Application and Network Measurements For Multiple ... - CiteSeerX

Application and Network Measurements For Multiple Description Coded Video Sequences F. Fitzek ∗ Aalborg University Head of Future Visions Niels Jernes Vej 12, room A5 204 DK-9220 Aalborg Denmark Phone: +45 9635 8678 Fax: +45 9815 1583 email: [email protected] July 13, 2004 Technical Report R-04-1005 ISBN 87-90834-46-1 ISNN 0908-1224 Multiple description coding has received a lot of interest lately, because it allows to split one source of information in multiple entities, where each of the entities is decodable. The more of these entities are received the more information of the original source can be restored. This opens the door for new applications in many fields of mobile communications. Unfortunately the multiple description coding comes with the price of higher bandwidth requirements. This overhead is caused by the coding process itself and the related network transport. In this work we present a solid investigation of the overhead arising from the splitting process for video coding. Several well known and new video sequences in the QCIF and CIF video format were encoded for multiple descriptors using H.263 and H.26L encoder. The overhead caused by the encoding and underlying network protocols is given separately. Furthermore we investigate the quality of the decoded video in dependency of the received percentage of descriptors. Our results show that the overhead depends dramatically on the video codec chosen and that for eight descriptors the overhead differs from 5% to 200%. A main conclusion we draw here is that we have to consider the scenario carefully whether or not multiple description coding should be chosen over single description coding. ∗

The work of F. Fitzek was supported by the SAMSUNG project Joint Advanced Development Enabling 4G (JADE).

c

Aalborg University.

R-04-1005

Page 1

Contents 1 Introduction 1.1 Structure Of This Document . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10 12

2 Methodology and Measurement Setup

13

3 Metric and Terminology

16

4 Coding Overhead Measurements 4.1 H.263 . . . . . . . . . . . . . . . . . . 4.1.1 QCIF . . . . . . . . . . . . . . 4.1.2 CIF . . . . . . . . . . . . . . . 4.2 H.26L . . . . . . . . . . . . . . . . . . 4.2.1 QCIF . . . . . . . . . . . . . . 4.2.2 CIF . . . . . . . . . . . . . . . 4.3 Impact of the Quantization Parameter

18 18 18 18 18 21 22 22

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

5 Network Overhead

25

6 Quality Measurements

28

7 Conclusion

31

8 Acknowledgment

31

A Additional Overhead Measurements A.1 H263 . . . . . . . . . . . . . . . . A.1.1 QCIF . . . . . . . . . . . A.1.2 CIF . . . . . . . . . . . . A.2 H26L . . . . . . . . . . . . . . . . A.2.1 QCIF . . . . . . . . . . . A.2.2 CIF . . . . . . . . . . . .

34 34 34 37 37 39 42

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

B Additional Quantization Measurements

44

C Additional Quality Measurements

50

c

Aalborg University.

R-04-1005

Page 2

List of Figures 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

The carphone sequence in the QCIF format. . . . . . . . . . . . . . . . . . . . . . The bridge–close sequence in the CIF format. . . . . . . . . . . . . . . . . . . . . Splitter and encoding chain for J=3. . . . . . . . . . . . . . . . . . . . . . . . . . Merger for J=3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Merger for J=3, where one descriptor is missing. . . . . . . . . . . . . . . . . . . Overhead of selected H.263 video sequences in the QCIF format for different number of sub–streams. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Number of INTRA, INTER, and INTER4V macro-blocks per frame for the bridgefar video sequence in the QCIF format. . . . . . . . . . . . . . . . . . . . . . . . Overhead of selected H263 video sequences in the CIF format for different number of sub–streams. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Number of INTRA, INTER, and INTER4V macro-blocks per frame for the bridgefar video sequence in the CIF format. . . . . . . . . . . . . . . . . . . . . . . . . Overhead of selected H.26L video sequences in the QCIF format for different number of sub–streams. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Percentage of I frames in one descriptor for different video lengths. . . . . . . . . Overhead of selected H26L video sequences in the CIF format for different number of sub–streams. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the container video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the container video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the foreman video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the foreman video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Transport of multiple description coded video and voice over RTP, UDP, and IP network layer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Network Overhead (RTP/UDP/IPv4) for the foreman video sequence and two different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Network Overhead (RTP/UDP/IPv6) for the foreman video sequence and two different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PSNR measurements using the foreman video sequences in the QCIF format versus percentage of successful received sub–streams (random approach). . . . . . . . . PSNR measurements using the foreman video sequences in the QCIF format versus percentage of successful received sub–streams (best case). . . . . . . . . . . . . . PSNR measurements using the foreman video sequences in the QCIF format versus percentage of successful received sub–streams (worst case). . . . . . . . . . . . . . PSNR values versus time for the bad case and a random case. . . . . . . . . . . . Comparison of PSNR degradation for different video sequences for J = 15. . . . . Overhead of multiple description coded video sequence bridge-close in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence bridge-far in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . .

c

Aalborg University.

R-04-1005

13 13 14 15 15 19 19 20 20 21 22 23 23 23 24 24 25 26 27 28 29 30 30 31 34 34

Page 3

27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49

Overhead of multiple description coded video sequence carphone in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence claire in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence container in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence foreman in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence grandma in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence highway in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence mthr dotr in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence news in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence salesman in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence silent in the QCIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence bridge-close in the CIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence bridge-far in the CIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence highway in the CIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence mobile in the CIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence paris in the CIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence tempete in the CIF format using the H.263 encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence bridge-close in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence bridge-far in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence carphone in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence claire in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence container in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence foreman in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence grandma in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . .

c

Aalborg University.

R-04-1005

34 34 35 35 35 35 36 36 36 36 37 37 37 37 38 38 39 39 39 39 40 40 40

Page 4

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72

Overhead of multiple description coded video sequence highway in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence mthr dotr in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence news in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence salesman in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence silent in the QCIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence bridge-close in the CIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence bridge-far in the CIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence highway in the CIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence mobile in the CIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence parisparis in the CIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead of multiple description coded video sequence tempetetempete in the CIF format using the H.26L encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the bridge-close video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the bridge-close video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the bridge-far video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the bridge-far video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the carphone video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the carphone video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the claire video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the claire video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the container video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the container video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the foreman video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the foreman video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

c

Aalborg University.

R-04-1005

40 41 41 41 41 42 42 42 42 43 43 44 44 44 44 45 45 45 45 46 46 46 46

Page 5

73 74 75 76 77 78 79 80 81 82 83 84 85

86

87

88

89

90

91

Overhead for the grandma video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the grandma video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the highway video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the highway video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the mthr dotr video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the mthr dotr video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the news video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the news video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the salesman video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the salesman video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overhead for the silent video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bandwidth requirements for the silent video sequence in the QCIF format for different quantization values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 2 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 3 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 5 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 10 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 15 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 20 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 2 . . . . . . . . . . . . . . . . . . . . . . . . .

c

Aalborg University.

R-04-1005

47 47 47 47 48 48 48 48 49 49 49 49

50

51

51

52

52

53

53

Page 6

92

93

94

95

96

97

98

99

100

101

102

103

104

105

106

PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 3 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 5 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 10 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 15 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 20 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 2 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 3 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 5 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 10 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 15 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 20 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 2 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 3 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 5 . . . . . . . . . . . . . . . . . . . . . . . . . PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 10 . . . . . . . . . . . . . . . . . . . . . . . . .

c

Aalborg University.

R-04-1005

54

54

55

55

56

56

57

57

58

58

59

59

60

60

61

Page 7

107 PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 15 . . . . . . . . . . . . . . . . . . . . . . . . . 108 PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 20 . . . . . . . . . . . . . . . . . . . . . . . . . 109 PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 2 . . . . . . . . . . . . . . . . . . . . . . . . . 110 PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 3 . . . . . . . . . . . . . . . . . . . . . . . . . 111 PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 5 . . . . . . . . . . . . . . . . . . . . . . . . . 112 PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 10 . . . . . . . . . . . . . . . . . . . . . . . . . 113 PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 15 . . . . . . . . . . . . . . . . . . . . . . . . . 114 PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 20 . . . . . . . . . . . . . . . . . . . . . . . . . 115 PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 2 . . . . . . . . . . . . . . . . . . . . . . . . . 116 PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 3 . . . . . . . . . . . . . . . . . . . . . . . . . 117 PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 5 . . . . . . . . . . . . . . . . . . . . . . . . . 118 PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 10 . . . . . . . . . . . . . . . . . . . . . . . . . 119 PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 15 . . . . . . . . . . . . . . . . . . . . . . . . . 120 PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 20 . . . . . . . . . . . . . . . . . . . . . . . . .

61

62

62

63

63

64

64

65

65

66

66

67

67

68

List of Tables 1

YUV QCIF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

c

Aalborg University.

