Chapter 2

Review of CAVLC, Arithmetic Coding,

and CABAC
2.1 Introduction of CAVLC
In the Baseline and Extended proles of H.264/AVC, except the xed-length cod-
ing, two VLC techniques are supported including: CAVLC for quantized transform
residues and Exp-Golomb coding for other syntax elements [Wiegand03]. In the pre-
vious standards, entropy coding of residues is based on the forward zig-zag scanned
run-length coding and xed variable-length coding. The H.264/AVC standard
adopts backward the zig-zag scanned run-length coding with adaptive VLC.
After intra/inter prediction and 44 transform and quantization of prediction
residuals, CAVLC [Bjntegaard02] is used to encode the quantized transformed
coefcients, and Universal Variable Length Coding (UVLC) is used to encode other
types of SEs.
The motivations of using CAVLC for SEs of transform coefcients are as fol-
lows: (1) Context adaptive coding can provide better coding efciency over the
whole quality range of the standard than UVLC technique; and (2) Separation of the
coding of Run and Level, due to possession of weak correlation, after zig-zag scan
can achieve better adaptivity, coding efciency, and lower the demand on memory
than that of Run-Level pair coding of UVLC.
After transformed and zig-zag scanned, the non-zero coefcients are often
sequences of 1. The CAVLC coder indicates the number of high frequency 1
coefcients by using trailing 1s (T1s). For non-zero coefcients, coding efciency
can be improved by using different VLC tables.
An example of CAVLC encoding of one 44 block is illustrated in Fig. 2.1. After
zig-zag scan, the reordered coefcients are coded as sequence of SEs, including
coeff_token, sign_T1s, 2 values for level, total_zeros, and 4 values for run_before.
These 5 types of SEs will be discussed as follows.
1. coeff_token is used to code both the total number of non-zero coefcients
(total_coeff) and the number of trailing ones (T1s). Note that the maximum
T1s allowed is 3. Any extra 1-valued coefcient is treated as normal non-zero
coefcient. In this case coeff_token = 0000100
2. sign of T1s is used to code the sign bit of each T1 in reverse zig-zag scanned
order. sign_of_T1 = 011
29 X. Tian et al., Entropy Coders of the H.264/AVC Standard, Signals and
Communication Technology, DOI 10.1007/978-3-642-14703-6_2,
C
Springer-Verlag Berlin Heidelberg 2011
30 2 Review of CAVLC, Arithmetic Coding, and CABAC
0 3 1 0
0 1 1 0
1 0 0 0
0 0 0 0
0,3,0,1,1,1,0,1,0...
Coefficients after
Zig-zag scan :
Total coef: 5
Trailing 1s: 3
Coeff_token
Coded SE SE Value Code
0000100
Sign_T1s +,, 011
Level +1
Level +3
Total_zeros 3
1
001,0
1 Run_before 0
111
10 Run_before 1
1 Run_before 0
01 Run_before 1
Run_before 1 Code not required
Fig. 2.1 CAVLC encoding of one transform coefcient block [Richardson03]
3. level (in sign and magnitude) is used to code the value of each of the remaining
non-zero coefcients in reverse order, with the highest frequency non-zero coef-
cient being encoded rst. level(1) = 1 (prex); and level(2) = 001 (prex), and
0 (sufx)
4. total_zeros is used to code the total number of zeros preceding the last non-zero
coefcient. total_zeros is 3 and coded by 111
5. run_before is used to code the number of successive zeros preceding each non-
zero coefcient in reverse zig-zag scanned order. run_before is coded with 10, 1,
1, 01.
The resulting encoded bit-stream is 0000100_011_1_001_0_111_10_1_1_01.
As shown in Fig. 2.2, both CAVLC decoding and encoding must be executed in
6 steps to completely process the 5 types of SEs of each coefcient block.
