MACROBLOCK SKIPPING ALGORITHMS FOR HIGH
DEFINITION H.264/AVC VIDEO CODING IN THE BASELINE
PROFILE
Susanna Spinsante, Ennio Gambi and Damiano Falcone
Dipartimento di Elettronica, Intelligenza artificiale e Telecomunicazioni
Universit
´
a Politecnica delle Marche, via Brecce Bianche 12, I-60131 Ancona, Italy
Keywords:
High Definition, macroblock skipping, complexity reduction.
Abstract:
This paper discusses different macroblock skipping algorithms to be applied in the H.264/AVC Baseline pro-
file, in order to facilitate the adoption of High Definition video coding in real time applications. Moving from
Standard to High Definition video coding, there is six times as much data to process: this motivates the search
for suited Mode Decision strategies, to reduce complexity while preserving an acceptable video quality for the
final user. The proposed schemes permit to speed up significantly the Mode Decision procedure, by forcing
the selection of the SKIP mode over each frame, without affecting significantly the final quality.
1 INTRODUCTION
In the framework of the H.264/AVC video coding
(ITU-T, 2005), a huge amount of fast Mode Deci-
sion strategies have been proposed (see, for example,
(Richardson, 2006)), with the aim of optimizing the
process in terms of computational resources and time,
without affecting the final perceived video quality.
Macroblock (MB) skipping algorithms usually play
a prominent role in fast Mode Decision schemes, as
they permit to increment the number of MBs encoded
as SKIP, that require the lower amount of computa-
tional resources. However, most of these proposals
for MB skipping and fast Mode Decision address the
Standard Definition (SD) video formats, whereas not
so much work has been carried out in the context of
High Definition (HD) video coding.
The H.264/AVC standard foresees the so-called
Fidelity Range Extension (FREXT) profile to support
specific applications, like content distribution, studio
editing and post processing. FREXT has also been ad-
dressed as the most proper profile for the management
of HD coding, but, given its extremely high number
of available options and its complexity, it cannot be
used in constrained environments, like those related
to the provisioning of real-time services (e.g. video-
conferencing). On the contrary, the aim of this pa-
per is to provide algorithms that can help introduc-
ing HD video coding in constrained environments, by
selecting the H.264/AVC Baseline profile. As a mat-
ter of fact, the Baseline profile seems the only one
able to meet requirements on the time delay, even
if not specifically conceived for supporting HD cod-
ing. Based on such premises, this paper begins show-
ing that the Baseline profile can be actually used for
HD coding, instead of the FREXT profile, without a
significant quality degradation, but with savings in
encoding time. Then, several MB skipping strate-
gies are presented and tested, that could be included
in the Baseline profile to permit further reduction of
the processing time, thus making real time HD cod-
ing more feasible, without significantly affecting the
final delivered video quality. HD formats provide
more visual information than any Standard Defini-
tion (SD) format; the increased pixel count inherent
in the HD formats supports better picture quality and
makes viewing images on larger screens clearer and
easier to watch. In videoconferencing, this enhances
the overall viewing experience and eliminates meet-
ing fatigue. Colors are more vibrant and realistic, and
movements are sharp and smooth. The higher amount
of available information, with respect to SD formats,
makes it reasonable to test the applicability of skip-
ping strategies, to reduce the computational burden
of the encoder, without affecting the perceived final
quality.
59
Spinsante S., Gambi E. and Falcone D. (2007).
MACROBLOCK SKIPPING ALGORITHMS FOR HIGH DEFINITION H.264/AVC VIDEO CODING IN THE BASELINE PROFILE.