R-04-1005

13

Page 8

2

YUV CIF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

c

Aalborg University.

R-04-1005

14

Page 9

1 Introduction MDC has emerged as an attractive coding scheme for robust transmission over channels with transient bursty error characteristics such as wireless systems during a deep fade. MDC is a great solution to this problem at the price of redundancy. The packet losses, which is most probable and can be correlated over wireless links, can cause severe degradation in the received video quality. The application of MDC can be classified into two main groups, namely path diversity and heterogeneous terminals. We give a short introduction of the related work. Path diversity Various network infrastructures such as multi–cast services in cellular networks, multi–hop networks and wired networks utilize MDC to provide path diversity. The study carried out in [1] combats errors encountered in the wireless channel by coding the highly compressed video data into multiple independently decodable streams send over different paths. An unbalanced MDC scheme to provide path diversity is proposed in [2]. This scheme creates unbalanced Multiple Description (MD) streams by adjusting the frame rate, quantization or spatial resolution of each stream and can achieve unbalanced rates of up to 2:1. Authors compare the conventional MDC system with their proposed system. Their results show that unbalanced MDC scheme provides improved reliability over multiple paths with unequal bandwidths. Authors examine the effectiveness of unbalanced MDC and path diversity for transmission of foreman and bus sequences. The recovered video quality for unbalanced rates of up to 1.75:1 and 1.76:1 are compared with that of SD scheme. Their results show that the proposed unbalanced MDC scheme is more robust to bursty errors and network congestions than SD scheme. The effectiveness of the combination of MDC with multiple path transport (MPT) for video qand image transmission over multi–hop wireless links is studied in [3]. Some applications have high bandwidth and stringent reliability requirements. For supporting these kind of applications, the sender may need to send the data stream over several paths to the destination [3]. The study in [3] addresses this need and the authors compare their proposed MDC–MPT system with a system which uses layered coding and asymmetrical paths for the base and enhancement layers. Apart from the standard Quarter Common Intermediate Format (QCIF) and Common Intermediate Format (CIF) sequences given in [4], authors present average Picture Signal to Noise Ratio (PSNR) for MDC for Susie, Football and Flower Garden video sequences. In [5], an MDC system for multi–hop networks is proposed. The application and network layers cooperate to provide more robustness against severe network conditions. The authors propose a multi–path selection method that chooses a set of paths to achieve more robust media transmission. The media content is transmitted over these intelligently selected multiple paths instead of shortest path or maximally disjoint paths. The simulation results in this paper show an average PSNR improvement of up to 8.1 dB than conventional MDC scheme. The challenge that the authors faced in this paper was finding a set of paths to minimize the cost function that estimated average streaming distortion in terms of network statistics, media characteristics and application requirements [6]. To remedy this problem, they proposed a fast heuristics–based solution by exploiting the infrastructure features of the internet in [6]. Their simulations run over several random internet topologies show that the proposed heuristic approach is capable of finding a good set of paths in a much shorter time than the methods they used in [5]. This heuristic is therefore very well suited for time critical applications such as VoIP and video conferencing. The application of MDC for multi–cast networks has been addressed in [7]. Multi–cast services addresses to a wide variety of receiver terminals having different bandwidth requirements. The study in [7] explores layered multiple description codes where base layer descriptions are trans-

c

Aalborg University.

R-04-1005

Page 10

mitted to users having low bandwidth capacities whereas the enhancement layer descriptions are transmitted to users having high bandwidth capacities. MDC is also being exploited in wired infrastructures for increased link capacity and reliability. For example, the wired connection to the base stations can have more capacity if MDC is used. MDC divides the data stream into several streams whereby the probability of buffer flow in wired links decreases [8]. Multiple path transport (MPT) scheme is proposed for wired links to provide path diversity. The source coding and traffic splitting is jointly designed for MPT schemes [9]. MDC and layered coding (LC) are the two options for source coding in MPT scheme [9]. Heterogeneous terminals MDC is exploited to support heterogeneous terminals. 4G mobile systems are envisioned to offer wireless services to a wide variety of mobile terminals ranging from cellular phones and Personal Digital Assistants (PDAs) to Laptops [10]. The flexibility of the bandwidth assigned to each descriptor and the number of descriptors assigned to end users makes MDC a very attractive coding scheme for 4G networks. Terminals use the same bandwidth and decide on the number of descriptors that they can receive [11]. In [12], a new MDC technique called Multiple Description Scalable Coding (MDSC) is proposed. MDSC can simply be described as a combination of MDC and scalable coding. This technique addresses receiving device characteristics and bandwidth variations of the channel and also enables tradeoffs between throughput, redundancy and complexity which is not possible with non–scalable MDC schemes. In MDSC, the number and the composition of descriptions are changed dynamically to make the proposed system very robust to changing channel characteristics [12]. For example, in [?] layered multiple description scheme is proposed. Layered codes incorporate sequences of layers in order to provide bandwidth heterogeneity and to cope with dynamic network congestion [?]. Authors use base layer descriptions for low bandwidth clients whereas they use both base and enhancement layer descriptions for high bandwidth clients. Therefore, we give plots of overhead versus sub–streams generated, video content or number of sub–streams successfully received. Other Work in this field and their shortcomings! In [13], the performance of a single stream based voice communication is compared with that of multiple description coding (multistream communications) using Forward Error Correction (FEC) scheme. Therefore, the payload data rate of single stream and multistream schemes are identical and MDC does not increase the overall data rate of the payload compared to single stream scheme. In this way, the added overhead of packet headers as a result of transmitting multiple streams are compared. The motion vector data is the most important part of the compressed video stream [14]. It is therefore very crucial to send it without errors. In [14], a multiple description motion coding (MDMC) algorithm is proposed. This algorithm ensures robust delivery of the motion vector data to the destination. Authors present their simulation results only for the Foreman video sequence. After making a literature search, we realized that there is a shortage of performance results of MDC with all of the video sequences given in [4]. We present the performance investigations of MDC for wireless networks using twelve of the video sequences given in [4]. The goal of this paper is to show the drawbacks of MDC in terms of overhead and quality measurements. In literature, authors usually give exact results for the overhead generated by the MDC process. For example, in [13], it was reported that the overall data rate of multiple streaming is 11 % higher than single–streaming. However, for better understanding and more solid investigation into the overhead generated by MDC, no exact overhead values should be given as it is strongly dependent on the video content. We are investigating the overhead generated by encoding and

c

Aalborg University.

R-04-1005

Page 11

network for IPv6. In the following section our measurement setup and the terminology used is introduced.

1.1 Structure Of This Document The paper is structured as followed: In Section 2 we introduce the methodology for our results achieved. The metric and terms which are used throughout the paper are addressed in Section 3 Our results are presented in Section 4 for the overhead measurements and in Section 6 for the quality measurements.

c

Aalborg University.

R-04-1005

Page 12

2 Methodology and Measurement Setup Our investigations are based on raw video sequences in the 4:2:0 YUV format, which are widely used in the research community and available at [15]. A video frame consists out of multiple pixels. Each pixel can be represented in two different color spaces: RGB or YUV. In the RGB color space, the pixel is composed of three colors: Red (R), Green (G) and Blue (B). In the YUV color space, the pixel is represented by its luminance (Y), chrominance (U), and saturation (V) [16]. Our investigations are based on raw video sequences in the YUV format, which are widely used in the research community and available at [4]. Two video formats are under investigation, namely Quarter Common Intermediate Format (QCIF) (an example is given in Figure 1) and Common Intermediate Format (CIF) (an example is given in Figure 2). QCIF and CIF are representing two different video formats with 176x144 pixels and 352x288 pixels, respectively [17]. Twelve QCIF and six CIF video sequences with different video length and content were used. The video sequences with frame length and a short information are given in Table 1 for the QCIF format and 2 for CIF format.

Figure 1: The carphone sequence in the QCIF format. Figure 2: The bridge–close sequence in the CIF format.

seq. name bridge-close bridge-far carphone claire container foreman grandma highway mthr dotr news salesman silent

no. of frames 2000 2101 382 494 300 400 870 2000 961 300 449 300

Table 1: YUV QCIF info Charles Bridge (Karl` uv most) oldest bridge in Prague. Charles Bridge (Karl` uv most) oldest bridge in Prague. Man talking at the phone. Female talking to the camera. Ship leaving the habor. Man speaking to the camera. Grandma in front of the camera. Driving over highway. Mom and daughter speaking to the camera. News studio and two speakers. Salesman in his office. Woman doing sign language.

As MDC splits the video stream into multiple descriptors, we have chosen the frame based

c

Aalborg University.

