C ODING EFFICIENCY COMPARISON FOR DIFFERENT REFERENCE PICTURE NUMBER

number

In this section we show the efficiency of multiple references in the symmetric tree prediction structure. Five cases are compared:

¾ N64_6B1_ref1 (1 forward and 1 backward reference)

¾ N64_6B1_full (all references)

¾ N64_6P1_ref1 (1 forward reference)

¾ N64_6P1_ref2 (2 forward references)

¾ N64_6P1_full (all references).

All the cases have six levels with one branch at each level. The first two cases use all B-pictures. The “6B1_ref1” case use one forward and one backward reference frame.

The “6B1_full” case uses the entire reference frame in the buffer. The last three cases use all P-pictures. The “6P1_ref1” and “6P1_ref2” cases use one and two forward reference frame, respectively. The “6P1_full” case uses the entire reference frame in the buffer.

From the simulation results, we found that for all B-pictures, in the more static sequences such as Foreman and Mobile, “6B1_ref1” has around 0.2 dB loss compare with “6B1_full”. And there is no difference for the non-static sequences such as Bus and Football. For all P-pictures, “6P1_ref2” is almost identical with “6P1_full” for all sequences. In the more static sequences such as Foreman and Mobile, “6P1_ref1” has around 0.2 dB loss compare with “6P1_full”. And there is no difference for the non-static sequences such as Bus and Football.

The results show the improvement from more reference is rare in symmetric tree prediction structure. This might because in the proposed structure, the reference frame

at higher index has much longer prediction distance compare with the reference frame at lower index, therefore limited their prediction efficiency. From these analyses, in the proposed structure, we suggest to use 1 forward and 1 backward reference for B-pictures, and 1 or 2 forward reference for P-pictures.

Bus

Figure 4-9 the RD curve of Bus with different numbers of references

Foreman

0 384 768 1152 1536 1920

bitrate

Figure 4-10 the RD curve of Foreman with different numbers of references

Mobile

0 384 768 1152 1536 1920 2304 2688 3072 3456 3840 4224 4608 4992 5376

bitrate

Figure 4-11 the RD curve of Mobile with different numbers of references

FOOTBALL

27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

0 384 768 1152 1536 1920 2304 2688 3072 3456 3840

bitrate

SnrY

N64_6B1_full N64_6B1_ref1 N64_6P1_ref1 N64_6P1_ref2 N64_6P1_full

Figure 4-12 the RD curve of Football with different numbers of references

4.4 Coding efficiency comparison with normal GOP structure

In this part, we compare the coding efficiency between normal structure and our symmetric tree prediction structure. From 4.2, we know that using B picture only at the last 3 levels has good coding efficiency and small encoder buffer size in symmetric tree structure. So we use this structure to compare with normal GOP structure with M=2.

Normal GOP structure:

¾ N64_M2

¾ N32_M2

¾ N16_M2

Symmetric tree prediction structure:

¾ N64_3P1_3B1

For normal GOP structure, the maximum decoding delay is a linear function of the GOP size, therefore usually normal GOP size will not choose large GOP size to prevent problem in achieving VCR functionality at decoder. Table 4-1 shows the maximum decoding delay for normal GOP structure and symmetric tree prediction structure. We can found that even with largest GOP size, symmetric tree prediction structure still provide shortest maximum decoding delay.

Table 4-1 max and average delay comparison with normal structure and symmetric tree structure

Delay

N64_M2 N32_M2 N16_M2 N64_3P1_3B1

Max 32 16 8 7

average 16.23077 8.212121 4.176471 3.892308

Structures

From the simulation results, we first compare the coding efficiency between the normal GOP with N=16 with the symmetric tree prediction structure, which has similar maximum decoding delay as shown in Table 4-1. For the static sequence, symmetric tree prediction structure provides at most 1.2dB gain in Mobile sequence, and at most 0.3dB gain in Foreman sequences. In the fast motion sequences, it has around 0.1 dB losses for Bus sequence, and around 0.3dB to 0 dB losses from low bitrate to high bitrate for Football sequence.