In Proceedings of the Second International Conference on Signal Processing and Multimedia Applications, pages 59-66
DOI: 10.5220/0002135600590066
Copyright
c
SciTePress
The paper is organized as follows: Section 2
presents the results of a performance comparison be-
tween Baseline and FREXT profiles, when applied
to encode different HD video sequences. Section 3
provides an overview of Mode Decision and SKIP in
H.264/AVC, and discusses a number of macroblock
skipping algorithms, with the aim of drawing some
basic ideas, and tailoring them for use with HD se-
quences. Section 4 presents the performances ob-
tained by testing the new solutions within the Baseline
profile of the Reference Software version JM11.0; fi-
nally, the last section provides some remarks that con-
clude the paper.
2 APPLICABILITY OF THE
BASELINE PROFILE FOR HD
CODING
Before HD encoding/decoding, several video commu-
nication services, like videoconferencing, adopted the
Common Interchange Format (CIF), and other for-
mats such as 4CIF and 16CIF, with “full resolution”
being defined as 16CIF. Now, for HD videoconfer-
encing, the ITU-T recommends the formats listed in
Table 1.
Table 1: ITU-T recommended HD formats.
Format frames per second Resolution (16:9)
1080p 24,30 1290x1080
720p 24,30,60 1280x720
The progressive scan format (denoted by suffix p)
is an alternative to interlacing that improves picture
quality on larger screens, reduces jagged pictures, and
smoothes movement on large monitors.
In this section, a comparison between the Baseline
and the FREXT profiles, applied to encode a series of
High Definition video sequences, is presented. The
two profiles are described in detail in the standard and
the related documents; only the features of interest are
consequently reminded in the following. Comparison
is based on the evaluation of the computational time
required to encode the HD sequences, and the result-
ing Peak Signal to Noise Ratio (PSNR) of the Luma
(Y) component. As the computational time depends
on the test platform adopted, it has been expressed in
relative units (r.u.). The selected encoder configura-
tion is as follows:
Total number of frames: 100
Frame rate: 30 fps
Hadamard transform: Not Used
Image format: 1280x720p
Search range: 16
Number of reference frames:1, P-slice reference:1
Quantization Parameter: 28
Entropy coding: CAVLC (Baseline), CABAC
(FREXT)
RD Optimization: Enabled
Number of B frames: 0 (Baseline), 49 (FREXT)
A subset of the results obtained over a number of
different video sequences is reported in Table 2, for
the Shields, Car, and Park Run HD sequences, se-
lected as representatives for their motion features. As
shown, for each sequence, the resulting PSNR values
obtained with the two profiles differ by less than 0.3
dB, but the total encoding time required by FREXT is
more than 1.6 times that required by Baseline. This
confirms the idea that it is possible to adopt Baseline
also for HD coding, when dealing with real time ser-
vices, because the final quality remains acceptable,
whereas a significant reduction of the coding time
is achievable. The same conclusions can be drawn
by performing a subjective (i.e. visual) comparison
among the sequences encoded with the two profiles.
Table 2: Quality and computational time comparison be-
tween Baseline and FREXT profiles, for different HD video
sequences.
PSNR Y (dB) Time (r.u.)
Sequence Baseline FREXT Baseline FREXT
Shields 35.02 35.03 1.07 1.68
Car 37.26 37.48 1 1.60
Park Run 32.84 32.92 1.04 1.88
The above comparison shows that the adoption of
the Baseline profile for HD coding can be a viable so-
lution for introducing HD in real time services. How-
ever, the problem of making the coding process more
efficient, in order to meet strict time constraints, still
holds, as the total time required by the Baseline pro-
file to encode a HD sequence remains, typically, al-
most three times larger than that required for encod-
ing its corresponding CIF version.
Among the actions that can be performed to make
a video encoder work faster, a prominent role can be
played by the adoption of MB skipping algorithms
that can force the selection of the SKIP mode during
the Mode Decision process. The peculiarities of the
HD format (higher resolution and redundancy), and
the limitation of the human visual perceptivity, can be
taken into account to design ad hoc fast strategies.