R-04-1005

Page 13

seq. name bridge-close bridge-far highway mobile paris tempete

Table 2: YUV CIF info Charles Bridge (Karl` uv most) oldest bridge in Prague. Charles Bridge (Karl` uv most) oldest bridge in Prague. Driving over highway. Train is moving Two people talking to each other. Moving Cam

no. of frames 2000 2101 2000 300 1065 260

approach. By means of a splitter, the raw video sequences are splitted into J sub–sequences, such that the first sub-sequence contains picture 1,J+1,2J+1, and so on. The J sub-sequences are feeded into different video encoder. In this work we only use the H.263 and H.26L. An illustration of the splitting and encoding process is given in Figure 3 for J=3. ..... J+1

1

H.26x Encoder

D1

..... J+2

2

H.26x Encoder

D2

..... J+3

3

H.26x Encoder

D3

raw video sequence J+3 J+2 J+1

J

.....

3

2

1

Figure 3: Splitter and encoding chain for J=3.

The properties of the encoded sub-streams such as amount of data and robustness are measured and evaluated. In the first step we are interested in the overhead that arises from the splitting and encoding process. Obviously the inter frame differences will increase with larger J, which in turn results in smaller compression gains as the video encoder has more data to work on. In our former work we have reported about the bandwidth requirements of the example video sequences for H.263 [18] and H.26L [19] for a single stream. Than we investigate how sensitive the overhead is in dependency of the encoder’s settings. Once the video is encoded with multiple description we are measuring the quality in terms of PSNR in case of missing or corrupted descriptors. For this purpose we need a merger as given in Figure 4. Each descriptor is decoded individually and than merged to a single stream. This stream will be conveyed towards the application. To calculate the PSNR values in dependency of the percentage of successfully received descriptors we select a sub–set of descriptors and feed this to the merger. In case descriptors are lost, frames will be missing in the video sequence. As a a very simple error concealment, we are freezing the last successful received frame until we have a new one. By freezing we mean that we make a copy of the last received frame for the next frame until an update frame is received. This procedure is shown in Figure 5 in case descriptor 2 is missing. In this case frame 1, J + 1, . . . will displayed two times. In case the descriptors are error prone they are decoded and merged afterwards.

c

Aalborg University.

R-04-1005

Page 14

D1

H.26x Decoder

..... J+1

1

D2

H.26x Decoder

..... J+2

2

D3

H.26x Decoder

..... J+3

3

J+3 J+2 J+1

J

.....

3

2

1

J

.....

3

1

1

Figure 4: Merger for J=3.

D1

H.26x Decoder

..... J+1

1

D2

D3

J+3 J+2

H.26x Decoder

..... J+3

J

3

Figure 5: Merger for J=3, where one descriptor is missing.

c

Aalborg University.

R-04-1005

Page 15

3 Metric and Terminology For the performance evaluation we refer to the following terminology: single descriptor stream The single descriptor stream is the encoded version with J = 1. multiple descriptor stream The multiple descriptor stream is the encoded version with J > 1. Thus multiple sub–streams exists, where each of the sub–streams is decodable for itself. sub–streams One encoded entity of the splitted video sequence. quantization parameter (QP) The Discrete Cosine Transform (DCT) is used to convert a block of pixels into a block of transform coefficients. The transform coefficients represent the spatial frequency components of the original block. Then the quantization is achieved as follows [16]: 1. Coefficients with an absolute value smaller than the quantizer threshold are set to zero (i.e., they are considered to be in the so–called dead zone. 2. Coefficients with an absolute value larger than or equal to the quantizer threshold are divided by twice the QP and rounded to the nearest integer. The larger the step size, the larger the compression gain, as well as the loss of information. The trade–off between compression and decodable image quality is controlled by setting the QP and quantizer threshold [16]. For the performance evaluation we refer to the following metric: bandwidth BSD represents the bandwidth of single description coding. Bi is the bandwidth of each descriptor i. overhead The overhead is defined as the amount of data by which the splitted streams of one video sequence are increased in comparison to the single stream. In all results we refer to the overhead produced by the encoding process. Further overhead such as the network overhead are addressed in Section 5. The overhead θ is defined as PJ

θ=

i=1 Bi

BSD

−1

(1)

video quality We use the picture signal to noise ratio (PSNR). The PSNR represents the objective video quality each video frame by a single number. A video frame is composed by N · M pixels (where M is the length and N the height of the frame). Each pixel is presented by one luminance value and a set of pixels by two chrominance values. Because the human eye is more sensible to the change in luminance we focus only on this parameter. The

c

Aalborg University.

R-04-1005

Page 16

mean squared error (MSE) and the PSNR in decibels are computed by the following two equations: P

M SE =

[f (i, j) − F (i, j)]2

∀i,j

N· M  255 P SN R = 20 · log10 √ , M SE

(2) (3)

where f (i, j) represents the original source frame and F (i, j) the reconstructed possibly error–prone frame containing N by M pixels. As mentioned before, wireless links have smaller bandwidth than wired links, therefore we concentrate on the transmission of videos in the CIF and QCIF format.

c

Aalborg University.

R-04-1005

Page 17

4 Coding Overhead Measurements The results are presented for the video sequences given in Table 1 and 2 with values of J=1 . . . 20. For illustration purposes, we highlight only some results in this section and refer to the appendix for complete results. Two different encoders are used, namely H.263 [20] and H.26L [21, 22, 23]. In the first subsection we present the results of the H.263 encoder.

4.1 H.263 For the encoding process we used mainly the default settings of the tmn software []. We changed the value for the skip frames to zero (-S 0) such that each frame of the video sequences was encoded. Furthermore unrestricted motion vector mode (annex D), syntax-based arithmetic coding (annex E), and advanced prediction mode were switched on. This was motivated by former work we have done in this field achieving comparable results. The quantization parameter was not changed and therefore equal to 10.1 The encoded streams contain only one I frame at the beginning and P frames are used afterwards. 4.1.1 QCIF In Figure 6 the overhead of selected H.263 video sequences in the QCIF format for different number of sub–streams is given. We report two main insights: i.) Increasing J results in a larger overhead and ii.) The overhead itself depends dramatically on the video content chosen. While the container sequence has an overhead three times larger than single encoded video in case J ≥ 18, the overhead of the bridge-far sequence is below 100% for all J ≤ 20. For all video sequences the overhead is less than 50% if only two sub-streams are used. H.263 allows for several types of macro-block coding schemes, such as INTRA mode (coding non-predicted DCT coefficients), INTER mode (predictive coding using 1 motion vector) and INTER4V (predictive coding using 4 motion vectors). In Figure 7 the number of INTRA, INTER, and INTER4V macro-blocks per frame for the bridge-far video sequence in the QCIF format is given. 4.1.2 CIF In Figure 8 the overhead of all H.263 video sequences in the CIF format for different number of sub–streams is given. Figure 9 the number of INTRA, INTER, and INTER4V macro-blocks per frame for the bridgefar video sequence in the CIF format is given.

4.2 H.26L For the overhead calculation of H.26L encoded video streams we used a quantization parameter of 31 as default parameter2 . Later we will investigate even the impact of this parameter by varying these values between 1 and 51. We have agreed on a group of picture (GoP) structure

1

Note, that the quantization parameters chosen differ between H.263 and H.26L. This is intended as the H.263 uses quantization parameter between 1 and 31, while the range for H.26L is between 1 and 51. 2 The give quantization value QP refers to the I frame, the P frames are encoded than with QP-1, thus resulting in higher quality. This decision is based on the default settings of the decoder.

c

Aalborg University.

R-04-1005

Page 18

Comparison QCIF H263 3.5 3

overhead

2.5 2 1.5 1 0.5 0 0

2

4

6

foreman salesman

8 10 12 number of sub-streams

14

container bridge-close

16

18

20

bridge-far

Figure 6: Overhead of selected H.263 video sequences in the QCIF format for different number of sub–streams.

Bridge-Far QCIF H263

mean number of macroblock per frame

10

1

0.1

0.01

0.001

0.0001 0

5

inter

10 number of sub-streams inter4v

15

20

intra

Figure 7: Number of INTRA, INTER, and INTER4V macro-blocks per frame for the bridge-far video sequence in the QCIF format.

c

Aalborg University.

R-04-1005

Page 19

Comparison CIF H263 2 1.8 1.6

overhead

1.4 1.2 1 0.8 0.6 0.4 0.2 0 0

2

4

6

bridge-close bridge-far

8 10 12 number of streams

14

highway mobile

16

18

20

tempete paris

Figure 8: Overhead of selected H263 video sequences in the CIF format for different number of sub–streams.

Bridge-Far CIF H263

mean number of macroblock per frame

100

10

1

0.1

0.01

0.001 0

5

inter

10 number of sub-streams inter4v

15

20

intra

Figure 9: Number of INTRA, INTER, and INTER4V macro-blocks per frame for the bridge-far video sequence in the CIF format.

c

Aalborg University.