From the implementation viewpoint, the coeff_token is often obtained by looking
up in a 2-D LUT. The selection of LUT of luminance block is based on the number
of non-zero coefcients of the previously coded blocks positioning on top and on the
left of the block at issue. The sign_T1s are output with 1-bit/sign. The Level values
of the remaining coefcients are coded using selected 1-D LUTs. Each Level-LUT
Coeff_token (1)
Sign_T1s (2)
Prefix
Suffix
Level (3)
Total_zeros (4)
Run_before (5)
Coeff Block Reconstruction
(6)
Decoding Steps
Block scan to
generate SEs
(1)
Coeff_token (2)
Prefix
Suffix
Level (4)
Total_zeros (5)
Run_before (6)
Encoding Steps
Sign_T1s (3)
(a) (b)
Fig. 2.2 Flowchart for
CAVLC codec: (a) Decoder
and (b) Encoder
2.2 Pre-CABAC Arithmetic Coders 31
consists of Exp-Golomb codewords [Golomb66] for levels with small values (asso-
ciated with high probability), and escapes codewords for levels with large values
(low probability). The Run values of the block are coded in two steps in LUT: First,
the coding of total_zeros before the last non-zero coefcient, and second, the coding
of run_before each non-zero coefcient in the reverse scanned direction.
The coding efciency of the scheme is high, as the probability is high that one
or more low frequency coefcients have no Runs, especially in high bit rate coding.
Because the maximum of total_zeros is related to non-zero coefcient number, the
1D-LUTs of total_zeros are allocated separately for different coefcient number.
For run_before coding, the number of zeros left to be coded is utilized to select
LUT of zeros_left. Context adaptivity is introduced to CAVLC coding of all SEs,
except sign_T1s.
2.2 Pre-CABAC Arithmetic Coders
Compared to the previously reported lossless variable length coders (VLC)
[Moffat02] including Elias, Golomb and Rice, Shannon-Fano, and Huffman
[Huffman52], the distinct difference of arithmetic coding is that codewords can be
represented using fractional number of bits, while in other VLCs, each codeword
must occupy integer number of bits. Shannon rst mentioned the possibility of such
coding technique in 1948 [Shannon48]. Elias explores the idea of successive sub-
division of coding interval [Abramson63] in 1960s. Complete scheme of arithmetic
coding was proposed by Rissanen [Rissanen76] and Pasco [Pasco76] independently
in 1976, in which nite-precision arithmetic coding was implemented. Further
research works include hardware-oriented arithmetic coders [Langdon84] of IBM,
and software-oriented arithmetic coders by Witten et al.[Witten87], which made it
practical in the image and video compression applications.
Multiplication of Range
LPS
(1.38) of subdivision of arithmetic requires high
computation and limits the arithmetic coding throughput. Before CABAC, three
types of multiplication-free binary arithmetic coders have already been proposed
for image compression, which signicantly reduce computation complexity of arith-
metic coding. The pre-CABAC arithmetic coders include Q-coder [Pennebaker88],
and its variants QM coder [Mitchell93] and MQ coder [Taubman00, Taubman02].
QM coder is adopted as the entropy coder in Joint Bi-level Image Experts Group
(JBIG) standard [JBIG] and Joint Image Experts Group (JPEG) standard [JPEG],
while MQ coder is adopted in JPEG2000 standard [JPEG2000]. The Q-coder, QM
coder, and MQ coder are reviewed as follows.
2.2.1 Q-Coder
Q-coder is an adaptive binary arithmetic coder suitable for both HW and SW imple-
mentations. It has simple adapting mechanism which operates on the estimated
32 2 Review of CAVLC, Arithmetic Coding, and CABAC
probability of coding bin during the coding process. Q-coder avoids the increasing-
precision problem using xed-precision [Rissanen76]. Renormalization is applied
after interval subdivision to maintain the size and precision of coding interval.
Carry propagation during renormalization is solved by bit stufng. When 8 suc-
cessive bits of 1s are output, a 0 is stuffed as a carry gap. Multiplication is
avoided by using approximated value 1 of Range, because Range is in [0.75, 1.5).
Probability estimation is only taken when renormalization is needed for the encod-
ing bin. Computation reduction of interval subdivision and renormalization helps
accelerate encoding. However, the precision of arithmetic coding is sacriced and
compression efciency is degraded. The probability estimation is based on state
transition of Finite-state machine (FSM), which can be implemented as simple table
lookup operation.
2.2.2 QM Coder
QMcoder enhances the coding efciency of Q-coder by incorporation of conditional
exchange, and state machine for probability estimation using Bayesian estimation
principle for rapid initial learning. The differences between QM coder and Q-coder
include [Slattery98]: (a) MQ coder has higher precision with 3 more bits allocated
for the coding interval and other intermediate parameters; (b) Q-coder is hardware-
based, and QM coder is software-based; (c) Carry propagation is solved at decoder
of Q-coder, while it is solved at encoder of QM coder. Bit stufng is used in Q-
coder to prevent carry propagating in the encoder, which is faster than byte-stufng
of QM coder. However, the carry still needs to be solved in the decoder of Q-coder.
