1. CABAC Based Bit Estimation for Fast H.264 RD Optimization Decision
Li Liu Xinhua Zhuang
CCNC 2009

2. Outline Introduction Based on CAVLC Based on CABAC
[1] Mohammed Golam Sarwer and Lai-Man Po, “Fast Bit Rate Estimation for Mode Decision of H.264/AVC,” Circuits and Systems for Video Technology, IEEE Transactions on,Vol.17, no. 10, pp , Nov
Based on CABAC
Experimental Results
Related paper

3. Introduction
Rate distortion optimization(RDO) is used to decide the best coding mode of a macroblock in H.264/AVC.
Efficient to improve coding performance.
Considerable complexity increase of the encoder.
Lagrange cost function:
Fast bit estimation is adopted to avoid entropy coding computation during the mode decision.

4. Introduction RDO computation Entropy coding DCT’ DCT & Q & Q’
Residue data
Rate
Distortion
RD cost calculation

5. Based on CAVLC
Estimate the number of bits for each of five different types of symbols of CAVLC separately.
CAVLC encoding of a block of transform coefficients proceeds as follow:
Encode the number of coefficients and trailing ones.
Encode the sigh of each TrailingOne.
Encode the levels of the remaining nonzero coefficients.
Encode the total number of zeros before the last coefficient.
Encode each run of zeros.

6. CAVLC example
Step1: Encode the number of coefficients and
trailing ones.
Step2: Encode the sigh of each TrailingOne.
Step3: Encode the levels of the remaining
nonzero coefficients.
Step4: Encode the total number of zeros
before the last coefficient.
Step5: Encode each run of zeros.
4x4 Block
Reordered Block: 0,3,0,1,-1,-1,0,1,0,0,0,0,0,0,0,0
TotalCoeff = 5; TotalZeros=3; T1s =3 (max value)
Encoding 1 2 3 4 5
Transmitted bitstream  (24 bits)

7. Based on CAVLC Encode the number of coefficients and trailing ones.
Four VLC tables are used for encoding coefficient token.
Bit consumption to encode the coefficient token is increased with number of coefficients.
Bit consumption to encode the coefficient token is decreased with number of trailing .
x-axis: number of nonzero coefficients
y-axis: true rate of Coeff_token

8. Based on CAVLC
The number of bits required to encode the coefficient token is :
is the total number of nonzero coefficients.
is the number of trailing .
are weighting constants.
Experiments show that at and better RD performance was found.

9. Encode the sigh of each TrailingOne
For each , a single bit encodes the sigh(0+, 1-).
The number of bits to encode the trailing ones is :

10. Encode the levels of the remaining nonzero coefficients
From the observation of level-VLC tables, it is shown that bit requirement is increased with magnitude of nonzero coefficients.
The number of bits to encode the level information is :
is the absolute values of all levels of quantized transform residual block.
is the sum of absolute values of all levels of quantized transform residual block.
is a positive constant.
Better results were found with
x-axis: SAT
y-axis: true rate of level

11. Encode the total number of zeros before the last coefficient.
From the observation of total zero VLC tables, it is shown that bit consumption to encode the total zero is increased with number of total zero.
The number of bits for total zero is :
is a positive constant.
is the total number of zeros before the last nonzero coefficients.
Better results were found with

12. Encode each run of zeros
After DCT, the high-frequency coefficient usually has small energy. By quantization, more zeros are found at the high-frequency position of quantized transform block, so the value of run for high-frequency nonzero coefficients is larger.
From the observation of run VLC tables, it is shown that more bits are required for large value of run, so bit consumption is higher to encode the run of high- frequency nonzero coefficient.
4x4 Block
Reordered Block: 0,3,0,1,-1,-1,0,1,0,0,0,0,0,0,0,0

13. Encode each run of zeros
The rate for run of each nonzero coefficients is :
is the frequency of th nonzero coefficient reordered block.
Example:
A string of coefficients [0,3,0,1,-1,-1,-1,0,1,0,0…]
frequency of (3) is 1
frequency of (1) is 7
Better results were found with