R-04-1005

Page 20

with one I frame and eleven P frames. As for H.263 we distinguish the same two video formats, namely QCIF and CIF. 4.2.1 QCIF In Figure 10 the overhead of selected H.26L video sequences in the QCIF format for different number of sub–streams is given. As for the H.263 encoder the overhead increase with larger number of descriptors. The highest gain is achieved switching from one descriptor to a second one. In general the overhead caused by the H.26L is less than for the H.263 encoder. Because of different encoder settings this comparison my be unfair, but it is not the main goal to compare the two encoders, but to look into the overhead. Therefore we want to note this observation.

Comparison QCIF H26L 1.6 1.4 1.2

overhead

1 0.8 0.6 0.4 0.2 0 0

2

foreman salesman

4

6

8 10 12 number of sub-streams container bridge-close

14

16

18

20

bridge-far

Figure 10: Overhead of selected H.26L video sequences in the QCIF format for different number of sub–streams. In contrast to the H.263 measurements the curves are not that smooth. This effect can be explained by the used GoP structure. While the encoded video stream with only one descriptor consists out of exactly 1/12 I frames and 11/12 P frames, the encoded stream with multiple descriptors has a much higher ratio of I frames as given in Figure 11. The percentage λ of I frames in one descriptor is calculated as given in Equation 4: h

λ=

N GoP ∗J

i

, (4) N/J where N is the number of frames of the original video sequence, GoP the number of frames per GoP, and J the number of descriptors of the video sequence. The operator [x] returns the smallest integer that is not less than its argument.

c

Aalborg University.

R-04-1005

Page 21

For a video length of 300 frames (as for container, silent, news, etc) the percentage of I frames increases by 44% for 12 descriptors. This explains the gaps in Figure 10. The effect vanish with larger video sequences and lower quantization values (see investigation of the quantization parameters).

Dependency of the video length

percentage of I frames in one descriptor

0.14

0.13

0.12

0.11

0.1

0.09

0.08 1

2

3

300

4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 number of sub-streams 2000

infinity

Figure 11: Percentage of I frames in one descriptor for different video lengths.

4.2.2 CIF In Figure 12 the overhead of selected H26L video sequences in the CIF format for different number of sub–streams is given for all six video sequences introduced in Table 2.

4.3 Impact of the Quantization Parameter In Figure 69 the overhead for the container video sequence in the QCIF format for different quantization values between 1 and 51 is given. Simultaneously in Figure 70 we present also the bandwidth requirements for the container video sequence in the QCIF format for the same quantization values. In Figure 71 and 72 the overhand and the bandwidth requirements are given now for the foreman video sequence.

c

Aalborg University.

R-04-1005

Page 22

Comparison CIF H26L 1.2 1 0.8

overhead

0.6 0.4 0.2 0 -0.2 -0.4 0

2

4

6

8 10 12 number of streams

bridge-close bridge-far

14

16

highway mobile

18

20

tempete paris

Figure 12: Overhead of selected H26L video sequences in the CIF format for different number of sub–streams.

Container QCIF Container QCIF

1e+07

1.8 1.6

bandwidth [bytes]

1.4

overhead

1.2 1 0.8

1e+06

100000

0.6 0.4 0.2 10000 0

0 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

4

6

QP 11

QP 21

8 10 12 number of sub-streams QP 31

14

QP 41

16

18

20

QP 51

QP 51

Figure 13: Overhead for the container video sequence in the QCIF format for different quantization values.

c

Aalborg University.

2

20

Figure 14: Bandwidth requirements for the container video sequence in the QCIF format for different quantization values.

R-04-1005

Page 23

Foreman QCIF Foreman QCIF

1e+07

1.6 1.4

bandwidth [bytes]

1.2

overhead

1 0.8 0.6

1e+06

100000

0.4 0.2 10000 0

0 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

4

6

QP 11

QP 21

8 10 12 number of sub-streams QP 31

14

QP 41

16

18

20

QP 51

QP 51

Figure 15: Overhead for the foreman video sequence in the QCIF format for different quantization values.

c

Aalborg University.

2

20

Figure 16: Bandwidth requirements for the foreman video sequence in the QCIF format for different quantization values.

R-04-1005

Page 24

5 Network Overhead After investigating the overhead caused by the encoding approach, we want to highlight the overhead caused by the underlying communication system. Once the multiple descriptors are available they have to be conveyed to the receiver. In the IP world this is mostly done by using the real time protocol (RTP) in conjunction with the real time control protocol (RTCP). Both protocols transport their packet via the user datagram protocol (UDP) to the underlying Internet protocol (IP) layer. As given in Figure 17 the multiple description coded sources arrive at the RTP layer. Each data segment is packatized into an RTP packet. The different streams are distinguished by the SS

Application

Video MD stream 1

Video MD stream 2

Video MD stream ...

Video MD stream J

Audio MD stream 1

Audio MD stream 2

RTCP rep

RTCP rep

RTCP rep

socket

socket

RTP RTCP rep

RTCP rep

RTCP rep

UDP socket

socket

socket

socket

IP Figure 17: Transport of multiple description coded video and voice over RTP, UDP, and IP network layer.

Each layer will add an additional overhead to each data chunk that is transmitted. Using the RTP/UDP/IP suite will result in an overhead of 40 byte. Header compression schemes such as Van JAcobsen Header Compression (VJHC) ?? or Robust Header Compression (ROHC) ?? will mitigate the overhead. Interested readers are referred to ?? for header compression overview. The use of ROHC, disabling the UDP checksum and doing some more refinements, would decrease the overhead from 40 bytes to 2 bytes per frame. Without these refinements the compressed header is about 6 byte as it is shown in [24]. Is has to be

c

Aalborg University.

R-04-1005

Page 25

mentioned here, that header compression can be used only in case of single hop communication. In case of multi hop some forms of tagging is needed to be capable to route the packet through the network. In Figure 18 and 19 the network overhead for the foreman video sequence and three different quantization values (namely 1, 31, and 51) is given for IP version 4 and IP version 6, respectively. It can be seen that the overhead increases dramatically due to the network overhead and not due to the encoder overhead if large quantization values are used. For small quantization values the impact becomes less. Splitting the video sequence into 20 sub–streams lead to an overhead of 1.3 introduced by the encoder. If the network overhead of IPv6 is also taken under consideration the overhead increases up to 9.2. Therefore, header compression is needed.

Foreman QCIF 9 8

network overhead

7 6 5 4 3 2 1 0 0

2

4

QP 1 QP 31

6

8 10 12 number of sub-streams QP 51 QP 1 + Network

14

16

18

20

QP 31 + Network QP 51 + Network

Figure 18: Network Overhead (RTP/UDP/IPv4) for the foreman video sequence and two different quantization values.

c

Aalborg University.

R-04-1005

Page 26

Foreman QCIF 12

network overhead

10

8

6

4

2

0 0

2

4

QP 1 QP 31

6

8 10 12 number of sub-streams QP 51 QP 1 + Network

14

16

18

20

QP 31 + Network QP 51 + Network

Figure 19: Network Overhead (RTP/UDP/IPv6) for the foreman video sequence and two different quantization values.

c

Aalborg University.

R-04-1005

Page 27

6 Quality Measurements After the overhead measurements we investigate the quality of the video in dependency of successfully received sub–streams. As metric for the video quality we use the picture signal to noise (PSNR). Here the differences between two pictures are taken into account pixel–wise. For the quality measurements we focus only the H.26L encoder. The MDC is done by splitting the video stream on the basis of frames. Video quality is given by the Picture to Signal Noise (PSNR) values. The calculation is done as given in [17]. The PSNR calculations were done using the videometer tool [25]. The videometer tool calculates the PSNR values and is able to freeze video frames in case the following frames are lost. This is important to have some sort of error resilience. In Figure 20 the PSNR measurements using the foreman video sequences in the QCIF format versus percentage of successful received sub–streams is given. The successful received sub– streams are chosen randomly. For each combination (number of sub–streams and percentage), we have repeated the simulations multiple times with a confidence interval of 99%.

Foreman PSNR (Random) 36

PSNR [dB] of Y component

34 32 30 28 26 24 22 20 18 0

2

0.1

0.2

3

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received substreams 5

10

15

0.9

1

20

Figure 20: PSNR measurements using the foreman video sequences in the QCIF format versus percentage of successful received sub–streams (random approach).

By these results we observe that the PSNR values degrades for a larger number of J even if the same percentage of sub–streams is received. As an example we look at the point were 50% of the sub–streams are received for J = 2 and J = 20. With J = 2 the PSNR value is almost 31 dB, while J = 20 leads to a value slightly above 29 dB. The reason is that for J = 2 we alternating loose and receive a frame, while for J = 20, we might end up in not receiving any frame for a long time and than receiving ten frames in a row. The latter case we refer to as worst case, while the former one is referred to as best case. If we now could force the system operate in the best case, results as in Figure 21 can be

c

Aalborg University.