In comparison, QM coder resolves the carry propagation at encoder by byte stufng
technique, and processing of carry is not needed at decoder.
In the QM coder, adaptive probability estimation is implemented by estimation
of next Range
LPS
when renormalization is taken for the coding bin. The estimation
is performed by a table lookup operation, and the values of MPS and LPS can be
exchanged when the probability of LPS reaches 0.5. For binary arithmetic coding
of QM coder, bit assignment of Low and Range is shown in Fig. 2.3. Low regis-
ter consists of one carry bit, 8 bits of output byte, 3 bits of spacer, and 16 bits of
active region. Range register also has 16 bits of active region. During renormaliza-
tion process, a counter counts the number of valid bits in the output byte section of
Low register. When the output byte is full, the byte can be output or recorded as an
outstanding byte based on carry bit and whether the output byte is 0xFF. Carry bit of
Low register: c bbbb bbbb sss xxxx xxxx xxxx xxxx
Range register: 0 aaaa aaaa aaaa aaaa
O
u
t
p
u
t
b
y
t
e
A
c
t
i
v
e
r
e
g
i
o
n
S
p
a
c
e
r
b
i
t
s
C
a
r
r
y

b
i
t
Fig. 2.3 Bit assignment of
low and range in QM coder of
JPEG
2.3 CABAC of H.264/AVC 33
Low can only inuence the last byte of bytes in the output buffer. Three spacer bits
are allocated in Low between output byte and active region to reduce probability of
carry overow to the output buffer.
2.2.3 MQ Coder
In JPEG2000 [JPEG2000, Taubman02], MQ coder is utilized as the entropy coder
of the Embedded Block Coding with Optimized Truncation (EBCOT) to encode bit
planes of code block of quantized image subbands after Discrete Wavelet Transform
(DWT) and quantization. MQ coder is closely related to QM coder of JPEG. One
distinct difference between the two coders is that the carry is fully resolved in
QM coder of JPEG indicating a carry in the Low register will propagate to the
nearest bit 0 in the output bits. In comparison, carry propagation in MQ coder is
restricted by the mechanism that one bit of 0 is stuffed when the output byte is
0xFF. Any carry from the following renormalization operations cannot inuence
previous output bytes. Compared to the renormalization and output bit packing
mechanism of Q-coder and QM coder, bit stufng of MQ coder is more efcient
in bandwidth and processing time as only one redundant bit stufng is needed when
byte 0xFF is output, and the counter for outstanding bytes is not necessary. ASIC
design [Pastuszak05] is reported to accelerate the coding throughput of MQ coder
and EBCOT.
2.3 CABAC of H.264/AVC
Although Q-coder, MQ coder, and QM coder are also binary arithmetic coders with
statistics adaptivity features, the Context-based Adaptive Binary Arithmetic Coder
(CABAC) proposed by Marpe et al. in 2001 [Marpe01a, Marpe01b, Marpe01c]
is adopted as the entropy coding tool for the Main prole and higher proles of
the H.264/AVC standard. Before CABAC, LUT-based VLC are generally utilized
for the entropy coding stage in the hybrid block-based video coding standards
including H.263, MPEG-2, MPEG-4 Part 2. The weakness of VLCs is that cod-
ing event with probability larger than 0.5 cannot be efciently represented, and
the coding procedure is not adaptive to the actual symbol statistics as the values
of LUTs are xed [Marpe03a]. The only arithmetic coder adopted in video stan-
dard is of Annex E of H.263 [ITU-T Rec. H.263], in which coding efciency of
entropy coding is not signicantly improved, because it directly uses SEs of VLC
for arithmetic coding without redenition. Before CABAC proposals by [Marpe01a,
Marpe01b, Marpe01c] of H.264/AVC, similar arithmetic-coder-based implemen-
tations were rst investigated and applied to the non-block-based video coding
[Marpe99, Heising01], such as DWT.