14. Total estimated bits
From previous analysis, the total estimated bits needed to encode a 4x4 residual block are:
Perfectly match
x-axis: Estimation error of a symbol is the
absolute difference between actual bit
rate and estimated bit rate of that
symbol.
Most of estimation error

15. Experimental Results JM 8.3 CAVLC is enabled Frame rate is 30
Motion vector search range is 32
Number of frame is 100
Percentage of difference coding time

16. Experimental Results All frames are I frames Foreman sequence
[9] C. H. Tseng, H. M.Wang, and J. F. Yang, “Enhanced intra 4x4 mode decision for H.264/AVC coders,” IEEE Trans. Circuits Syst. Video Technol., vol. 16, no. 8, pp. 1027–1032, Aug
Experimental Results
All frames are I frames
Foreman sequence

17. Experimental Results All frames are I frames: 47% time reduction
[15] Z. Chen, P. Zhou, and Y. He, “Fast integer pel and fractional pel motion
estimation for JVT,” Joint Video Team (JVT) Docs, JVT-F017, Dec
Experimental Results
All frames are I frames: 47% time reduction
IPP + fast motion estimation algorithm[15]
34% time reduction
IBPBP + fast motion estimation algorithm[15]
32% time reduction

18. Experimental Results
All I frames
IPP
IBPBP

19. Experimental Results Experiments with full search motion estimation
IPP + number of frames is set to 50
17% time reduction

20. Experimental Results Comparison with other method[9]
[9] C. H. Tseng, H. M.Wang, and J. F. Yang, “Enhanced intra 4x4 mode decision for H.264/AVC coders,” IEEE Trans. Circuits Syst. Video Technol., vol. 16, no. 8, pp. 1027–1032, Aug
Experimental Results
Comparison with other method[9]
Only intra 4x4 modes are used

21. Based on CABAC
The CABAC encoding process of a symbol consists of three elementary steps:
Binarization
Context modeling
Binary arithmetic coding

22. Flow diagram of the CABAC

23. Coding of residual data with CABAC
For each block:
coded_block_flag
0: The coded block flag is insignificant, no further information is transmitted.
1: a significance map specifying the positions of significant. The absolute value of the coefficient level as well as the sign is encoded for each significant transform coefficient.
Number of bits: 1 bit for each block.

24. Coding of residual data with CABAC
For each coefficient:
significant_coeff_flag
1: current scanning position contains a significant coefficient.(nonzero)
0: current scanning position contains a insignificant coefficient.(zero)
last_significant_flag
1: current significant coefficient is the last one inside the block.
0: current significant coefficient is not the last one inside the block.
Number of bits:
is the total number of zero coefficients before the last nonzero coefficient.
is the total number of nonzero coefficients.
and are weighted coefficients.

25. Coding of residual data with CABAC
For each significant coefficient:
coeff_abs_level_minus1
The absolute value of the significant coefficient level minus one.
coeff_sign_flag
The sigh of the significant coefficient level.
Number of bits:

26. Coding of residual data with CABAC
Estimate total number of bits to encode all symbol coeff_abs_level_minus1
is the absolute value of the i-th nonzero coefficient.
is estimated number of bits to encode .
stands for average number of bits encoding absolute level value k calculated from previous frames.
Total number of bits:

27. Coding of residual data with CABAC
The total number of bits estimated to code a coefficient block using CABAC entropy coding as:
where

28. Experimental Results
JM 10.1
Saving ratio of total encoding time

29. Experimental Results

30. Experimental Results 30% RDO time saving 0.02 dB less
0.22% bit rate reducing

31. Related paper
Jongmin Hahm and Chong-Min Kyung, “Efficient CABAC Rate Estimation for H.264/AVC Mode Decision, IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 20, NO. 2, FEBRUARY 2010
Arithmetic coding is replaced by a table lookup .