R-04-1005

Page 28

received. Now the PSNR values are mainly the same if the percentage of received sub-streams is also the same.

foreman PSNR (best case) 36 34

video quality [PSNR]

32 30 28 26 24 22 20 18 0

2

0.1

0.2

3

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams 5

10

15

0.9

1

20

Figure 21: PSNR measurements using the foreman video sequences in the QCIF format versus percentage of successful received sub–streams (best case).

On the other side if we have a look at the worst case we see the results in Figure 22. In case of J = 20 and 50% of the sub–streams are received the PSNR value is only 27 dB. As an example we compare the case where one stream out of two is received with the example where five out of ten are received (even though if they have the same percentage of received streams). The first example will always result in better quality as between two missing frames one correct frame will be received. In the second example the gap can be five frames. By means of the results achieved, we come up with an optimization scheme, where the receiver informs the sender or the sender is already aware (e.g. analyzing the RTP sender/receiver reports) which streams are received successfully and rearrange the streams at sender side. As an example if the first ten out of twenty streams are received successfully, while the other half is missing, the sender could rearrange the streams in that way that i.) the streams are mixed randomly or ii.) sending the odd frames over the existing channels and the even frames are skipped. For the foreman example this would lead to an improvement of 2.25 dB if the rearrangement is done randomly and 3.32 dB if the rearrangement is done in smarter way by assuring that the odd frames are received and the even frame are skipped. In Figure 23 the PSNR values versus time for the bad and a random case are given. It can be seen that that the random approach has significant higher PSNR values if the bad case do not receive any information resulting in a dramatic degradation of the PSNR values. In Figure 24, PSNR measurements versus percentage of successfully received sub–streams for a total of J = 15 descriptors are given. The results are obtained using 6 different video

c

Aalborg University.

R-04-1005

Page 29

foreman PSNR (worst case) 36 34

video quality [PSNR]

32 30 28 26 24 22 20 18 0

0.1

2

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams

3

5

10

15

0.9

1

20

Figure 22: PSNR measurements using the foreman video sequences in the QCIF format versus percentage of successful received sub–streams (worst case).

PSNR comparison 36

PSNR [dB] of Y component

34 32 30 28 26 24 22 20 18 10

15

20

25

30 frame

dedicated

35

40

45

50

random

Figure 23: PSNR values versus time for the bad case and a random case.

c

Aalborg University.

R-04-1005

Page 30

sequences. The foreman video sequence includes high motion. Whereas clarie and container video sequences include relatively lower motion than foreman. As we can observe in the figure, if the motion in a given video sequence is high, the slope of the PSNR degradation curve is also high. For example, for a loss of 60 % out of J = 15 descriptors, the PSNR degradation is 6 dB for foreman, 2 dB for claire and 0.5 dB for container video sequence. Thus, we conclude that the more the motion in a given video sequence, the more important is to receive as many descriptors as possible for a given J.

Compare PSNR (Random) for 15 descriptors 40 38 PSNR [dB] of Y component

36 34 32 30 28 26 24 22 20 0

0.1

0.2

foreman highway

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received substreams silent container

0.9

1

claire carphone

Figure 24: Comparison of PSNR degradation for different video sequences for J = 15.

7 Conclusion 8 Acknowledgment This work was done within the Jade project supported by SAMSUNG, Korea. The script for splitting and Merging was done by Patrick Seeling. Parts of the introduction were taken from one of our WWRF contribution in 2004. We further would like to thank Axel Meyer for his advice in video post-processing. Furthermore we would to thank Sterica Rein for the technical support generating the video sequences highway, bridge-far, and bridge-close in Eastern Europe. Also we would like to thank all anonymous people which helped us to realize this work.

c

Aalborg University.

R-04-1005

Page 31

References [1] J. G. Apostolopoulos, “Error Resilient Video Compression Through The Use of Multiple States,” in IEEE Int’l Conf. on Image Processing, Vancouver, BC, Canada, September 2000, vol. 3, pp. 352–355. [2] J. G. Apostolopoulos and S. J. Wee, ,” . [3] N. Gogate, D. M. Chung, S. S. Panwar, and Y. Wang, “Supporting Image and Video Applications in a Multihop Radio Environment Using Path Diversity and Multiple Description Coding,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 12, no. 9, pp. 777–834, September 2002. [4] “Video traces for network performance evaluation: http://trace.eas.asu.edu/yuv/yuv.html.

YUV 4:2:0 Video Sequences,”

¨ Erg¨ [5] A.C. Beˇ g en, Y. Altunba¸sak, and O. un, “Multi-path Selection for Multiple Description Encoded Video Streaming,” in IEEE Int’l Conf. on Communications, Anchorage,Alaska,USA, May 2003, vol. 3, pp. 1583–1589. ¨ Erg¨ [6] A. C. Beˇ g en, Y. Altunba¸sak, and O. un, “Fast Heuristics for Multi-Path Selection for Multiple Description Encoded Video Streaming,” in IEEE Int’l Conf. on Multimedia and Expo, Baltimore, Maryland, July 2003, pp. 517–520. [7] P. A. Chou, H. J. Wang, and V. N. Padmanabhan, “Layered Multiple Description Coding,” in In Proc. Packet Video Workshop, April 2003. [8] Vivek K Goyal, “Multiple Description Coding: Compression Meets the Network,” IEEE SIGNAL PROCESSING MAGAZINE, vol. 01, pp. 74–93, September 2001. [9] Yao Wang, Shivendra Panwar, Shunan Lin, and Shiwen Mao, “Wireless video transport using path diversity: multiple description vs layered coding,” Image Processing Proceedings, vol. 1, pp. I–21 – I–24, September 2002. [10] R. Prasad and S. Hara, Multicarrier Techniques for 4G Mobile Communications, Universal Personal Communications. Artech House, 2003. [11] Frank H.P. Fitzek, Ba¸sak Can, Ramjee Prasad, DS Park, and Youngkwon Cho, “Application of Multiple Description Coding in 4G Wireless Communication Systems,” WWRF 8 bis, February 2004. [12] M. Schaar and D. S. Turaga, “Multiple Description Scalable Coding Using Wavelet-Based Motion Compensated Temporal Filtering,” in IEEE Int’l Conf. on Image Processing, Barcelona, Spain, September 2003, vol. 3, pp. 489–492. [13] Y. J. Liang, E. G. Steinbach, and B. Girod., “Multi-stream Voice over IP Using Packet Path Diversity,” in IEEE 4th Workshop on Multimedia Signal Processing, 2001. [14] C. S. Kim and S. U. Lee, “Multiple Description Coding of Motion Fields for Robust Video Transmission,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 11, no. 9, pp. 999–1010, September 2001.

c

Aalborg University.

R-04-1005

Page 32

[15] Arizona State University, http://trace.eas.asu.edu.

“Video traces for network performance evaluation,”

[16] Frank Fitzek, Patrick Seeling, and Martin Reisslein, Wireless Internet, chapter Video Streaming in Wireless Internet, Number ISBN: 0849316316, Electrical Engineering & Applied Signal Processing Series. March 2004. [17] F.H.P. Fitzek, P. Seeling, and M. Reisslein, Wireless Internet, chapter Video Streaming in Wireless Internet, Electrical Engineering & Applied Signal Processing Series. CRC Press, 2004, To be published. [18] F.H.P. Fitzek and M. Reisslein, “MPEG–4 and H.263 Video Traces for Network Performance Evaluation,” Tech. Rep., Technical University of Berlin, 2000, TKN–00–06. [19] P. Seeling, F. Fitzek, M. Reisslein, and A. Wolisz, “H26L Video Traffic Measurements,” Tech. Rep. TKN-02-000, TKN, August 2002, www.acticom.info. [20] ITU-T/SG15, “Recommendation H.263, video coding for low bitrate communication,” 1996. [21] JVT, “JM / TML Software Encoder/decoder,” http://bs.hhi.de/ suehring/tml/, 2002, H.26L Software Coordination by Carsten Suehring. [22] Th. Wiegand, “H.26L Test Model Long–Term Number 9 (TML-9) draft0,” ITU-T Study Group 16, Dec. 2001. [23] Thomas Wiegand, “H26lstandard,” Tech. Rep., JVT, 2002. [24] P. Seeling, F.H.P. Fitzek, S. Hendrata, and M. Reisslein, “Video Quality Evaluation for Wireless Transmission with Robust Header Compression,” in Fourth International Conference on Information, Communications & Signal Processing and Fourth IEEE Pacific-Rim Conference on Multimedia (ICICS-PCM 2003). IEEE, december 2003. [25] F.H.P. Fitzek, P. Seeling, and M. Reisslein, “VideoMeter tool for YUV bitstreams,” Tech. Rep. acticom-02-001, acticom – mobile networks, Germany, October 2002.

c

Aalborg University.

