On Packetization of Embedded Multimedia Bitstreams

Xiaolin Wu, Samuel Cheng, and Zixiang Xiong

IEEE Transactions On Multimedia, March 2001IEEE Transactions On Multimedia, March 2001

Outline

Introductionpacketization

Problem formulation Optimal Packetization

High Bit RateLow Bit Rate

Result Conclusion

Introduction

Problem of multimedia communicationpacket dropping corrupted packets

Techniques to alleviate or recovererror detection codesautomatic repeat request (ARQ)forward error correction (FEC)

• Error Concealment

Introduction (cont.)

Resynchronizationperiodic symbols insert into the

compressed source bit-streams.

Confine errors to local segment of long message

Error resilience

dataRecyn

error

Recyn data

Packetization

packet independentdata partition

bit-stream compressionRLC - make bit-stream different size

Question :

How to pack variable length bit-streams into packets of a fixed size

Packetization (cont.)

One of the solution is to fill the packets with the bit-streams sequentially.defeating the purpose of resynchronization

stream 1 stream 2 stream 3

stream 3 stream4 stream 5

stream 6 .........

packet 1

packet 2

packet 3

Recynchronization marker

Packetization (cont.)

Another solution is to enforce the alignment of the bitstreamnot allowing any bit-stream to start in the

middle of a packet.packetization inefficiency

stream 1

stream 2

stream 3

packet 1

packet 2

packet 3

Problem Formulation

Embedded bit-streamGiven a traversal, the resulting binary

sequence is a so-called embedded bitstream.• several pass such as bit-plane coding

Scalability in reconstruction quality.• can be truncated at any location

Problem Formulation (cont.)

K sample blocks S1, S2, ......., SK

compressed independently of each other Compressed bitstream Bi, 1 i K

• scalable in rate-distortion. Ni Length of Bi, 1 i K

M packet of payload LMLN

Kkk

1

Optimal Packetization(OP)

We want to select ML bits to be packeted into M packets.To minimize the damage of packet los

s by packet alignment constraintsTo minimize the distortion under pack

et alignment constraints

High Bit Rate Case

one bitstreams have to occupy an integer number of packet

pre-defined function: : the distortion of first a bits of Bk

)(adk

)()(),( ladadla kkk

High Bit Rate Case(cont.)

Original greedy approach

sort all Δ value in descending orderpick the M largest distortion reducti

ons

1/0,1),,( LNiKkLiL kk

Question :Not contiguous subsequence from first bit of the embedded bitstream

High Bit Rate Case(cont.)

Improved algorithm:

Maintain a pointer pk for each bitstream

Bk, 1 k K.

Initialize pk = 1, 1 k K ; m=0;

repeat

find j such that Δj (pj, L) = max 1 k K Δk(pk, L)

pack this L bits into packet m;

pk = pk + L;

m = m + 1;

until m = M;

Add L bits of bitstream bk will reduce the most distortion

High Bit Rate Case(cont.)

nonconvex operation R-D function

solve D(M, K) in bottom-up Dynamic programming

Low Bit Rate Case

We often have M < K allow more than one embedded bitstre

ams to be packed into one packet If k bitstreams are to share a packet, t

hey have to be completely contained in that packet.

Low Bit Rate Case (cont.)

NP complete

If we impose an order for bitstreams Bk, and allow a packet to contain only consecutive bitstreams in this order, this problem is solvable.

Low Bit Rate Case (cont.)

minimum distortion of Bu,...,Bv Dynamic programming function

),( vu

Result:

Result (cont.)

Conclusion

Optimal Packetization is addressed under both low and high bit rate case

Using dynamic programming for nonconvex distortion