Error Control Techniques for Interactive Low-Bit Rate Video Transmission

Home , Internet video

Error Control Techniques for Interactive Low-bit Rate Video Transmission

over the Internet

Injong Rhee

Department of Computer Science

North Carolina State University

Raleigh, NC 27695-7534

E-mail: [email protected]

Abstract The e ect of packet loss is also far-reaching. Packet

loss degrades not only the quality of the frames con-

tained in lost packets, but also the quality of their sub-

A new retransmission-based error control technique is

sequent frames. This is b ecause of motion comp ensation

presented that does not incur any additional latency in

employed by video co decs. Motion-comp ensated video

frame playout times, and hence are suitable for inter-

co decs remove temp oral redundancy in a video stream

active applications. It takes advantage of the motion

by enco ding only pixel value di erence prediction er-

prediction loop employed in most motion compensation-

ror b etween an image to b e enco ded inter-frame and a

basedcodecs. By correcting errors in a referenceframe

previously transmitted image reference frame. Packet

causedbyearlier packet loss, it prevents error propaga-

loss can intro duce an error in a reference frame, which

tion. The technique rearranges the temporal dependency

can b e propagated to its subsequent frames and get am-

of frames so that a displayedframe is referenced for the

pli ed as more packets are lost.

decoding of its succeeding dependent frames much later

than its display time. Thus, the delay in repairing lost

Error propagation can be controlled by more fre-

packets can be e ectively masked out. The developed

quently adding intra frames which are co ded temp o-

technique is combined with layered videocoding to main-

rally indep endently. However, the ratio of the com-

tain consistently good video quality even under heavy

pression eciency of an intra-frame over an inter-frame

packet loss. Through the results of extensive Internet

is as large as 3 to 6 times. Increasing the frequency

experiments, the paper shows that layeredcoding can be

of intra-frames could increase the bandwidth require-

very e ective when combined with the retransmission-

menttoomuch for video transmission over a bandwidth-

based error control technique for low-bit rate transmis-

constraint network. Nonetheless, the severe degradation

sion over best e ort networks where no network-level

of image quality due to error propagation has forced

mechanism exists for protecting high priority data from

several p opular video conferencing to ols, suchas nv[7 ],

packet loss.

vic[13 ] and CU-SeeMe[6 ], to adopt an even more drastic

approach. Using a technique called conditional replen-

ishment, these to ols lter out the blo cks that have not

1 Intro duction

changed much from the previous frame and intra-co de

the remaining blo cks. Since all the co ded blo cks are

temp orally indep endent, packet loss a ects only those

Transmitting interactive video over packet switching

frames contained in lost packets. However, this en-

networks is time-constrained. If a packet containing a

hanced error resilience comes at the cost of low com-

video frame is not delivered b efore its scheduled play-

pression eciency. Additional compression can always

out time, and the frame cannot be displayed and ac-

be obtained if temp oral redundancy is removed from

cordingly, the overall quality of video degrades. The

each co ded blo ck i.e., by co ding only their prediction

diculty of meeting this time constraint lies mainly in

error.

frequent o ccurrences of packet loss which are commonly

caused by network congestion. The recovery of lost

The goal is to develop an error recovery scheme

packets b efore their display times is not always feasi-

that solves the error propagation problem without los-

ble b ecause of the delay asso ciated with detecting and

ing much compression eciency. Retransmission-based

repairing the lost packets.

and forward error correction schemes are the two ma jor

error control schemes found in the literature.

Retransmission-based error recovery REC can pro-

vide go o d error resilience without incurring much band-

width overhead b ecause packets are retransmitted only

when some indications exist that they are lost. However,

retransmission always involves additional transmission

delay, and thus has b een widely known ine ective for in-

teractive real-time video applications such as video con- frames temp orally dep end only on the essential signals

ferencing. Many researchers prop osed use of extended of their reference frames. This dep endency ensures that

control or playout times to allow retransmitted packets the distortion of the enhancement signals of a frame

to arrive in time for display [21 , 18 , 12 , 23 ]. This implies do es not a ect the quality of its succeeding frames.

that the playout time of a frame is delayed by at least Since essential signals are protected by FEC, this dep en-

three one-way trip times after its initial transmission dency will signi cantly reduce error propagation. How-

two for packet transmissions and one for a retransmis- ever, QAL has its own limitations. First, since frames

sion request. Under the currentInternet environment, are motion-comp ensated only to the essential signals of

this delaywould b e intolerable for interactive video ap- their reference frames, the temp oral redundancy present

plications, and can severely reduce the interactivityof in the enhancement signals are not exploited at all, re-

the application. sulting in low compression eciency. Second, under

heavy packet loss, even essential signals can be lost,

Retransmission, however, can be still a very e ec-

causing error propagation.

tive technique to improve error resilience in interactive

real-time video conferencing. In this pap er, we study a

These limitations, however, can b e greatly reduced

REC scheme, called periodic temporal dependency dis-

when QAL is combined with PTDD. In this pap er, we

tance PTDD, that can be used for interactive video

show that a motion-comp ensated layered co ding tech-

applications. In particular, the schemes do not require

nique such as QAL can b e very e ective for transmis-

any arti cial extension of control time and play-out de-

sion over a lossy packet-switched network even when no

lays, and thus are suitable for interactive applications.

network-level mechanism to protect a prioritized stream

In the scheme, frames are simply displayed at their nor-

exists. When used with PTDD, the layering technique

mal playout times without any delay, as they are de-

yields reasonably go o d compression eciency since p e-

co ded. Thus, if a packet arrives after the playout time

rio dic frames temp orally dep ends on b oth the essential

of its frame, the frame will be displayed with errors.

and enhancement signals of their reference frames. This

However, this \late" packet can b e used to remove error

dep endency is safe b ecause retransmission can recover

propagation. Rather than discarding the late packet,

lost packets containing b oth signals. In addition, p eri-

PTDD uses it to restore its frame although the frame

o dic frames reduce error propagation b ecause their im-

is displayed. Because the frame is used as a reference

mediately succeeding non-p erio dic frames use them as

frame for its succeeding frames, restoring the reference

a reference frame.

frame stops error propagation.

We conducted b oth transatlantic and transpaci c In-

To allow enough time for retransmitted packets to

ternet transmission tests from the East Coast of the

arrive b efore their frames are referenced for the recon-

US to show the utility of the REC-based techniques for

struction of their dep endent frames, the PTDD scheme

transmission over a long distance network. The exp er-

extends the temporal dependency distance TDD of

imental results indicated that the prop osed technique

frames. The TDD of a frame is de ned to b e the min-

achieve go o d error resilience and compression eciency.

imum numb er of frame intervals or inter-frame delay

In Section 2, we describ e the related work, and in

between that frame and another frame on which it is

Section 3 we describ e PTDD and QAL schemes. Section

temp orally dep endent. In the PTDD scheme, every p-

4 contains a discussion of the exp erimental results, and

th frame whichwe call periodic frame has an extended

Section 5 contains the conclusion.