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
60
3 MACROBLOCK SKIPPING
ALGORITHMS FOR FAST
MODE DECISION
One of the key functions in contributing to the compu-
tational time required by a video encoder is the Mode
Decision process. H.264 Mode Decision (Ivanov,
2006) utilizes a Lagrangian rate-distortion cost func-
tion J when computing mode costs for available MB
segmentation options:
min
{
J (s, c, MODE|QP, λ
MODE
)
}
(1)
J (s, c, MODE|QP, λ
MODE
) = SSD (s, c, MODE|QP) +
λ
MODE
· R (s, c, MODE|QP)
(2)
where QP is the MB quantization parameter, λ
MODE
is the Lagrangian multiplier, s and c represent the
original and the reconstructed MBs, respectively.
MODE is the mode chosen from the set of poten-
tial modes (SKIP, 16 × 16, 16 × 8, 8 × 16, 8 × 8,
8 × 4, 4 × 8, 4 × 4, Intra 4 × 4, Intra 16 × 16).
R(s, c, MODE|QP) is the transmitted bit rate asso-
ciated with MODE and QP, SSD is the Sum of the
Squared Differences. The Mode Decision algorithm
implemented in the Reference JM encoder sequen-
tially calculates J for all segmentation options. The
best coding mode, i.e. the one found only by an ex-
haustive evaluation of all the available modes, is that
giving the minimum rate-distortion (RD) cost, which
can be defined according with a Sum of Absolute Dif-
ferences (SAD) or SSD criterion.
In order to make a SKIP decision, after computing
all the modes, JM performs the following additional
checks:
the best mode selected is Inter 16 × 16
the reference frame is the previous one
the Motion Vector (MV) found is equal to the pre-
dicted MV
all the quantized transform coefficients are equal
to zero
The exhaustive search, necessary to detect the best
coding mode, can be a big burden for the encoder, es-
pecially when it has to deal with HD sequences and
their frame formats. On the other hand, experimental
tests show that, typically, more than half the MBs in
each frame of a sequence could be encoded as SKIP,
without any relevant loss in the final quality. Even if
not motivated by an analytical approach, this empiri-
cal consideration is commonly accepted as true, and
adopted as a guideline in the design of fast Mode De-
cision techniques. Taking into account the fact that
the SKIP mode is also the one having the lowest com-
putational complexity, as it does not encode any mo-
tion or residual information, it seems reasonable to
optimize the SKIP mode selection during the Mode
Decision process, to get a significant reduction in the
overall encoding time. While in the case of CIF and
QCIF sequences this could lead to an unacceptable
impact on quality, the larger amount of information
available in HD formats should compensate the risk
of quality loss. These considerations inspire the MB
skipping algorithms proposed in the paper: they rely
on some basic ideas, borrowed from the technical lit-
erature, that are briefly exposed in the following.