R-04-1005

Page 33

A Additional Overhead Measurements A.1 H263 A.1.1 QCIF In this section the overhead of multiple description coded video sequence for 12 famous sequences in the QCIF format are given.

Bridge-Close QCIF H263

Bridge-Far QCIF H263

1.5

1 0.8

1

0.6

overhead

overhead

0.4 0.5

0

0.2 0 -0.2 -0.4

-0.5

-0.6 -0.8

-1

-1 0

5

10 number of sub-streams

total data

15

20

0

data per stream

5

10 number of sub-streams

total data

15

20

data per stream

Figure 25: Overhead of multiple description Figure 26: Overhead of multiple description coded video sequence bridge-close in coded video sequence bridge-far in the QCIF format using the H.263 enthe QCIF format using the H.263 encoder. coder.

Claire QCIF H263

Carphone QCIF H263 3.5

2

3 1.5 2.5 2 overhead

overhead

1

0.5

0

1.5 1 0.5 0

-0.5 -0.5 -1

-1 0

5

total data

10 number of sub-streams

15

0

20

5

total data

data per stream

10 number of sub-streams

15

20

data per stream

Figure 27: Overhead of multiple description Figure 28: Overhead of multiple description coded video sequence claire in the coded video sequence carphone in QCIF format using the H.263 enthe QCIF format using the H.263 encoder. coder.

c

Aalborg University.

R-04-1005

Page 34

Container QCIF H263

Foreman QCIF H263

3.5

2

3 1.5 2.5 1 overhead

overhead

2 1.5 1 0.5

0.5

0

0 -0.5 -0.5 -1

-1 0

5

10 number of sub-streams

total data

15

20

0

data per stream

5

10 number of sub-streams

total data

15

20

data per stream

Figure 29: Overhead of multiple description Figure 30: Overhead of multiple description coded video sequence container in coded video sequence foreman in the the QCIF format using the H.263 enQCIF format using the H.263 encoder. coder.

Highway QCIF H263

Grandma QCIF H263 1.5

3 2.5

1 2

overhead

overhead

1.5 1 0.5

0.5

0

0 -0.5 -0.5 -1

-1 0

5

total data

10 number of sub-streams

15

0

20

5

total data

data per stream

10 number of sub-streams

15

20

data per stream

Figure 31: Overhead of multiple description Figure 32: Overhead of multiple description coded video sequence highway in the coded video sequence grandma in QCIF format using the H.263 enthe QCIF format using the H.263 encoder. coder.

c

Aalborg University.

R-04-1005

Page 35

News QCIF H263 3

2.5

2.5

2

2

1.5

1.5 overhead

overhead

Mthr_Dotr QCIF H263 3

1 0.5

1 0.5

0

0

-0.5

-0.5

-1

-1 0

5

10 number of sub-streams

total data

15

20

0

data per stream

5

10 number of sub-streams

total data

15

20

data per stream

Figure 33: Overhead of multiple description Figure 34: Overhead of multiple description coded video sequence mthr dotr in coded video sequence news in the QCIF format using the H.263 enthe QCIF format using the H.263 encoder. coder.

Silent QCIF H263

Salesman QCIF H263 2.5

3.5 3

2

2.5 1.5 overhead

overhead

2 1.5 1

1 0.5

0.5 0 0 -0.5

-0.5

-1

-1 0

5

total data

10 number of sub-streams

15

0

20

5

total data

data per stream

10 number of sub-streams

15

20

data per stream

Figure 35: Overhead of multiple description Figure 36: Overhead of multiple description coded video sequence silent in the coded video sequence salesman in QCIF format using the H.263 enthe QCIF format using the H.263 encoder. coder.

c

Aalborg University.

R-04-1005

Page 36

A.1.2 CIF In this section the overhead of multiple description coded video sequence for 6 famous sequences in the CIF format are given.

Bridge-Far 1.2

1

1

0.8

0.8

0.6

0.6

0.4

0.4 overhead

overhead

Bridge-Close 1.2

0.2 0

0.2 0

-0.2

-0.2

-0.4

-0.4

-0.6

-0.6

-0.8

-0.8

-1

-1 0

5

10 number of sub-streams

total data

15

20

0

data per stream

5

10 number of sub-streams

total data

15

20

data per stream

Figure 37: Overhead of multiple description Figure 38: Overhead of multiple description coded video sequence bridge-close in coded video sequence bridge-far in the CIF format using the H.263 enthe CIF format using the H.263 encoder. coder.

Mobile

Highway 0.8

2

0.6 1.5 0.4 0.2 overhead

overhead

1

0.5

0 -0.2 -0.4

0

-0.6 -0.5 -0.8 -1

-1 0

5

total data

10 number of sub-streams

15

0

20

5

total data

data per stream

10 number of sub-streams

15

20

data per stream

Figure 39: Overhead of multiple description Figure 40: Overhead of multiple description coded video sequence mobile in the coded video sequence highway in the CIF format using the H.263 encoder. CIF format using the H.263 encoder.

A.2 H26L

c

Aalborg University.

R-04-1005

Page 37

Paris

Tempete

2

1 0.8

1.5

0.6 0.4 overhead

overhead

1

0.5

0

0.2 0 -0.2 -0.4 -0.6

-0.5

-0.8 -1

-1 0

5

total data

10 number of sub-streams

15

20

0

data per stream

5

total data

10 number of sub-streams

15

20

data per stream

Figure 41: Overhead of multiple description Figure 42: Overhead of multiple description coded video sequence paris in the coded video sequence tempete in the CIF format using the H.263 encoder. CIF format using the H.263 encoder.

c

Aalborg University.

R-04-1005

Page 38

A.2.1 QCIF In this section the total amount of data and the amount per sub-stream for 12 famous video sequences in the QCIF format are given.

Bridge-Close QCIF H26L

Bridge-Far QCIF H26L

0.8

0.2

0.6 0 0.4 -0.2 overhead

overhead

0.2 0 -0.2 -0.4

-0.4

-0.6

-0.6 -0.8 -0.8 -1

-1 0

5

10 number of sub-streams

total data

15

20

0

data per stream

5

10 number of sub-streams

total data

15

20

data per stream

Figure 43: Overhead of multiple description Figure 44: Overhead of multiple description coded video sequence bridge-close in coded video sequence bridge-far in the QCIF format using the H.26L the QCIF format using the H.26L encoder. encoder.

Carphone QCIF H26L

Claire QCIF H26L

1.2

1.5

1 0.8

1

0.6

overhead

overhead

0.4 0.2 0 -0.2

0.5

0

-0.4 -0.5

-0.6 -0.8 -1

-1 0

5

total data

10 number of sub-streams

15

20

0

data per stream

5

total data

10 number of sub-streams

15

20

data per stream

Figure 45: Overhead of multiple description Figure 46: Overhead of multiple description coded video sequence carphone in coded video sequence claire in the the QCIF format using the H.26L QCIF format using the H.26L enencoder. coder.

c

Aalborg University.

R-04-1005

Page 39

Container QCIF H26L

Foreman QCIF H26L

2

1.5

1.5

1

overhead

overhead

1

0.5

0.5

0

0 -0.5

-0.5

-1

-1 0

5

10 number of sub-streams

total data

15

20

0

data per stream

5

10 number of sub-streams

total data

15

20

data per stream

Figure 47: Overhead of multiple description Figure 48: Overhead of multiple description coded video sequence container in coded video sequence foreman in the the QCIF format using the H.26L QCIF format using the H.26L enencoder. coder.

Highway QCIF H26L 1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

overhead

overhead

Grandma QCIF H26L 1

0 -0.2

0 -0.2

-0.4

-0.4

-0.6

-0.6

-0.8

-0.8 -1

-1 0

5

total data

10 number of sub-streams

15

0

20

5

total data

data per stream

10 number of sub-streams

15

20

data per stream

Figure 49: Overhead of multiple description Figure 50: Overhead of multiple description coded video sequence highway in the coded video sequence grandma in QCIF format using the H.26L enthe QCIF format using the H.26L coder. encoder.

c

Aalborg University.

R-04-1005

Page 40

Mthr_Dotr QCIF H26L

News QCIF H26L

2

1.5

1.5

1

overhead

overhead

1

0.5

0.5

0

0 -0.5

-0.5

-1

-1 0

5

10 number of sub-streams

total data

15

20

0

data per stream

5

10 number of sub-streams

total data

15

20

data per stream

Figure 51: Overhead of multiple description Figure 52: Overhead of multiple description coded video sequence mthr dotr in coded video sequence news in the QCIF format using the H.26L enthe QCIF format using the H.26L coder. encoder.