TDD while the other frames have TDD 1. Note that

the extension of TDD do es not a ect the playout times

of frames since the playout times of frames are deter-

2 Related Work

mined solely byinter-frame delays and the initial con-

trol time which is used to reduce delay jitter. In the

PTDD scheme, the TDD of p erio dic frames is deter-

We rst discuss the related work on retransmission-

mined by the estimated delaybetween the sender and

based techniques for video transmission over packet-

the receiver.

switched networks and then discuss related work on lay-

ered video co ding.

In PTDD, non-p erio dic frames with TDD 1 are not

protected by retransmission. If packet loss o ccurs for a

Retransmission-based error recovery. Dempsey et al.[5 ]

non-p erio dic frame, the frame will b e displayed with er-

applied retransmission for the recovery of audio packets.

rors which can b e propagated to its subsequent frames

They showed that by adding some delay b efore the play-

until the next p erio dic frame is received. The receiv-

out of each received audio packet, retransmission can

ing video quality will p erio dically uctuate. To remedy

b e used to protect audio data from packet loss. Their

this, we apply a layered co ding technique, called quality

work hinges on the earlier b ehavior study by Brady [3 ]

assurance layering QAL [22 , 15 , 8 ], for non-p erio dic

showing that although less than 200 ms round trip delay

frames which divides video signals into essential and en-

is required for high qualityvoice applications, delays up

hancement signals. Essential signals are protected bya

to 600 ms can b e tolerable byhuman ears.

simple forward error correction FEC technique. Since

Ramamurthy and Raychaudhuri [21 ] applied a sim-

the amount of essential signals is often much smaller

ilar technique to video transmission over ATM. They

than that of the entire signals, with only a little FEC

analyzed the p erformance of video transmission over an

redundant information, we can maintain relatively go o d

ATM network when b oth retransmission and error con-

video quality.

cealment are used to repair errors o ccurring from cell

QAL is di erent from the technique adopted in vic

loss. They analytically showed that for a coast-to-coast

[14 ] in that QAL uses motion comp ensation. In QAL,

techniques. It was shown that by utilizing priority pack- ATM connection, 33 ms to 66 ms play-out delay is su-

etization, the basic image quality can be maintained cient to see a signi cant improvement in image quality.

if the HP stream is guaranteed to be received. How-

Padop oulos and Parulkar [18 ] prop osed an imple-

ever, these techniques do not solve the error propagation

mentation of an ARQ scheme for continuous media

problem b ecause the essential signals of a frame still de-

transmission. Various techniques including selective

p end on b oth the essential and enhancement signals of

rep eat, retransmission expiration, and conditional re-

its reference frame. Since the enhancement signals are

transmission are implemented inside a kernel. Their

less reliably transmitted, frames that dep end on these

exp erimentover an ATM connection showed the e ec-

signals also carry along the same error

tiveness of their scheme.

Intra-H.261 is used in a p opular Internet video con-

Most recently, Li et al. [12 ] and Xu et al. [23 ]

ferencing to ol called vic that enco des each frame as

used retransmission in the recovery of lost packets for

an intra-frame [13 ]. Using a conditional replenishment

video multicasting. Li et al [12 ] prop osed a novel

technique, the co dec is shown to give excellent error re-

scheme for distributing an MPEG-co ded video over a

silience over packet loss. Another b ene t of the co dec is

b est-e ort network. By transmitting di erent frame

the reduced computation in enco ding since it do es not

typ es I, P and B frames of MPEG to di erentmulti-

involve a motion prediction lo op. Alayering scheme for

cast groups, they implemented a simple layering mech-

Intra-H.261 using a bit plane division technique is also

anism in which a receiver can adjust frame play-out

discussed in [2 , 14 ].

times during congestion by joining or leaving a mul-

Quality assurance layering QAL was studied in ticast group. For instance, consider an MPEG picture

[22 , 15 ]. In this scheme, errors due to the loss of LP pattern: IBBPBBPBBPBB. Their scheme delays the

packets do not propagate b ecause each frame is tem-

play-out times of frames for one frame interval. They

p orally dep endent only on the essential signals of its

call this delayed play-out time adaptive playback point.

reference frame which are assumed to be reliably re-

When a receiver leaves the B frame group, the adaptive

ceived. Shimamura et al. [22 ] also studied the e ect of

playback p oint is additionally extended by three frame

various priority packetization techniques including fre-

intervals b ecause the next frame to b e displayed after a

quency truncation.

P frame is at three frame intervals away. The scheme is

shown e ective for non-interactive real-time video ap-

Ghanbari [8 ] prop osed similar layering techniques as

plications.

QAL. In the techniques, each frame is rst decimated to

pro duce a low-resolution image, then the low-resolution In a video conference involving a large number of

image is co ded using H.261 or a DCT-based co dec and participants, di erent participants may have di erent

packetized as HP data. The original image is then com- service requirements. While some participants may re-

pared with the deco ded frame of its low-resolution frame quire real-time interactions with other participants, oth-

to pro duce a di erence image which is co ded using a ers may simply want to watch or record the confer-

di erent co dec and packetized as a LP data. This co d- ence. Xu et al. [23 ] contended that retransmission can

ing scheme will have similar error resilience as the QAL be e ectively used for the transmission of high qual-

technique since the LP data dep end only on the HP ity video to the receivers that do not need a real-time

data. MPEG-2 adopts an approach similar to this. transfer of video data. They designed a new proto col

called structure-orientedresilient multicast STORM in

Khansari et al. [10 ] applied QAL to video trans-

which senders and receivers collab orate to recover lost

mission over a mobile network to solve the \fading"

packets using a dynamic hierarchical tree structure.

problem commonly exp erienced during a mobile host

hand-o p erio d. By keeping the size of the HP stream

large ab out 83 of the total bandwidth, video quality

Layered co ding. Leicher [11 ] applied a forward error

even under fading can b e kept relatively high. They also

correction scheme known as priority encoding transmis-

studied the e ects of di erent packetizing techniques.

sion PET [1 ] to a hierarchically enco ded MPEG video.