3.1 Macroblock Skipping Strategy
based on Skip Mode Distortion
Estimation
The algorithm presented in (Richardson, 2004) aims
at predicting MBs that are likely to be skipped by
the encoder. MBs that can be skipped with little
or no increase in distortion are not coded, resulting
in substantial computational savings, without signifi-
cantly affecting RD performance. Correct prediction
of skipped MBs can save significant computational
resources, since all of the subsequent processing of
the MB (Motion Estimation, transform and quantiza-
tion, entropy coding) can be avoided. For each MB,
the algorithm estimates the increase in distortion at
the decoder, if the MB were skipped, compared to the
distortion if the MB were coded. The distortion is
approximated by an energy measure, the Sum of Ab-
solute Errors (SAE), which is typically computed by
a video encoder, so it does not require further opera-
tions. Provided that:
SAE
MB
=
i, j
a
i, j
b
i, j
(3)
being a
i, j
an original Luma sample value, b
i, j
the
corresponding decoded Luma sample value, and i, j
ranging from 0 to 15, SAE increases with the decoded
image distortion. Introducing SAE
di f f
, the difference
in SAE between a skipped MB and a coded MB, as an
estimate of the increase in distortion due to skipping
a MB, we can say that a low value of SAE
di f f
implies
little advantage in coding and transmitting the MB,
whereas a high value of SAE
di f f
implies a significant
benefit in coding the MB, compared with skipping it:
SAE
di f f
= SAE
SKIP
SAE
noSKIP
(4)
SAE
SKIP
is the sum of absolute errors between the
uncoded MB and the luma values in the same po-
sition, in the reference frame. It is an approximate
measure of the distortion if the MB is skipped, since
MACROBLOCK SKIPPING ALGORITHMS FOR HIGH DEFINITION H.264/AVC VIDEO CODING IN THE
BASELINE PROFILE
61
the decoder copies the luma samples from the refer-
ence frame. SAE
noSKIP
is the SAE of the decoded
MB (compared with the original, uncoded MB) and
is an approximate measure of distortion, if the MB is
not skipped. SAE
SKIP
is calculated as the first step
of a Motion Estimation algorithm in the encoder and
is readily available at an early stage of processing of
each MB. SAE
noSKIP
is not normally calculated dur-
ing coding or decoding, and cannot be calculated if
the MB is actually skipped. Therefore, a model for
SAE
noSKIP
is used, to estimate SAE
di f f
. More specif-
ically, given a MB at position i in frame n, the value
of SAE
noSKIP
is set equal to the SAE of the most re-
cent available decoded MB in position i. Experimen-
tal results have shown this is a good predictor for
SAE
noSKIP
; the encoder has to compute and store this
value for each coded MB. Older values of SAE
noSKIP
for position i are replaced when a new MB is coded.
The MB skipping model compares SAE
di f f
to a
threshold: when lower, the current MB is skipped,
otherwise it is encoded. The threshold controls the
proportion of skipped MBs: a higher threshold results
in an increased number of skipped MBs, but also an
increased distortion, due to incorrectly skipped MBs.
3.2 Active Macroblock Skipping based
on Flexible Macroblock Order
The MB skipping strategy described in (Beesley,
2005) exploits one of the error resilience tools avail-
able in H.264/AVC, the Flexible Macroblock Order
(FMO), that allows a picture to be partitioned into
one or more slices, so that every MB in a missing
slice is likely to have all of its neighbours, from a cor-
rectly decoded slice, available for concealment pur-
poses at the decoder. According with the proposed
method, MBs selected for skipping are deliberately
removed during the encoding stage, in the knowledge
that they can be effectively concealed during the de-
coding phase. This provides a significant reduction
in bit stream size, with a resulting effect on quality
that, obviosuly, depends on the concealment options
activated at the decoder.
The proposed algorithm computes individual MB
PSNR values, along with the PSNR values should
each MB be concealed using weighted pixel value av-
eraging, i.e. one of the several concealment strategies
available. The resulting PSNR values are then com-
pared: if the concealed MB has an improved PSNR
over its decoded equivalent, then the MB is marked
for possible removal from the encoded bit stream. Ac-
tually, the MB is really removed only in the case all
its neighbouring MBs are not removed, because they
are necessary to perform concealment at the decoder.
The selection of the removable MBs is performed by
means of a list of all possible candidates, ordered by
their bit stream size, so that the MBs with the largest
potential savings are given the highest priority.
3.3 A Fast Algorithm for Inter-mode
Selection
In (Yu, 2004), the authors present an improved ver-
sion of the so-called Modified Fast Inter-mode selec-
tion (MFInterms) algorithm, to provide a more effi-
cient prediction of Mode Decision. The strategy in-
cludes temporal similarity detection, and the detec-
tion of different moving features within a MB.