CABAC [ISO/IEC 14496-10, Marpe03a] is the rst successful arithmetic cod-
ing scheme deployed in the video coding standard, with signicant compression
improvement compared to the previous entropy coding tools. As shown in Fig. 2.4,
34 2 Review of CAVLC, Arithmetic Coding, and CABAC
Bin String
Bypass
Bin
Bin &
Context
model
Coded
Bit stream
Bin Value
for
Context
Model
Update
Binarization
(1)
Input SE
Regular Bin
Coding
Engine
Bypass Bin
Coding
Engine
Binary
Arithmetic
Coding
(3)
(2)
Regular
Bin
Context
Modeling
Context Model
Selection & Access
Fig. 2.4 Block diagram of
the CABAC encoder of
H.264/AVC
CABAC encoding process consists of three elementary steps: binarization, context
modeling, and binary arithmetic coding (BAC). The input SEs are binarized into bin
strings, in which regular bins and bypass bins are encoded separately by the encod-
ing engines of BAC. For regular bin coding, the context model (probability model)
of the bin is prepared by the step of context modeling. Details of the three steps are
discussed in the following sections.
2.3.1 Binarization
Binarization maps the non-binary valued SE into bin string, which is a sequence of
binary decision (bin). Three types of bins are generated in the binarization step: reg-
ular bin, bypass bin, and terminate bin for the bins with unequal (variable), equal,
or dominant probabilities of value 1 and 0, respectively. Advantages of binarization
[Marpe03a] include: (a) the probability of non-binary SE can be represented by the
probabilities of individual coding bins, while compression efciency is not inu-
enced; (b) low-complexity BAC can be utilized; (c) context modeling at sub-symbol
(sub-SE) level provides more accurate probability estimation than context modeling
at the symbol level, and the alphabet of the encoder is reduced.
2.3 CABAC of H.264/AVC 35
Five binarization schemes are used in CABAC: Unary (U), Truncated Unary
(TU), kth order Exp-Golomb (EGk), concatenation of the rst and third scheme
(UEGk), and xed length binarization (FL). Kth ordered Exp-Golomb binarization
(EGk) [Teuhola78] a derivative of Golomb coding [Golomb66] is proved to be
an optimal prex-free coding for geometrically distributed sources. EGk codeword
consists of prex and sufx bin strings, with total length of 2l+k+1 bits. EGk prex
is a Unary codeword, with l bits of 1 and one terminating bit 0. The length l of string
of bit 1 is represented as:
l =

log
2

x
2
k
+1

(2.1)
The length of sufx binary string is equal to l + k, and the value of the sufx
string is:
EGk
sufx
= x +2
k
2
k+l
(2.2)
UEGk is combined binarization scheme of TU and EGk. It is utilized for
binarization of SEs of absolute value of residual coefcient level and MVD. TU
generates prex of the bin string, and EGk is adopted to generate the sufx with k
set to 0 and 3 for coefcient level and MVD, respectively. TU is simple and it per-
mits fast adaptation of probability of coding symbol. However, it is only benecial
for small SE values. For large SE values, sufx bin string generated by EGk pro-
vides a good t to the probability distribution, and bypass bin coding is also utilized
to reduce computation complexity.
2.3.2 Context Modeling
The context modeling step shown in Fig. 2.4 implements two sub-functions:
context model selection and context model access. The statistics of the coded SEs
are utilized to update the probability models (context model) [Wiegand03] of reg-
ular bins. For regular bin coding, one context model is chosen, and fetched from a
pre-dened set of context models to provide the probability of regular bin to be MPS
or LPS, and the context model is updated after bin coding based on bin value. The
context index (CtxIdx) is calculated to select context model, which is the sumof con-
text offset (CtxOffset) and context index increment (CtxIdxInc). CtxOffset locates the
context model set of processed SE, while CtxIdxInc selects one context model from
the set based on the values of coded bins or coded SEs of neighboring coded blocks.
The idea of multiplication-free arithmetic coding of H.264/AVC is based on the
assumption that estimated probability of each context model can be represented by a
sufciently limited set of representative values. In CABAC, the number of the repre-
sentative values is set to 64 to enable accurate estimation, which is larger than the 30
of Q-coder. Each context model contains a 1-bit tag of MPS value, and a 6-bit prob-
ability state index (pStateIdx) that addresses one of 64 representative probability
36 2 Review of CAVLC, Arithmetic Coding, and CABAC
values of LPS from p
0
to p
63
in the range of [0.01875, 0.5]. The probability values
of LPS are derived from (2.3). The ratio of two neighboring probability values is a
constant value , which is approximately equal to 0.949.
