals which do not
   support ULP to discard the extra packets. In case of UXP, however,
   two separate streams with and without erasure protection have to be
   sent, which increases the bandwidth.

   Next, both approaches offer different mechanism to adjust packet
   sizes, if necessary: UXP allows to adjust the packet sizes
   arbitrarily. This is an advantage in case the loss probability is
   dependent on the packet length, which happens, for example, if the
   end-to-end connection contains wireless links. In this case proper
   adjustment of the packet size is one essential network adaption
   technique. In addition, if a preencoded stream is sent over the
   network, the packet size can be adjusted independently of slice
   structures.
   Since ULP does not change the existing packetization scheme, this
   flexibility does not exist.

   The ability of UXP to adjust the packet size arbitrarily can be
   especially exploited in a streaming scenario, if a delay of several
   hundred milliseconds is acceptable. It is then possible to fill
   several video frames into a single TB of desired size, e.g. a group
   of pictures consisting of I-frame, P-frames and B-frames. The
   redundancy scheme can thus be selected in such a way as to guarantee
   the following property: In case of packet loss, the streams for P-
   frames are only recoverable, if the I-frame, on which the decoding of
   P-frames depends, is recoverable. The same is true for B-frames,
   which can only be decoded if the respective P-frames are recoverable.
   This prevents situations in which, for example, the B-frames have
   been received correctly, but the P-frames have been lost, i.e.
   assures a gradual decrease in application quality also on the frame
   level. Of course, a similar encoding is possible with ULP. But in
   this case one might have to send several frames within one packet
   which leads to large packet sizes.

                  Finally,

   Furthmore, decoding delay is also a crucial issue in communications.
   Again, both approaches have different delay properties: UXP
   introduces a decoding delay because a reasonable amount of correctly
   received packets are necessary to start decoding of a TB. The delay
   in general depends on the dimensions of the interleaver. This should
   be considered for any system design which includes UXP.
   With ULP, every correctly received media packet can be decoded right
   away. However, a significant delay is introduced, if packets are
   corrupted, because in this case one has to wait for several
   redundancy packets. Thus, the delay is in general dependent on the
   actual ULP-FEC-packet scheme and cannot be considered in advance
   during the system design phase.

   Finally, we want to point out that UXP uses RS-codes which are known
   to be the most efficient type of block codes in terms of erasure
   correction capability.

10. Intellectual Property Considerations

   Siemens AG has filed patent applications that might possibly have
   technical relations to this contribution.
   On IPR related issues, Siemens AG refers to the Siemens Statement on
   Patent Licensing, see http://www.ietf.org/ietf/IPR/SIEMENS-General.

11. References

   [1] J. Rosenberg and H. Schulzrinne, "An RTP Payload Format for
   Generic Forward Error Correction", Request for Comments 2733,
   Internet Engineering Task Force, Dec. 1999.

   [2] A. Albanese, J. Bloemer, J. Edmonds, M. Luby, and M. Sudan,
   "Priority encoding transmission", IEEE Trans. Inform. Theory, vol.
   42, no. 6, pp. 1737-1744, Nov. 1996.

   [3] Shu Lin and Daniel J. Costello, Error Control Coding:
   Fundamentals and Applications, Prentice-Hall, Inc., Englewood
   Cliffs, N.J., 1983.

   [4] W. Li: "Fine Granularity Scalability Using Bit-Plane Coding of
   DCT Coefficients", ISO/IEC JTC1/SC29/WG11, Doc. MPEG98/M4204, Dec.
   1998.

   [5] G. Blaettermann, G. Heising, and D. Marpe: "A Quality Scalable
   Mode for H.26L", ITU-T SG16, Q.15, Q15-J24, Osaka, May 2000.

   [6] F. Burkert, T. Stockhammer, and J. Pandel, "Progressive A/V
   coding for lossy packet networks - a principle approach", Tech.
   Rep., ITU-T SG16, Q.15, Q15-I36, Red Bank, N.J., Oct. 1999.

   [7] Guenther Liebl, "Modeling, theoretical analysis, and coding for
   wireless packet erasure channels", Diploma Thesis, Inst. for
   Communications Engineering, Munich University of Technology, 1999.

12. Acknowledgments

   Many thanks to Thomas Stockhammer, who initially came up with the
   idea of unequal erasure protection to improve progressive video
   transmission over lossy networks.

13. Author's Addresses

   Guenther Liebl, Thomas Stockhammer
   Institute for Communications Engineering (LNT)
   Munich University of Technology
   D-80290 Munich
   Germany
   Email: {liebl,tom}@lnt.e-technik.tu-muenchen.de

   Minh-Ha Nguyen, Frank Burkert
   Siemens AG - ICM D MP RD MCH 83/81
   D-81675 Munich
   Germany
   Email: {minhha.nguyen,frank.burkert}@mch.siemens.de

   Marcel Wagner, Juergen Pandel, Wenrong Weng, Gero Baese
   Siemens AG - Corporate Technology CT IC 2
   D-81730 Munich
   Germany
   Email:
   {marcel.wagner,juergen.pandel,wenrong.weng,gero.baese}@mchp.siemens.
   de

Full Copyright Statement

   "Copyright (C) The Internet Society (date). All Rights Reserved.
   This document and translations of it may be copied and furnished to
   others, and derivative works that comment on or otherwise explain it
   or assist in its implementation may be prepared, copied, published
   and distributed, in whole or in part, without restriction of any
   kind, provided that the above copyright notice and this paragraph
   are included on all such copies and derivative works. However, this
   document itself may not be modified in any way, such as by removing
   the copyright notice or references to the Internet Society or other
   Internet organizations, except as needed for the purpose of
   developing Internet standards in which case the procedures for
   copyrights defined in the Internet Standards process must be
   followed, or as required to translate it into languages other than
   English.

   The limited permissions granted above are perpetual and will not be
   revoked by the Internet Society or its successors or assigns.

   This document and the information contained herein is provided on an
   "AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING
   TASK FORCE DISCLAIMS ALL WARRANTIES; EXPRESS OR IMPLIED; INCLUDING
   BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF INFORMATION HEREIN
   WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF
   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.