He used a temp oral layering scheme in which reference

Prioritylayering techniques are also applied to still

frames e.g., I-frames and P frames are given higher

JPEG image transmission. Posnak et al. [20 ] used a

priority than other temp orally dep endent frames e.g.,

frequency truncation layering technique that partitions

B frames. Since B frames temp orally dep end on P and

DCT blo cks of JPEG enco ded frames into essential and

I frames which are more reliably received, this technique

enhancementlayers. Han and Polyzos [9] also studied

can e ectively suppress error propagation. However,

the e ectiveness of layered co ding through the hierarchi-

in a low bit rate conferencing, the frequency of I and

cal mo de of JPEG and presented a statistical analysis

P frames must be kept very low b ecause of their low

showing that the overhead of the co ding metho d can b e

compression eciency. Thus, the resulting images can

insigni cant.

be very jerky as packets are b eing dropp ed to a ect B

frames.

3 Error Recovery Techniques

Pancha and El Zarki applied a priority packetization

scheme to an MPEG enco ded video stream for transmis-

sion over ATM [17 ]. A frequency truncation technique

3.1 Retransmission-Based Error Control REC

is applied in which a xed numb er of DCT co ecients

Our schemes are based on a careful observation on how of each DCT blo ck are allo cated to the HP data cf.

video frames are enco ded in most motion comp ensation- [16 , 4, 20 ]. DeCleene et al. [4] also studied the p er-

based co decs. We base our discussion mostly on H.261 formance of MPEG under various priority packetization Packet received -1 -1 -1 -1 ZZVLC Q DCT

F FFFFF F 0 1 2 3 4 5 6

periodic frame periodic frame Current/reference Reference/Current S0 S1 Periodic Temporal Dependency Distance time Reference Frame R1 Base Reference Frame R0

Current Prediction

Perio dic Temp oral Dep endency Distance Frame (CP) Figure 3:

ceived, sends a retransmission request NACK to the

Inverse Motion

sender. The sender gets the NACK at time t and re-

Compensation 2

(construct display

ket 3. The retransmitted packet arrives at frame) transmits pac

Display

time t which is b efore frame 3 is displayed. Packet3is

now used to restore the R-frame of frame 3, so frame 3

can b e deco ded and displayed without an error.

Figure 2: H.261 Deco der mo di ed to handle the recov-

ery of reference frames.

Figure 2 shows a H.261 deco der mo di ed to handle

the recovery of R-frames using retransmitted packets.

The only di erence from the original H.261 deco der is

from which many motion comp ensation-based video

one additional frame bu er added to handle the recov-

standards such as MPEG are designed. In H.261, a

ery. When a packet is received and deco ded into an

video sequence consists of two typ es of video frames:

image blo ck, the deco der determines whether the blo ck

intra-frame I-frame and inter-frame P-frame. I-

b elongs to the current frame b eing deco ded or its R-

frame removes only spatial redundancy presentinthe

frame. If it is for the current frame, then the blo ckis

frame. P-frame is enco ded through motion estimation

stored into frame bu er CP along with its motion vec-

using another P-frame or I-frame as a reference frame

tor. If it is for the R-frame, the blo ck is added with

R-frame. For each image blo ck in a P-frame, mo-

its temp orally dep endent blo ck in frame bu er R and

tion estimation nds a closely matching blo ck within its

stored into R . Note that CP contains only the pre-

R-frame, and generates the distance between the two

diction error and motion vectors of the current frame

matching blo cks as a motion vector. The pixel value

while R contains the fully motion comp ensated image

di erences b etween the original P-frame and a motion-

of the R-frame of the current frame, and R contains

predicted image of the P-frame obtained by simply cut-

the R-frame of R . We call R a base reference frame

1 0

and-pasting the matching image blo cks from its R-frame

bu er. At the next display time, the current frame is

are enco ded along with the motion vectors.

constructed using the information in CP and R . After

displaying the current frame, R is copied to R and

1 0

Most of the previously prop osed retransmission

the displayed image are copied to R . In this scheme,

schemes work as follows. When a packet containing the

as long as the packets b elonging to R arrive b efore the

enco ding of a frame is lost at a receiver, the receiver de-

construction of the current frame, the packet can be

tects the loss after receiving a subsequent packet of the

used to help remove errors in the current frame. The

lost packet and sends a retransmission request to the

dead line of a packet can b e informally de ned to b e the

sender. After receiving the request, the sender retrans-

arrival time of the packet at the receiver after whichit

mits the packet. We de ne the display time of a packet

is not useful for deco ding any frame. Note that the de-

to b e the time that the frame whose enco ding is con-

co der in Figure 2 extends the deadline of packets by one

tained in the packet is displayed at the receiver. If the

frame interval without delaying frame play-out times.

retransmitted packet arrives b efore its display time, the

frame can b e fully restored. Otherwise, it is discarded

We can easily generalize the ab ovescheme to extend

and the displayed image contains some error. All the

packet deadlines beyond one frame interval. The pe-

subsequently deco ded frames will carry the same error

rio dic temp oral dep endency distance scheme PTDD

until a new I-frame is received.

generalizes the ab ove scheme. See Figure 3. In the

scheme, every i-th frame has an extended temp oral de-

Our scheme di ers from others in that retransmit-

p endency distance TDD i we call this frame aperi-

ted packets arriving after their display times are not

odic frame while all the other inter-frames have TDD

discarded but instead used to reduce error propagation.

1. The TDD of p erio dic frames in Figure 3 is 4. In fact,

In motion comp ensation-based co decs, the correct im-

the pattern of the temp oral dep endency is very simi-

age reconstruction of a currently displayed image de-

lar to a picture group pattern of MPEG. All the frames

p ends on a successful reconstruction of its R-frames.

with TDD 4 can b e regarded as P-frames while the other

The scheme allows that while a frames is b eing recon-

frames as B-frame except the rst frame. Thus, this

structed, the \late" packets of an R-frame can b e de-

scheme can b e easily incorp orated into MPEG. PTDD

co ded and used for restoring the R-frame. This will stop

do es not incur much computational overhead and do es

error propagation b ecause the next frame reconstructed

not require many additional frame bu ers. Only two

would not carry over an error from the R-frame.

additional bu ers are needed to store the R-frames of

Figure 1 illustrates the error recovery using retrans-

the next p erio dic frame.

mission in a video stream containing two packets per

frame. Packet 3 is lost, and the receiver receives packet One drawback of PTDD is that only p erio dic frames