p

= p
1
for : = 1, . . . , 63, =

0.01875
0.5

1/63
, and p
0
= 0.5
(2.3)
Probability update of context model is based on the rule in (2.4), in which p
old
and p
new
are the probabilities for the bin to be LPS before and after bin coding. If
the coding bin is MPS, the probability of LPS decreases by simply multiplying the
ratio , while for the LPS bin, the update probability of MPS is calculated rst, and
then the probability of LPS is obtained.
p
new
=

max( p
old
, p
62
), if bin = MPS
1 (1 p
old
), if bin = LPS
(2.4)
By mapping the update probability value of LPS of (2.4) to the closest value in
the aforementioned set of representative values, multiplication of probability esti-
mation of CABAC is replaced by simple table lookup for the pStateIdx of next
probability state according to pStateIdx of the current bin, and based on whether it is
MPS or LPS. This probability value estimation of context state is actually the func-
tion state transition of FSM with 64 predened states. This type of probability FSM
is rst utilized in Q-coder, and adopted in QM coder and MQ coder. Compared to
Q-coder, QM coder, and MQ coder, the representative LPS probability values need
not to be stored in CABAC. Instead, the approximation of the products of coding
interval Range and the LPS probability of (1.37) are stored. In order to be more
adaptive to the coding context, the values of MPS and LPS can be exchanged when
the probabilities of MPS and LPS are equal and the coding bin is LPS.
For particular regular bins of CABAC, multiple context models are allocated for
single bin to more precisely represent probabilities of bin in different coding con-
texts. Four types of context model selection techniques are supported in CABAC,
based on (a) neighboring coded SE values of the current SE, (b) values of prior
coded bins of SE bin string, (c) position of the to-be-encoded residual coefcient
in the scanning path of residual block coefcients, and (d) level values of encoded
coefcients of residual block.
2.3.3 Binary Arithmetic Coding (BAC)
BAC performs arithmetic coding of each bin based on bin value, type, and the corre-
sponding context model of the bin. BAC is a recursive procedure of coding interval
subdivision and selection, as shown in Fig. 2.5.
2.3 CABAC of H.264/AVC 37
MPS
MPS
LPS
MPS
Range
Low
Range
Low
CABAC Bin Encoding Flow
Fig. 2.5 Coding interval subdivision and selection procedure of CABAC
Coding interval subdivision mechanism of CABAC is different from that of
QM and MQ coders. In QM and MQ coders, calculation of Range
LPS
of (1.38) is
simplied by using the approximated value 1 for Range, and multiplication is elim-
inated. In comparison, Range is also utilized for Range
LPS
calculation of CABAC.
Figure 2.6 shows the reference pseudo C-program of interval subdivision and selec-
tion of regular bin, in which the 2 higher-order bits of Range (bit 7 and bit 6) and
the index of probability state (pStateIdx) of LPS are used to look up for the pre-
calculated product of Range and p
LPS
of (1.37) in a 2-D LUT. Although the product
of LUT is of limited precision, the precision of Range
LPS
calculation, interval subdi-
vision is improved, and computational complexity is reduced in CABAC, compared
to those of QM and MQ coders.
RangeIdx = (Range >> 6) & 3; //Range[7:6]
Range
LPS
= rangeTableLPS[pStateIdx][RangeIdx];
Range
MPS
= Range Range
LPS
;
if(bin == MPS) //bin is MPS
Range = Range
MPS
;
else { //bin is LPS
Range = Range
LPS
;
Low = Low + Range
MPS
;
}
Fig. 2.6 Coding interval subdivision and selection of regular bin of CABAC
Because Range and Low of coding interval are represented by nite number of
bits (9 bits for Range and 10 bits for Low), it is necessary to renormalize (scale
up) the interval to prevent precision degradation, and the upper bits of Low are
output as coded bits during renormalization. Coding interval renormalization and
bit output of CABAC is based on Wittens algorithm [Witten87], as illustrated in the
reference pseudo C-program of Fig. 2.7. The coding interval of (Low, Low + Range)
is renormalized when Range is smaller than the threshold value 256 (0x100), which
is
1
/
4
of the maximum range of coding interval.