Silent QCIF H26L

Salesman QCIF H26L 1.2

1

1

0.8

0.8

0.6

0.6

0.4

0.4 overhead

overhead

1.2

0.2 0

0.2 0

-0.2

-0.2

-0.4

-0.4

-0.6

-0.6

-0.8

-0.8 -1

-1 0

5

total data

10 number of sub-streams

15

0

20

5

total data

data per stream

10 number of sub-streams

15

20

data per stream

Figure 53: Overhead of multiple description Figure 54: Overhead of multiple description coded video sequence silent in the coded video sequence salesman in QCIF format using the H.26L enthe QCIF format using the H.26L coder. encoder.

c

Aalborg University.

R-04-1005

Page 41

A.2.2 CIF In this section the total amount of data and the amount per sub-stream for 6 famous video sequences in the CIF format are given.

Bridge-Close

Bridge-Far 2.5

2

2

1.5

1.5 overhead

overhead

2.5

1 0.5

1 0.5

0

0

-0.5

-0.5

-1

-1 0

5

10 number of sub-streams

total data

15

20

0

data per stream

5

10 number of sub-streams

total data

15

20

data per stream

Figure 55: Overhead of multiple description Figure 56: Overhead of multiple description coded video sequence bridge-close in coded video sequence bridge-far in the CIF format using the H.26L enthe CIF format using the H.26L encoder. coder.

Highway

Mobile

1

1

0.8

0.8 0.6

0.6

0.4 overhead

overhead

0.4 0.2 0 -0.2

0.2 0 -0.2 -0.4

-0.4

-0.6

-0.6

-0.8

-0.8

-1 0

5

total data

10 number of sub-streams

15

20

0

data per stream

5

total data

10 number of sub-streams

15

20

data per stream

Figure 57: Overhead of multiple description Figure 58: Overhead of multiple description coded video sequence highway in coded video sequence mobile in the the CIF format using the H.26L enCIF format using the H.26L encoder. coder.

c

Aalborg University.

R-04-1005

Page 42

Paris

Tempete

2.5

1.5

2 1

overhead

overhead

1.5 1 0.5

0.5

0

0 -0.5 -0.5 -1

-1 0

5

total data

10 number of sub-streams

15

20

0

data per stream

5

total data

10 number of sub-streams

15

20

data per stream

Figure 59: Overhead of multiple description Figure 60: Overhead of multiple description coded video sequence parisparis in coded video sequence tempetetemthe CIF format using the H.26L enpete in the CIF format using the H.26L encoder. coder.

c

Aalborg University.

R-04-1005

Page 43

B Additional Quantization Measurements

Bridge-Close QCIF Bridge-Close QCIF

1e+08

0.7 0.6 1e+07 bandwidth [bytes]

0.5

overhead

0.4 0.3

1e+06

0.2 100000 0.1 0 10000 0

-0.1 0

2

QP 1

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

2

4

6

20 QP 11

QP 21

8 10 12 number of sub-streams QP 31

14

16

QP 41

18

20

QP 51

QP 51

Figure 61: Overhead for the bridge-close video sequence in the QCIF format for different quantization values.

Figure 62: Bandwidth requirements for the bridge-close video sequence in the QCIF format for different quantization values.

Bridge-Far QCIF Bridge-Far QCIF

1e+08

0.2 0.18 0.16

1e+07 bandwidth [bytes]

0.14

overhead

0.12 0.1 0.08

1e+06

0.06 100000

0.04 0.02 0

10000 0

-0.02 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

QP 11

4

6

QP 21

8 10 12 number of sub-streams QP 31

14

QP 41

16

18

20

QP 51

QP 51

Figure 63: Overhead for the bridge-far video sequence in the QCIF format for different quantization values.

c

Aalborg University.

2

20

Figure 64: Bandwidth requirements for the bridge-far video sequence in the QCIF format for different quantization values.

R-04-1005

Page 44

Carphone QCIF Carphone QCIF

1e+07

1.2

bandwidth [bytes]

1

overhead

0.8

0.6

1e+06

100000

0.4

0.2 10000 0

0 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

2

4

6

20 QP 11

QP 21

8 10 12 number of sub-streams QP 31

14

16

QP 41

18

20

QP 51

QP 51

Figure 65: Overhead for the carphone video sequence in the QCIF format for different quantization values.

Figure 66: Bandwidth requirements for the carphone video sequence in the QCIF format for different quantization values.

Claire QCIF Claire QCIF

1e+07

1.4 1.2

bandwidth [bytes]

overhead

1 0.8 0.6

1e+06

100000

0.4 0.2 10000 0

0 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

QP 11

4

6

QP 21

8 10 12 number of sub-streams QP 31

14

QP 41

16

18

20

QP 51

QP 51

Figure 67: Overhead for the claire video sequence in the QCIF format for different quantization values.

c

Aalborg University.

2

20

Figure 68: Bandwidth requirements for the claire video sequence in the QCIF format for different quantization values.

R-04-1005

Page 45

Container QCIF Container QCIF

1e+07

1.8 1.6

bandwidth [bytes]

1.4

overhead

1.2 1 0.8

1e+06

100000

0.6 0.4 0.2 10000 0

0 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

2

4

6

20 QP 11

QP 21

8 10 12 number of sub-streams QP 31

14

16

QP 41

18

20

QP 51

QP 51

Figure 69: Overhead for the container video sequence in the QCIF format for different quantization values.

Figure 70: Bandwidth requirements for the container video sequence in the QCIF format for different quantization values.

Foreman QCIF Foreman QCIF

1e+07

1.6 1.4

bandwidth [bytes]

1.2

overhead

1 0.8 0.6

1e+06

100000

0.4 0.2 10000 0

0 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

QP 11

4

6

QP 21

8 10 12 number of sub-streams QP 31

14

QP 41

16

18

20

QP 51

QP 51

Figure 71: Overhead for the foreman video sequence in the QCIF format for different quantization values.

c

Aalborg University.

2

20

Figure 72: Bandwidth requirements for the foreman video sequence in the QCIF format for different quantization values.

R-04-1005

Page 46

Grandma QCIF Grandma QCIF

1e+08

1 0.9 1e+07 bandwidth [bytes]

0.8

overhead

0.7 0.6 0.5 0.4

1e+06

100000

0.3 0.2 0.1

10000 0

0 0

2

QP 1

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

2

4

6

20 QP 11

QP 21

8 10 12 number of sub-streams QP 31

14

16

QP 41

18

20

QP 51

QP 51

Figure 73: Overhead for the grandma video sequence in the QCIF format for different quantization values.

Figure 74: Bandwidth requirements for the grandma video sequence in the QCIF format for different quantization values.

Highway QCIF Highway QCIF

1e+08

1 0.9 0.8

1e+07 bandwidth [bytes]

0.7

overhead

0.6 0.5 0.4

1e+06

0.3 100000

0.2 0.1 0

10000 0

-0.1 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

QP 11

4

6

QP 21

8 10 12 number of sub-streams QP 31

14

QP 41

16

18

20

QP 51

QP 51

Figure 75: Overhead for the highway video sequence in the QCIF format for different quantization values.

c

Aalborg University.

2

20

Figure 76: Bandwidth requirements for the highway video sequence in the QCIF format for different quantization values.

R-04-1005

Page 47

Mthr_Dotr QCIF Mthr_Dotr QCIF

1e+08

1.6 1.4 1e+07 bandwidth [bytes]

1.2

overhead

1 0.8

1e+06

0.6 100000 0.4 0.2 10000 0

0 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

2

4

6

20 QP 11

QP 21

8 10 12 number of sub-streams QP 31

14

16

QP 41

18

20

QP 51

QP 51

Figure 77: Overhead for the mthr dotr video sequence in the QCIF format for different quantization values.

Figure 78: Bandwidth requirements for the mthr dotr video sequence in the QCIF format for different quantization values.

News QCIF News QCIF

1e+07

1.4 1.2

bandwidth [bytes]

overhead

1 0.8 0.6

1e+06

100000

0.4 0.2 10000 0

0 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

QP 11

4

6

QP 21

8 10 12 number of sub-streams QP 31

14

QP 41

16

18

20

QP 51

QP 51

Figure 79: Overhead for the news video sequence in the QCIF format for different quantization values.

c

Aalborg University.

2

20

Figure 80: Bandwidth requirements for the news video sequence in the QCIF format for different quantization values.

R-04-1005

Page 48

Salesman QCIF Salesman QCIF

1e+07

1.2 1

bandwidth [bytes]

overhead

0.8 0.6 0.4

1e+06

100000

0.2 0 10000 0

-0.2 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

2

4

6

20 QP 11

QP 21

8 10 12 number of sub-streams QP 31

14

16

QP 41

18

20

QP 51

QP 51

Figure 81: Overhead for the salesman video sequence in the QCIF format for different quantization values.

Figure 82: Bandwidth requirements for the salesman video sequence in the QCIF format for different quantization values.