4 at time t and recognizing that packet 3 is not re- are protected by retransmission. An error in a non- 1 Frame interval

frame display frame1 frame 2 frame 3 at the sender Time

Retransmission of packet 3 packet p1 p2 p3 p4 p5 p6 p7 p8 Time transmission times t2

lost Time Arrival t1 t3 Times send NACK Although frame 2 is displayed, the packet can be used to restore the reference frame of frame 3. Thus frame 3 is displayed without error

frame 2 has error Time Playout control time frame 1frame 2 frame 3

Times

Figure 1: The recovery of frames using retransmission

enhancement signals LP stream

frame can propagate until the next p erio dic p erio dic Video In

Q ZZ VLC

ed. In the next section, we discuss how frame is receiv Comparator

e can overcome this drawback. In fact, there is an

w - DCT essential signal HP stream

teresting tradeo between the packet deadline and in ZZ VLC Q

Inter/intra

extent of error propagation. A long TDD of pe- the -1 -1

rio d frames can prolong error propagation among non-

However, it allows more time for p e- p erio dic frames. -1 -1

DCT DCT

vered. We plan to explore this rio dic frames to b e reco Inter/intra

+ tradeo in the future. +

Prediction

3.2 Layered Co ding with REC Frame Frame buffer 1 buffer 2

Motion Estimation

The error resilience of non-p erio dic frames can b e im-

proved by employing a layered co ding scheme that pack-

etizes enco ded frames into essential or high priority

and enhancement or low priority signals. Although

Figure 4: QAL Enco der

layered co ding is originally develop ed for a network

paradigm, such as ATM and RSVP, where a certain

t of bandwidth can be reserved, it can also be

amoun High-priority data

ver a b est-e ort network such as the Internet if

used o 132 N

y forward error cor-

the HP stream can b e protected b 1 DC

rection. The HP stream should b e kept small to reduce

the bandwidth required to protect the stream b ecause

the size of redundant information intro duced by FEC is

prop ortional to the size of the HP stream. End of data

Quality assurance layering techniques guarantee a

certain level of receiving video qualityeven though all

the enhancement signals are lost. Aversion of quality

assurance layering is shown in Figure 4. It is mo di ed

from H.261 and augmented with priority packetization

and conditional replenishment. After DCT co ecients

are quantized Q, they are partitioned into HP and LP

layers as describ ed in Figure 5. For a xed nonzero in-

Figure 5: Partitioning of DCT Co ecients from [16 ]

teger b less than 64, the rst b co ecients are allo cated

to the HP layer, and the remaining co ecients are allo-

cated to the LP layer. The subsequentinter-frame and

estimation and conditional replenishment to enco de the

the currently enco ded frame are used to p erform motion

signals of reference frames, causing error propagation.

Thus, QAL alone cannot b e very e ectiveover a b est

network where no guarantee on the packet loss

HP HP HP HP e ort

b ehavior can b e made. However, QAL combined with

PTDD e ectively suppresses error propagation. The er-

LP LP LP LP rors o ccurring in non-p erio dic frames due to loss in the

HP stream do not propagate b eyond the next p erio dic

frame if the lost packets for the current p erio dic frame

periodic non-periodic non-periodic periodic

vered by retransmission b efore the reconstruc-

frame frame frame frame are reco

tion of the next p erio dic frame.

Figure 6: A Temp oral Dep endency Chain in Layered

PTDD PTDD +QAL

4 Exp erimental Result

The main ob jective of the exp erimentistoshow that

next frame. Only the motion vectors and the HP co-

PTDD is an e ective error control scheme for a real-time

ecients of the current frame are used to reconstruct

interactive video transmission over the Internet. We

a predicted frame. The di erence between the subse-

show this through video transmission exp eriments over

quent frame and the predicted frame is enco ded. In the

transatlantic and transpaci c Internet connections from

scheme, each frame temp orally dep ends only on the es-

Emory University,Atlanta, USA. By mo difying an im-

sential signals of its reference frame. Since a frame is

plementation of H.261, we implemented three variants of

reconstructed from the essential signals of its reference

PTDD HP.261, HPF.261, HPL.261, each of which dif-

frame, an error in the enhancement signals of the ref-

fers from each other in the way that non-p erio dic frames

erence frame caused by packet loss do es not carry over.

are protected. HP.261 do es not provide any protection

Thus, even if all LP stream packets are discarded, a

for non-p erio dic frames; HPF.261 protects non-p erio dic

certain level of video quality can b e maintained.

frames by adding one Exclusive-OR parity packet size

QAL, however, has not yet b een proved to give good

512 bytes to each frame; and HPL.261 combines H.261

compression eciency. The diculty lies in the tradeo

and QAL and it protects the HP stream by adding

between compression eciency and the size of redun-

one Exclusive-OR parity packet size 512 bytes to each

dant information added by FEC. If the HP stream do es

frame.

not contain enough signals, there are not many tem-

The p erformance of the three PTDD co decs is com-

p orally redundant signals b etween the reference frame

pared to that of several other co decs. H.261 is used

reconstructed from the essential signals and the current

as a base case for the comparison. H.261 is combined

frame to be enco ded. On the other hand, as we add

with a FEC technique where one Exclusive-OR parity

more co ecients to the HP stream, the HP stream gets

packet size 512 bytes is added to each frame. We call

larger, and so do es the redundant information added by

this scheme HF.261. In addition, we implement QAL

FEC to protect the HP stream. Due to low compression

without PTDD based on H.261 as describ ed in Fig-

eciency, QAL has b een traditionally used in a situa-

ure 4. This implementation is called HL.261. We also

tion where a large p ortion of bandwidth can b e allo cated

implemented INTRA-H.261 which is adopted in vic.

to the HP stream ab out 50 to 80 of the total band-

INTRA-H.261 is known for good error resilience over

width [10 ]. Under the current Internet environment,

packet loss. For each frame, INTRA-H.261 intra-co des

this may not b e reasonable.

every image blo ckchanged signi cantly from the corre-

PTDD provides a good compromise for these con-

sp onding blo ck in the previous frame. The remaining

icting requirements. Since b oth the essential and en-

blo cks are not co ded. We also implemented INTRAL-

hancement signals of p erio dic frames can be restored

H.261 which combines layered co ding with INTRA-

by retransmission, we can allow p erio dic frames to ex-

H.261 where the DCT co ecients of each co ded blo ckis

ploit the temp oral redundancy present in b oth of the

divided into the HP and LP streams. Again, one parity

essential and enhancement signals of their reference

packet is added to each frame to protect the HP stream.

frames. This increases compression eciency. Even

All the layered enco ders co de 5 DCT co ecients of each

with a small amount of bandwidth allo cated to the HP

DCT blo ck as essential signals and the remaining as en-

stream, QAL combined with PTDD can achieve good

hancement signals.

compression eciency. Note that non-p erio dic frames

use only the essential signals of their reference frames in-

cluding the ones immediately following p erio dic frames.

4.1 Testing Metho dology

So, when a p erio dic frame is displayed with some er-

ror due to packet loss in the LP stream, its dep endent

A test video sequence is obtained from a typical video

non-p erio dic frames would not b e a ected. Figure 6 il-

conferencing session where one typical \talking head"

lustrates achain of temp oral dep endency in a layered

engages in a conversation with the other party. The

PTDD scheme.

video is sampled at 5 frames/sec rate and each frame

is captured in the color CIF YUV format 352 288.