As illustrated in Fig. 2.7, renormalization of Range and Low is an iterative pro-
cedure, and the maximum number of iterations is 6, as the smallest possible value
of Range is 6. For the processing of carry propagation and output of coding bits,
the coded bits of CABAC are not output until it is conrmed that further carry
38 2 Review of CAVLC, Arithmetic Coding, and CABAC
while (Range < 0x100) {
if (Low >= 0x100) {
if (Low >= 0x200) {
//Output bit 0 and following outstanding bits of 1
PutBit(1);
Low = Low 0x200;
}
else { // Low is between 0x100 and 0x200
Low = Low 0x100;
NumOutstandingBits ++; //Accumulate outstanding bits
}
}
else {
//Output bit 0 and following outstanding bits of 1
PutBit(0);
}
//scale up Range and Low values by left shift
Range = Range << 1;
Low = Low << 1;
}
Fig. 2.7 Pseudo C-program of renormalization and bit output of CABAC
propagation will not inuence bit values. Figure 2.7 illustrates that only when inter-
val length (Range) is smaller than the threshold 0x100, one bit can be output if the
interval is located within the top half [0x200, 0x400) or bottom half [0, 0x200) of
the maximum coding range, or an outstanding (OS) bit is accumulated when the
interval is within [0x100, 0x300). When a bit of value X is output in BAC, the accu-
mulated OS bits are output with the value 1X. Compared to the bit stufng or byte
stufng schemes of Q-coder, QM coder, and MQ coder, carry propagation is com-
pletely solved during renormalization of BAC, and no additional processing of bit
stream is needed at CABAC decoder. Moreover, as no bits or bytes are stuffed in
the bit stream, the compression efciency of CABAC is further improved. However,
the renormalization illustrated in Fig. 2.8 is a highly sequential operation, and as the
number of iterations is variable depending on the selected subinterval Range, it is
1
3/4
1/2
1/4
0
Output
Bit 1
Output
Bit 0
Accumulate
Outstanding
Bit
Bit output when
Range < 1/4 of max interval range
(0x0)
(0x200)
(0x300)
(0x400)
Ratio(Value)
(0x100)
Fig. 2.8 Decision of bit
output and accumulation of
outstanding (OS)bit
2.3 CABAC of H.264/AVC 39
challenging for SW or HW acceleration of renormalization and bit output of BAC.
In some situations, long delay may be experienced, when a large number of OS bits
are accumulated.
2.3.4 Comparisons of CABAC with Other Entropy Coders
The coding efciency of CABAC is higher than those of the Q-coder, QM coder, and
MQ coder, because of (a) the more precise multiplication of Range
LPS
, (b) larger
number of probability states for each probability model, and more precise proba-
bility estimation of coding bins; and (c) more context models (probability models)
deployed for various coding contexts of different types of SEs.
Because of the high computational complexity of CABAC, another entropy
coding tool CAVLC [Bjntegaard02] is deployed in the Baseline prole and
Extended prole of H.264/AVC targeting low bit-rate real-time video coding. It
offers compression-complexity tradeoff with lower complexity, and lower coding
efciency, compared to CABAC [Marpe03a]. It is employed to encode the quantized
transform coefcients of 44 residual blocks, while zero-order Exp-Golomb codes
[Teuhola78] (EG0) are used for all other types of non-residual SEs. Adaptivity is
introduced to CAVLC by allowing switching among multiple VLC tables based on
the already processed SEs, and the coding efciency of CAVLC is better than those
of the previous VLCcoders which used single VLCtable. Instead of coding data pair
of run-level as a single SE, run and level of the residual block are encoded separately
in CAVLC, so that the inter-symbol redundancy can be more efciently exploited.
However, compression efciency of CABAC is signicantly higher, with typically
bit rate reduction of 914% in the video quality range of 3038 dB [Marpe03a],
compared to CAVLC and EG0. This is because (a) in CABAC, encoding symbols
can be more precisely represented in non-integer number of bits, especially for the
symbol with probability higher than 0.5, and (b) CABAC encoder is more adaptive
to the non-stationary symbol statistics with efcient context modeling (probability
estimation) for the coding bins of all types of SEs.
Since the adoption of CABAC entropy coding in H.264/AVC [Marpe97, Marpe,
Heising01, Marpe03a, Mrak03], CABAC is also applied in many applications of
image and video processing including motion mode and residual data of 3Ddynamic
mesh [Muller05], prediction residual in lossless 4D medical image compression
[Sanchez08], SEs of 88 transform coefcients of AVS coding standard [Zhang07],
motion vector coding of scalable video coder [Wu07], parameters of depth and cor-
rection vectors in multi-view video coding [Sehoon07]. CABAC is also utilized
to encode afne motion vector [Kordasiewicz07], and MVD of 3-D DWT-based
subband video encoder [Golwelkar07].
http://www.springer.com/978-3-642-14702-9