Silent QCIF Silent QCIF

1e+07

1.2

bandwidth [bytes]

1

overhead

0.8

0.6

1e+06

100000

0.4

0.2 10000 0

0 0

QP 1

2

4

QP 11

6

QP 21

8 10 12 number of sub-streams QP 31

14

16

18

QP 1 QP 41

QP 11

4

6

QP 21

8 10 12 number of sub-streams QP 31

14

QP 41

16

18

20

QP 51

QP 51

Figure 83: Overhead for the silent video sequence in the QCIF format for different quantization values.

c

Aalborg University.

2

20

Figure 84: Bandwidth requirements for the silent video sequence in the QCIF format for different quantization values.

R-04-1005

Page 49

C Additional Quality Measurements

carphone PSNR comparison 36 35.5

video quality [PSNR]

35 34.5 34 33.5 33 32.5 32 31.5 0.5

Mode F/T

0.6 0.7 0.8 0.9 percentage of successfully received sub-streams Mode R

1

worst case

Figure 85: PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 2

c

Aalborg University.

R-04-1005

Page 50

carphone PSNR comparison 36

video quality [PSNR]

35

34

33

32

31

30 0.3

0.4

Mode F/T

0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 86: PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 3

carphone PSNR comparison 36 35

video quality [PSNR]

34 33 32 31 30 29 28 27 0.2

0.3

Mode F/T

0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 87: PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 5

c

Aalborg University.

R-04-1005

Page 51

carphone PSNR comparison 36

video quality [PSNR]

34

32

30

28

26

24 0.1

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams

Mode F/T

Mode R

0.9

1

worst case

Figure 88: PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 10

carphone PSNR comparison 36

video quality [PSNR]

34 32 30 28 26 24 22 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 89: PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 15

c

Aalborg University.

R-04-1005

Page 52

carphone PSNR comparison 36

video quality [PSNR]

34 32 30 28 26 24 22 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 90: PSNR comparison for different modes and the worst case investigating carphone video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 20

claire PSNR comparison 38.6 38.5

video quality [PSNR]

38.4 38.3 38.2 38.1 38 37.9 37.8 37.7 37.6 0.5

Mode F/T

0.6 0.7 0.8 0.9 percentage of successfully received sub-streams Mode R

1

worst case

Figure 91: PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 2

c

Aalborg University.

R-04-1005

Page 53

claire PSNR comparison 38.6 38.4

video quality [PSNR]

38.2 38 37.8 37.6 37.4 37.2 37 36.8 36.6 0.3

0.4

Mode F/T

0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 92: PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 3

claire PSNR comparison 38.5 38

video quality [PSNR]

37.5 37 36.5 36 35.5 35 34.5 0.2

0.3

Mode F/T

0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 93: PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 5

c

Aalborg University.

R-04-1005

Page 54

claire PSNR comparison 39

video quality [PSNR]

38 37 36 35 34 33 32 0.1

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams

Mode F/T

Mode R

0.9

1

worst case

Figure 94: PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 10

claire PSNR comparison 39 38

video quality [PSNR]

37 36 35 34 33 32 31 30 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 95: PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 15

c

Aalborg University.

R-04-1005

Page 55

claire PSNR comparison 39 38 37 video quality [PSNR]

36 35 34 33 32 31 30 29 28 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 96: PSNR comparison for different modes and the worst case investigating claire video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 20

container PSNR comparison 34.9 34.85

video quality [PSNR]

34.8 34.75 34.7 34.65 34.6 34.55 34.5 34.45 34.4 0.5

Mode F/T

0.6 0.7 0.8 0.9 percentage of successfully received sub-streams Mode R

1

worst case

Figure 97: PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 2

c

Aalborg University.

R-04-1005

Page 56

container PSNR comparison 34.9 34.8 34.7 video quality [PSNR]

34.6 34.5 34.4 34.3 34.2 34.1 34 33.9 33.8 0.3

0.4

Mode F/T

0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 98: PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 3

container PSNR comparison 34.8 34.6 34.4 video quality [PSNR]

34.2 34 33.8 33.6 33.4 33.2 33 32.8 32.6 0.2

0.3

Mode F/T

0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 99: PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 5

c

Aalborg University.

R-04-1005

Page 57

container PSNR comparison 35 34.5

video quality [PSNR]

34 33.5 33 32.5 32 31.5 31 30.5 30 0.1

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams

Mode F/T

Mode R

0.9

1

worst case

Figure 100: PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 10

container PSNR comparison 35

video quality [PSNR]

34 33 32 31 30 29 28 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 101: PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 15

c

Aalborg University.

R-04-1005

Page 58

container PSNR comparison 35 34

video quality [PSNR]

33 32 31 30 29 28 27 26 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 102: PSNR comparison for different modes and the worst case investigating container video sequences in the QCIF format versus percentage of successful received sub– streams. Number of descriptors is 20

foreman PSNR comparison 34.5 34

video quality [PSNR]

33.5 33 32.5 32 31.5 31 30.5 0.5

Mode F/T

0.6 0.7 0.8 0.9 percentage of successfully received sub-streams Mode R

1

worst case

Figure 103: PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 2

c

Aalborg University.

R-04-1005

Page 59

foreman PSNR comparison 35

video quality [PSNR]

34 33 32 31 30 29 28 0.3

0.4

Mode F/T

0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 104: PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 3

foreman PSNR comparison 35 34

video quality [PSNR]

33 32 31 30 29 28 27 26 25 0.2

0.3

Mode F/T

0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 105: PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 5

c

Aalborg University.

R-04-1005

Page 60

foreman PSNR comparison 36 34

video quality [PSNR]

32 30 28 26 24 22 20 0.1

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams

Mode F/T

Mode R

0.9

1

worst case

Figure 106: PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 10

foreman PSNR comparison 36 34

video quality [PSNR]

32 30 28 26 24 22 20 18 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 107: PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 15

c

Aalborg University.

R-04-1005

Page 61

foreman PSNR comparison 36 34

video quality [PSNR]

32 30 28 26 24 22 20 18 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 108: PSNR comparison for different modes and the worst case investigating foreman video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 20

highway PSNR comparison 37

video quality [PSNR]

36.5

36

35.5

35

34.5

34 0.5

Mode F/T

0.6 0.7 0.8 0.9 percentage of successfully received sub-streams Mode R

1

worst case

Figure 109: PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 2

c

Aalborg University.

R-04-1005

Page 62

highway PSNR comparison 36.5 36

video quality [PSNR]

35.5 35 34.5 34 33.5 33 32.5 0.3

0.4

Mode F/T

0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 110: PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 3

highway PSNR comparison 37

video quality [PSNR]

36

35

34

33

32

31 0.2

0.3

Mode F/T

0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 111: PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 5

c

Aalborg University.

R-04-1005

Page 63

highway PSNR comparison 37 36

video quality [PSNR]

35 34 33 32 31 30 29 0.1

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams

Mode F/T

Mode R

0.9

1

worst case

Figure 112: PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 10

highway PSNR comparison 37 36

video quality [PSNR]

35 34 33 32 31 30 29 28 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 113: PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 15

c

Aalborg University.

R-04-1005

Page 64

highway PSNR comparison 37 36

video quality [PSNR]

35 34 33 32 31 30 29 28 27 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 114: PSNR comparison for different modes and the worst case investigating highway video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 20

silent PSNR comparison 34.5

video quality [PSNR]

34

33.5

33

32.5

32 0.5

Mode F/T

0.6 0.7 0.8 0.9 percentage of successfully received sub-streams Mode R

1

worst case

Figure 115: PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 2

c

Aalborg University.

R-04-1005

Page 65

silent PSNR comparison 34.5 34

video quality [PSNR]

33.5 33 32.5 32 31.5 31 30.5 0.3

0.4

Mode F/T

0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 116: PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 3

silent PSNR comparison 35

video quality [PSNR]

34 33 32 31 30 29 28 0.2

0.3

Mode F/T

0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 117: PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 5

c

Aalborg University.

R-04-1005

Page 66

silent PSNR comparison 35 34

video quality [PSNR]

33 32 31 30 29 28 27 26 0.1

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams

Mode F/T

Mode R

0.9

1

worst case

Figure 118: PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 10

silent PSNR comparison 35 34

video quality [PSNR]

33 32 31 30 29 28 27 26 25 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 119: PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 15

c

Aalborg University.

R-04-1005

Page 67

silent PSNR comparison 36

video quality [PSNR]

34 32 30 28 26 24 22 0

0.1

Mode F/T

0.2

0.3 0.4 0.5 0.6 0.7 0.8 percentage of successfully received sub-streams Mode R

0.9

1

worst case

Figure 120: PSNR comparison for different modes and the worst case investigating silent video sequences in the QCIF format versus percentage of successful received sub–streams. Number of descriptors is 20

c

Aalborg University.

R-04-1005

Page 68