Under heavy or bursty packet loss, even the HP

This video sampling rate allows us to achieve a bit rate

stream loses packets since FEC cannot provide adequate

suitable for transatlantic and transpaci c transmission

protection for the HP stream. Unfortunately, packet

loss in the HP stream intro duces errors in the essential

uch load on the network. Con- without imp osing to o m "H.261"

"HP.261"

the long distance and limited bandwidth be- sidering "HPF.261"

34.5 "HPL.261"

ween the testing sites, this frame rate is not unusual. t "HL.261"

"INTRAL-H.261"

around 250 Kbits/sec. In addi-

The target bit rate is "INTRA-H.261"

to the controlled sample rate, we use conditional tion 34

replenishment for all the tested schemes to obtain a de-

bit rate. Adjusting quantization steps would be

sired 33.5

a more common way to control the bit rate. However,

since INTRA-H261 uses conditional replenishment, we

Average PSNR (dB)-> 33

use the same technique uniformly for all the schemes for

fairness. Finding the optimal video quality for a given

32.5

bit rate is outside the scop e of this pap er.

Ab out 40 second length video sequence total 190 32

2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 The video sequence is replayed sev-

frames is obtained. Average bytes/frame->

eral times for a ve minute p erio d for each exp eriment.

The replay do es not a ect the integrityof the exp eri-

Figure 7: Compression Eciency of Various Schemes

ment b ecause the rst frame is always completely intra-

co ded without any conditional replenishment. The

Compression Avg. bit FEC HP Avg.

95th frame is intra-co ded with conditional replenish-

scheme rate Kb/s PSNR

ment to remove any artifact due to the deco der drift

H.261 240.6 0 0 34.50

e ect. All the co decs applies a default quantization step

HF.261 245.0 8.4 0 34.49

size 8, and all the motion comp ensated co decs use the

HP.261 232.6 0 0 34.51

full exhaustive searchover search window size 15 by 15.

HPF.261 234.2 8.6 0 34.41

All PTDD co decs set the TDD of all the p erio dic frames

HL.261 252 8 33 33.97

to b e 5 frame intervals. Given 5 frames/s frame rate,

HPL.261 239.3 8.4 26 34.20

this TDD extends the deadlines of p erio dic frames up

INTRA-H.261 247.7 0 0 33.65

to 1 sec.

INTRAL-H.261 252.7 8.1 27 33.49

To see the compression eciency of various co decs

for the input test sequence, we measure the average

Table 1: Chosen data rates for network exp eriments,

p eak signal-to-noise ratio PSNR of deco ded frames

and their average PSNR's

over various data rates. The data rate is measured

by the average number of bytes required to compress

a frame Figure 7. It is clearly shown that for a given

rate sp eci ed in Table 1. The video quality of HF.261 is

data rate, INTRA-H.261 and INTRAL-H.261 show the

similar to HP.261, and is not shown here. Note that the

worst video quality while H.261 shows the b est. To

mean PSNR of HP.261 for the given data rates is 0.1 dB

achieve the same quality, INTRA-H.261 requires more

higher than that of H.261. This is due to the artifact

bandwidth than the other motion comp ensation-based

of the enco der. In general, PTDD gives slightly lower

co decs. For instance, to obtain ab out 34 dB PSNR,

compression eciency than H.261 as shown in Figure 7.

INTRA-H.261 requires 80 11KB/6KB more bits p er

The test video sequence is rst compressed us-

frame than H.261. It is known that more than 1 dB dif-

ing each compression scheme and then packetized into

ference in video quality is clearly visible byhuman eyes.

approximately 512-byte packets. The packetized se-

Although these measurements alone do not prove the

quences of all the PTDD schemes are transmitted over

low compression eciency of INTRA-H.261, it clearly

the Internet. For each transmission test, we obtain a 5

indicates one asp ect of that.

minute trace that records the sequence numb ers and ar-

HL.261 gives only slightly higher compression e-

rival times of all the received packets. The transmission

ciency than INTRA-H.261. This is b ecause the HP

tests are conducted during the p erio ds b etween Oct. 20

stream of HL.261 do es not contain many co ecients,

and only the essential signals of reference frames are

35 On the other used for motion comp ensation in HL.261. "H.261" 34.8 "HP.261"

"HPL.261"

HPL.261 gives excellent compression eciency. hand, "INTRA-H.261"

34.6 In HPL.261, although non-p erio dic frames are enco ded

34.4 in the same as in HL.261, p erio dic frames are motion-

34.2

comp ensated to the full signals of its reference frame,

resulting in higher compression eciency. PSNR->

33.8

Table 1 shows the chosen data rate of each co dec for

33.6 transmission along with the ratio of the bit rate of the

33.4

HP stream over the total bit rate, and that of the re-

33.2

dundant bit rate induced by FEC. HP.261 and HPL.261

en a little lower data rates b ecause the asso ciated

are giv 0 20 40 60 80 100 120 140 160 180 200

Frame #->

retransmission of lost packets may increase the actual

data rate. Figure 8 shows the PSNR of each frame com-

Figure 8: Video quality of enco ded sequences

pressed by four di erentschemes under the target bit packet frame packet # received received packet frame packet # packet frame packet # received received number number in a frame Y/N time (ms) number number in a frame number number in a frameY/N time (ms) 0 1 0 Y 201 0 1 0 0 1 0 Y 201 1 1 1 Y 450 1 1 1 1 1 1 Y 450 2 1 2 Y 600 2 1 2 2 1 2 Y 600 3 1 3 N * 3 1 3 3 1 3 N * 4 1 4 Y 980 4 1 4 4 1 4 Y 980 5 2 0 Y 1100 5 2 0 5 2 0 Y 1100 6 2 1 Y 1305 6 3 1 6 3 1 Y 1305 7 3 0 N * 7 2 2 7 2 2 N * 8 3 1 Y 1790 8 2 3 8 2 3 Y 1790 9 3 2 N * 9 3 0 9 3 0 N * 10 3 3 Y 2100 10 3 1 10 3 1 Y 2100 11 4 0 N * 11 3 2 11 3 2 N * 12 4 1 N * 12 3 3 12 3 3 N * 13 5 0 Y 2710 13 4 0 13 4 0 Y 2710 14 5 1 Y 3000 14 4 1 14 4 1 Y 3000

(a) (b) (c)

Figure 9: c is the result of mapping an actual trace a to a packet sequence b.

a packetized sequence S for those schemes as if the se- to Oct. 27, 1997, and b etween Jan 30 to Feb 6, 1998.