We then compare the coding efficiency between the normal GOP with N=64 with the symmetric tree prediction structure, which has more than 4 times different maximum decoding delay as shown in Table 4-1. For the static sequence, symmetric tree prediction structure has very close performance compare with normal GOP structure in Mobile sequence, and has at most 0.3dB loss in Foreman sequences. In the fast motion sequences, it has around 0.4 dB losses for Bus sequence, and around 0.3dB to 0.1 dB losses from low bitrate to high bitrate for Football sequence.

From the above analysis, we know that comparing with normal GOP structure, symmetric tree prediction structure provide up to 1.2dB gain and at most 0.3dB loss when there are similar maximum decoding delay. And it has at most 0.4dB loss when it has less then one-fourth maximum decoding delay.

Bus

0 384 768 1152 1536 1920 2304 2688 3072 3456 bitrate

Figure 4-13 RD curve of Bus sequence with normal and symmetric tree structure comparison

0 384 768 1152 1536 1920

bitrate

Figure 4-14 RD curve of Foreman sequence with normal and symmetric tree structure comparison

Mobile

0 384 768 1152 1536 1920 2304 2688 3072 3456 3840 4224 4608 bitrate

Figure 4-15 RD curve of Mobile sequence with normal and symmetric tree structure comparison

0 384 768 1152 1536 1920 2304 2688 3072 3456 bitrate

Figure 4-16 RD curve of Football sequence with normal and symmetric tree structure comparison

Chapter 5 Conclusion

This thesis proposes a symmetric tree prediction structure that can generate an AVC compliant bitstream with low complexity VCR functionality supported at decoder.

Comparing with the normal GOP structure, it reduces the maximum random access decoding delay in a GOP with size N from linear functionality of N to logarithmic function of N. The prediction structure is symmetric between forward and backward prediction, which makes the backward playback has the same low complexity with forward playback. It separate the pictures into several levels and make the picture can only reference the picture at lower level, which make the fast playback can skip the higher level pictures therefore no redundant pictures will be decoded. We also propose a decoder that can fully support the VCR functionality with low decoding complexity.

The symmetric tree prediction structure can be configured to trade-off among lower decoding delay, fewer encoder buffer size, and better coding efficiency. We have discus this trade-off and a suggested configuration is proposed that provide good coding efficiency with reasonable complexity. From the simulation results, the symmetric tree prediction structure with suggested configuration can provide -0.3dB to +1.2dB coding efficiency difference comparing with the normal GOP structure with similar maximum decoding delay.

Based on the proposed structure, many issues can be further investigated. Different bit allocation method can be used to make the frame at lower level has better quality, and hence improve the overall coding efficiency. Different motion estimation search range can be use for the reference picture at different distance. Different prediction

structure variation can be investigated to reduce the decoding delay and improve the coding efficiency.

Chapter 6 Bibliography

[1]C Fu, Y Chan, W Siu “Macroblock-based reverse play algorithm for MPEG video streaming” 2004 ISCAS 753 ~ 756

[2]Y Tan, Y Liang, J Yu “Video transcoding for fast forward/reverse video playback” 2002 ICIP 713~716

[3]C Lin, J Zhou, J Youn, M Sun” MPEG video streaming with VCR-functionality” 2001 CSVT 415~425

[4]K C Yang, C Huang, J Wang “Restructuring GOP Algorithm to Reduce Video Server Load on VCR Functionality” 2003 ICCP

[5]W. Sweldens, “A custom-design construction of biorthogonal wavelets,” J. Appl.

Comp. Harm. Anal., vol. 3 (no. 2), pp. 186-200, 1996.

[6] ISO/IEC JTC1, “Call for Proposals on Scalable Video Coding Technology,”

ISO/IEC JTC1/WG11 Doc. N5958, Oct. 2003.

[7] http://www.vcodex.com

[8]T. Wiegand, G. J. Sullivan, G. Bjontegaard, A. Luthra ”Overview of H.264/AVC Video Coding Standard” IEEE Trans. Circuit Syst. Video Technol. Vol. 13 No. 7 pp560_576, July 2003.

[9] http://iphome.hhi.de/suehring/tml/download/