quence would have b een transmitted in a real test. Each

Each packet of a frame is transmitted at a regular

packet p in trace T is mapp ed to a packet q that has

interval determined by the 5 frames/s frame rate and

the same sequence number as p. If packet p is received,

the numb er of packets within the frame. For example,

we record q as received and assign the receiving time of

for the frame interval of 200 ms, if one frame consists of

p to q . Otherwise, we record q as lost. Figure 9 illus-

10 packets, a packet in the frame is transmitted at 20 ms

trates this mapping. Figure 9a shows a sample trace

interval. Each transmitted packet is assigned a unique

ofaHP.261 sequence `' indicates the packet is not re-

sequence numb er. Retransmitted packets are given the

ceived. Those packets received indicated byYshow

same sequence numb ers as their original packets.

received times. Figure 9 b shows a sample of a pack-

etized H.261 sequence. Figure 9 c shows the result The ARQscheme employed during the exp eriment

of mapping a to b. This mapping technique pro- works as follows. The receiver sends one acknowledg-

vides a very accurate comparison of various transmis- ment to the sender for each received frame. An ac-

sion schemes b ecause the sequences of all the schemes knowledgment contains information ab out the missing

are mapp ed to the same traces. packets of the last p erio dic frame that the receiver re-

ceived. After retransmitting a packet, the sender do es

This mapping technique cannot capture the dynam-

not retransmit the same packet for ab out three frame

ics of acknowledgments and retransmissions in HP.261,

intervals. It may retransmit the packet if it receives an-

HPL.261 and HPF.261. This is b ecause in the exp er-

other acknowledgment after the p erio d indicating that

iment, the acknowledgment is transmitted by the re-

the packet is lost. The receiver also do es not request

ceiver only when a new frame is received and each se-

for the retransmission of packets whose deadlines are

quence may have a di erent number of packets in a

expired. These mechanisms reduce the numb er of un-

frame. Thus, acknowledgments and retransmitted pack-

necessary retransmissions.

ets might b e received at di erent times for di erent se-

quences. The dynamics can only b e captured through a Each trace is fed to an o -line deco der to measure

real transmission, whichiswhy the sequences of HP.261, the signal-to-noise of the received frames. To simplify

HPF.261, and HPL.261 are actually transmitted. The the exp eriment, we did not add any jitter control time

other schemes do not have this problem b ecause they for frame play-out. Each frame is considered to b e dis-

do not involveany retransmission. played at the arrival of the rst packet of its next frame

if that packet is received. If that packet is not received,

We obtained 168 traces of HP.261, 147 traces of

the frame is considered to b e displayed at 200 ms after

HPF.261, and 55 traces for HPL.261 Since many traces

its previous frame's play-out time. If no packet for the

are obtained, it is dicult to present the result of each

frame is received, any frame displayed last will b e dis-

trace indep endently. So we classify the traces into sev-

played for that frame. Retransmitted packets are not

eral loss rate groups and present only the average b e-

used for the displayof their frames, but used only to

havior of the traces in each group. Table 2 shows the

restore their corresp onding reference frames.

loss groups and their corresp onding loss ranges. Since

high loss cases are relatively infrequent, we set a larger For fair comparison, we run trace-driven simulations

range for high loss rates. by mapping each of the traces T to the packetized se-

quence of H.261, HL.261, INTRAL-H.261 and INTRA-

H.261 as follows. We rst obtain a 5 minute length of

Loss group 0.025 0.05 0.075 0.1 0.125 0.15

Loss range 0, 0.025 [0.025,0.05 [0.05,0.075 [0.075,0.1 [0.1, 0.125 [0.125, 0.15

Loss group 0.175 0.2 0.25 0.3 0.35 0.4

Loss range [0.15, 0.175 [0.175,0.2 [0.2, 0.25 [0.25,0.30 [0.3, 0.35 [0.35,0.40

Table 2: Loss Rate Groups and their loss ranges

ket loss, the mean PSNR of HP.261 drops "HP.261" and 20 pac

34 "INTRA-H.261"

w that of INTRA-H.261. This is due to the drop in "H.261" b elo

the REC recovery rates.

32 e can see that the PSNR of HPF.261

31 In Figure 11, w

is slightly improved from that of HP.261. HPF.261 gives

slightly b etter PSNR than INTRA-H.261 for all loss

groups except two. This improvement is mainly due 28

PSNR ->

to FEC. However, the improvement is still marginal.

Although many packets can b e recovered through FEC

ket loss, as packet loss gets severer, FEC

26 under small pac

e. From Table 3, under less than 10

25 b ecomes ine ectiv

ket loss, FEC could recover from 30 to 50 packet

24 pac

However, as more packets are dropp ed, the FEC

23 lost.

very rate drops b elow 10. 22 reco 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

loss rate (x 100%) ->

There is a clear correlation b etween the round trip

time and REC recovery rates. As the round trip time

Figure 10: Mean PSNR of H.261, INTRA-H.261 and

delays increase, the recovery rates of p erio dic frames

HP.261

also drop. When round trip times increase beyond

250ms, the recovery rates by REC are signi cantly re-

This is b ecause the increased network delay re- "HPF.261" duced.

34 "INTRA-H.261"

the probability for retransmitted packets to be "H.261" duces

received b efore their deadlines. Second, in HP.261,

there is no protection for non-p erio dic frames against

packet loss, and in HPF.261, the protection by FEC

as not so helpful in recovering lost packets for non-

30 w

Thus, as more packets are lost, more 29 p erio dic frames.

28 non-p erio dic frames su er from error propagation. PSNR ->

Overall, although the HP.261 and HPF.261 show

ver INTRA-h.261, the improve-

26 b etter error resilience o

t can be considered marginal. This is b ecause

25 men

hemes, little protection is provided for non-

24 in b oth sc

frames. Packet loss in a non-p erio dic frame

23 p erio dic

error propagation which is continued until the 22 starts

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35

p erio dic frame is received. The resulting video

loss rate -> next

quality uctuates substantially. This problem disap-

p ears when PTDD is combined with a b etter FEC

Figure 11: Mean PSNR's of H.261, INTRA-H.261, and

scheme such as QAL.

HPF.261

4.3 Performance of PTDD + QAL + FEC HPL.261

4.2 Performance of PTDD HP.261, and PTDD +

FEC HPF261

In this section, we rep ort the result obtained from the

HPL.261 traces. HPL.261 combines PTDD and QAL.

In this section, we rep ort the results obtained from the

The HP stream generated from QAL is protected by

HP.261 and HPF.261 traces. The sequences of H.261,

FEC. We ran trace-driven simulations of H.261, HL.261,

HF.261 and INTRA-H.261 are mapp ed to the traces

INTRA-H.261, and INTRAL-H.261 based on the traces

of HP.261 and HPF.261. Figures 10 and 11 show the

of HPL.261. Table 4 summarizes the result. Figure 12

average PSNR's of H.261, INTRA-H.261 and HP.261

shows the average PSNR's of various schemes. Wedo

for various loss groups. Table 3 summarizes the result

not have traces for loss groups higher than 20.

of H.261, HF.261, HPF.261, and INTRA-H.261.

In Figure 12, HPL.261 shows go o d error resilience,

In Figures 10 and 11, the mean PSNR of H.261 drops

showing clear separation ab out 1 dB from INTRA-

drastically even under small packet loss, showing the

H.261. 1 dB or higher PSNR di erence in video quality

e ect of error propagation. INTRA-H.261, HP.261, and

is clearly visible byhuman eyes. The go o d p erformance

HPF.261 exhibit generally go o d error resilience. They

of HPL.261 can b e b est explained from Figure 13. As

all show similar PSNRs for all loss groups. Between 12

Loss of H.261 HPF.261 INTRA-H.261 HF.261

Rate traces PSNR D. R. PSNR D. R. RTT PSNR D. R. PSNR D.R.

dB Kb/s dB Kb/s REC FEC ms dB Kb/s db Kb/s

2.5 24 32.91 247.0 34.16 247.9 89.7 48.88 188.5 33.12 259.0 33.62 246.4

5.0 42 30.39 246.5 33.48 249.1 96.21 42.23 197.7 32.46 258.5 31.66 246.0

7.5 32 28.72 245.3 32.71 249.6 96.64 33.77 223.5 31.83 257.3 30.07 244.8

10.0 14 28.17 245.7 32.13 250.8 84.56 24.46 239.1 31.41 257. 6 29.24 245.1

12.5 14 26.73 249.7 31.21 261.9 81.70 18.77 325.1 30.80 259. 7 27.55 249.5

15.0 9 26.58 249.5 30.60 263.3 57.94 11.78 388.1 30.4 259.0 27.26 249.9

17.5 6 25.45 250.2 29.19 263.7 50.74 9.34 437.7 29.94 259.3 25.91 250.6

25.0 5 26.27 250 29.31 261.7 41.9 5.11 359.3 29.53 260.1 2 6.69 249.8

35.0 1 25.68 239.8 29.35 245.4 24.96 6.14 243.5 28.84 251.5 26.9 239.3

Table 3: Exp erimental Data based on HPF.261 traces

Loss of INTRA-H.261 INTRAL-H.261 H.261 HPL.261 HL.261

Rate traces D.R. D. R. FEC D.R. D.R. REC FEC RTT D.R. FEC

Kb/s Kb/s Kb/s Kb/s ms Kb/s

2.5 22 249 243 60.7 243 238 89.9 61.3 242.6 243 63.5

5 16 247 240 59.8 240.8 239 78.0 57.8 280.4 238 57.9

7.5 7 249 242 52.4 242 240 73.2 52.1 310.5 240 54.4

10 3 249 241 35.5 240 240 57 39.1 305.6 239 41.7

12.5 2 249 243 48.5 242 245 84.4 48.7 347 244 54.5

Table 4: Exp erimental Data based on HPL.261 traces

35 35 "HPL.261" "HPL.261" 34 "INTRA-H.261" "HPF.261" "INTRAL-H.261" "HP.261" 33 "HL.261" 34 "INTRA-H.261" "H.261" 32 33 31

30 32 29

28 PSNR -> PSNR -> 31 27

26 30 25

24 29 23

22 28 0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2

loss rate -> loss rate ->

Figure 12: Mean PSNR's of H.261, HL.261, HPL.261, Figure 14: PSNR vs. Loss Rate Comparison of HP.261,

INTRA-H.261, and INTRAL-H.261 HPF.261, and HPL.261

in HP.261, the PSNR of HPL.261 drops when there is

not contain enough information to improve the receiv-

heavy loss for a frame. However, it quickly b ounces

ing video qualityeven though they are successfully re-

back when the loss rate of the subsequent frames quickly

ceived. Second, the compression eciency of INTRAL-

reduces. This is unlikeHP.261 where reb ound happ ens

H.261 is lower than that of INTRA-H.261. Thus, under

mostly around p erio dic frames. We clearly see many

no or little loss, its average PSNR is always a bit lower

plateaus indicating that packet loss do es not havemuch

than that of INTRA-H.261. The PSNR di erence b e-

impact on the video quality of non-p erio dic frames.

tween INTRAL-H.261 and INTRA-H.261 over various

loss groups seems to remain constant b ecause of these

HL.261 p erforms quite well under low packet loss.

two reasons.

As packet loss gets substantial, The PSNR of HL.261

quickly drops. This is b ecause b eyond low loss, the FEC

We also compare the mean PSNRs of all the PTDD

scheme protecting the HP stream b ecomes ine ective

schemes presented in this pap er in Figure 14. We use the

and the HP stream starts losing packets causing error

mean PSNR of INTRA-H.261 obtained from the traces

propagation.

of HPL.261 as a reference for comparison. From the g-

ure, we can see that HPL.261 seems to p erform b est. At

Surprisingly, INTRAL-H.261 p erforms slightly worse

the b eginning, when there is no or little loss, HPL.261

than INTRA-H.261. This can b e explained in twoways.

gives the worst PSNR among REC-based co decs. How-

First, the HP stream contains have only 5 DCT co ef-

ever, as packet loss gets larger, HP.261 and HPF.261

cients from each co ded blo ck. These co ecients do 40 "HPL.261" "H.261"

30 PSNR ->

20 6000 7000 8000 9000 10000 11000 12000 13000 14000 15000 sequence # -> 100 "lossrate"

loss rate -> 40

0 6000 7000 8000 9000 10000 11000 12000 13000 14000 15000

packet sequence # ->

Figure 13: One trace of HPL.261 with 10 packet loss { PSNR, and Loss and Recovery Rates

p ointed out in [23 ], in a multicast group, while some re- quickly drops while HPL.261 sustains go o d video qual-

ceivers want real-time video, others maywanttowatch ity.

or record the transmitted video. These observers may

want the highest quality that the video source could pro-

5 Discussion

vide although they can tolerate a longer play-out delay.

At the same time, they mayhave only a small amount

of bandwidth allo cated for the video.

Retransmission has widely b een known ine ective for

recovering lost packets in real-time interactive video ap-

Motion-comp ensated enco ding allows much b etter

plications. This pap er challenges the conventional wis-

video quality for a given data rate than intra-co ding.

dom and presents new retransmission-based error con-

Our scheme allows real-time receivers to view video in

trol techniques that do not incur any additional latency

a comparable qualityasintra-co ding schemes. At the

in the frame play-out time, and hence are suitable for

same time, it allows the same enco ded video to b e multi-

interactive applications.

casted to other non-real-time receivers. These receivers

can view the video in very high qualityby adding an

One main implication of our work is that many

additional delay i.e., control time b efore the display

motion comp ensation prediction-based co decs, suchas

of the rst frame. On the other hand, for intra-co ded

MPEG, and H.261, are still useful for Internet inter-

video to b e sentonalow bandwidth network, its qual-

active video transmission over a lossy network. Some

ity needs to b e reduced substantially. In this case, al-

of the disadvantages of the motion comp ensated co decs

though real-time receivers may get the video in similar

cited in the literature [14 , 13 ] include 1 computational

quality as the motion-comp ensation scheme with REC,

complexity, 2 error resilience, 3 tight coupling be-

non-real-time receivers will receive a p o or quality video.

tween the prediction state at the enco der and that at the

Note that this feature is di erent from media scaling [14 ]

deco der, and 4 compute-scalable deco ding. In this pa-

where receivers with a higher network capacity always

p er, we showed that the H.261 equipp ed with our REC

get a higher quality image. Here, this feature allows

schemes achieves comparable error resilience to that of

even the receivers with a low network capacity to get

INTRA-H.261. We also b elieve that some of the other

a high quality while trading interact-ability. The mo-

disadvantages can b e overcome with a simple mo di ca-

tion comp ensated co decs equipp ed with a REC scheme

tion to the co decs. For instance, the compute-scalable

can provide this feature as they generally give b etter

deco ding can also b e achieved by deco ding only p erio dic

compression eciency and lost packets can b e recovered

frames and shedding o the computational load for de-

through retransmission.

co ding non-p erio dic frames in PTDD.

Having said some of the disadvantages of motion-

References

comp ensated co decs, we would like to emphasize one

of their advantages over INTRA-H.261, which is high

[1] A. Albanese, J. Blomer, J. Edmonds, M. Luby, compression eciency. Although INTRA-H.261 gives

and M. Sudan, \Priority Enco ding Transmission," go o d error resilience, the low compression eciency of

IEEE Transactions on Information Theory,vol. 42, INTRA-H.261 will makeitvery dicult to obtain very

no. 6, Nov. 1996. high quality video for a low bit rate transmission. As

[16] N. Ohta, \Packet Video: Mo deling and Signal Pro- [2] Amir, E., McCanne, S., and Vetterli, M. \A Lay-

cessing," Artech House Publishers, 1994. ered DCT Co der for Internet Video," the Proceed-

ings of IEEE International Conference on Image

[17] P.Pancha and M. El Zarki,\MPEG Co ded Video

Processing. September, 1996. Lausanne, Switzer-

and BISDN," IEEE Communications Magazine,

land.

Vol. 32, No. 5, May 1994.

[3] P. Brady, \E ects of Transmission Delay on Con-

[18] C. Papadop oulos and G. Parulkar,

versational Behavior on Echo-Free Telephone Cir-

\Retransmission-based error control for continuous

cuits," Bel l System Technical Journal,vol. 50, no.

media applications," Proceedings of the Sixth Inter-

1, pp. 115 { 134, Jan. 1993.

national Workshop on Network and Operating Sys-

[4] B. DeCleene, P. Pancha, M. El Zarki, and H.

tem Support for Digital Audio and Video, pp. 5{12,

Sorensen, \Comparison of PriorityPartition Meth-

1996.

ods for VBR MPEG," Proceedings of IEEE Info-

[19] Priority Enco ding Transmission, Web Page:

com'94, Toronto, Canada, June 1994.

http://www.icsi.b erkeley.edu/PET/icsi-p et.html.

[5] B. Demsey, J. Lieb eherr, and A. Weaver, \On

retransmission-based error control for continuous

[20] E. J. Posnak, S. P. Gallindo, A. P. Stephens and H.

media trac in packet-switching networks," Com-

M. Vin. \Techniques for ResilientTransmission of

puter Networks and ISDN Systems, 1996.

JPEG Video Streams" the Proceedings of Multime-

dia Computing and Networking, San Jose, Volume

[6] T. Dorcey, \CU-SeeMe desktop video conferencing

2417, Pages 243-252, February 1995.

software," ConneXions, vol. 9, no. 3, Mar. 1995.

[21] G. Ramamurthy, and D. Raychaudhuri, \Perfor-

[7] R. Fredrick, Network Videonv, XeroxPalo Alto

mance of packet video with combined error re-

Research Center.

covery and concealment," Proceedings of INFO-

COM'95, pp. 753{761, April 1995.

[8] M. Ghanbari, \Twolayer co ding of video signals for

VBR networks," IEEE Journal on Selected Areas

[22] K. Shimamura, Y. Hayashi, and F. Kishino,

in Communications,vol. 7, no. 5, June 1989.

\Variable-bit-rate Co ding Capable of Comp ensat-

ing for Packet Loss," Proceedings of Visual Com- [9] K. Han and G. C. Polyzos, \Multi-Resolution Lay-

munications and Image Processing, Nov. 1988, pp. ered Co ding for Real-Time Image Transmission:

991 { 998. Architectural and Error Control Considerations,"

Real-Time Imaging," Academic Press to app ear.

[23] R. Xu, C. Myers, H. Zhang, R. Yavatkar, \Re-

[10] M. Khansari, A. Jalali, E. Dub ois, and P. Mermel-

silientmulticast supp ort for continuous-media ap-

stein, \Low bit-rate video transmission over fading

plications," Proceedings of the Sixth International

channels for wireless micro cellular systems," IEEE

Workshop on Network and Operating System Sup-

Transactions on Circuits and Systems for Video

port for Digital Audio and Video, St Louis, May

Technology,vol. 6, pp. 1-11, February 1996.

1997.

[11] C. Leicher, \Hierarchical enco ding of MPEG

sequences using priority enco ding transmission

PET," ICSI technical rep ort TR-94-058, Nov.

1994.

[12] X. Li, S. Paul, P.Pancha, and M. Ammar, \Lay-

ered video multicast with retransmission LVMR:

evaluation of error recovery schemes," Proceedings

of the Sixth International Workshop on Network

and Operating System Support for Digital Audio

and Video, St Louis, May 1997.

[13] S. McCanne and V. Jacobson, \vic: a exible

framework for packet video," in Proceedings of

ACM Multimedia'95, San Francisco, CA, Nov.

1995, pp. 511{522

[14] S. McCanne, M. Vetterli, and V. Jacobson, \Low-

complexity video co ding for receiver-driven lay-

ered multicast," IEEE Journal on Selected Areas

in Communications, vol. 16, no. 6, pp. 983-1001,

August 1997

[15] M. Normura, T. Fujii, N. Ohta, \Layered Packet-

Loss Protection for Variable Rate Co ding Using

DCT," Proceedings of International Workshop on

Packet Video, D7, Sept. 1988.