The basic idea exploited by the suggested algo-
rithm is that a mode having a smaller partition size
may benefit detailed areas, whereas a larger partition
size is more suitable for homogeneous areas. Gener-
ally, a homogeneous MB is more likely to require ex-
amination of fewer inter-modes, compared to a highly
detailed MB. Two measurements are included in the
algorithm, targeted at MBs encoded with SKIP mode,
and MBs encoded by the inter-modes with larger de-
composed partition size (greater than 8 × 8 pixels):
the temporal similarity between two MBs, and the
motion consistency of a MB. Macroblocks coded with
the SKIP mode can be easily detected by comparing
the residue between the current MB and the previ-
ously encoded MB with a threshold, as follows:
S
residue
=
m
n
|
B
m,n,t
B
m,n,t1
|
(5)
T (S
residue
) =
1, S
residue
< T h
ASV
0, S
residue
> T h
ASV
(6)
where S
residue
is the sum absolute difference between
B
m,n,t
and B
m,n,t1
, which represent current and previ-
ous MBs, respectively. The temporal similarity is im-
plemented by means of an adaptive spatially varying
threshold, Th
ASV
, which depends on a constant and
the sum absolute difference of four nearest encoded
neighbours. This is motivated by the fact that, gener-
ally, the skipped MBs tend to occur in clusters, like in
a part of a static background; a MB undergoes tem-
poral similarity detection if one of the encoded neigh-
bours is a skipped MB.
Besides the temporal similarity computation, the
proposed algorithm suggests checking the motion
vector of each 8 × 8 block decomposed from a highly
detailed MB. If consistency among motion vectors ex-
ists, the inter modes with partition size greater than
8 × 8 are checked, otherwise, all possible inter modes
are searched.
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
62
3.4 Macroblock Skipping Algorithms
for HD Coding
The different approaches for MB skipping, previously
reviewed, have been applied to HD sequences, with
the aim of improving the efficiency of the Mode De-
cision process, and reducing the required coding time
in the Baseline profile. In order to make the follow-
ing discussion clearer, the algorithm discussed in 3.1
will be referred to as Alg
1
, the algorithm discussed in
3.2 will be referred to as Alg
2
, finally the algorithm
discussed in 3.3 will be referred to as Alg
3
.
The implementation of Alg
1
, in the case of HD
coding, is based on a SAE computation performed on
the Luma samples only. The SAE
noSKIP
value is de-
termined by subtracting the Luma samples of the last
non skipped MB, in a given position, from the Luma
samples of the same MB in the original frame. As ex-
pected, the SAE value gets higher for those MBs be-
longing to high motion parts of the frame. The thresh-
old used to control the amount of skipping in a frame
is fixed to a constant value, determined by experimen-
tal observations over many different sequences.
In order to apply the basic idea of Alg
2
to HD
frames, they are virtually divided into 9 areas, located
by the first two MB rows and columns, and by the last
two MB rows and columns, as shown in Figure 1.
AREA 1
AREA 7
AREA 3
AREA 9
AREA 5
AREA 2
AREA 8
AREA 4 AREA 6
Figure 1: Nine areas in a 1280 × 720 HD frame.
This structure allows to have a high number of avail-
able reference MBs in the central part of the frame
(AREA 5), where it is more probable to find high mo-
tion MBs, i.e. MBs for which the SKIP coding option
is less probable to occur. All the MBs included in
AREA 5 have also up to four available neighbouring
MBs, so that decision on their coding mode can be
performed by considering two controlling conditions:
the number of skipped MBs in the last frame avail-
able as a reference;
the dominant coding mode of the neighbouring
MBs, in the current frame, according with a ma-
jority rule.
The MBs located in AREA 5 can also be subjected
to the Mode Decision strategy presented in Alg
3
, be-
cause all their neighbouring MBs are available, as
shown in Figure 2. The adaptive threshold depends
on the residues of the neighbouring MBs, and on a
constant C, which will assume different values:
T h
ASV
= C · (S
N1
, S
N2
, S
N3
, S
N4
) (7)
being S
Ni
the residue of the i-th neighbour.
N
1
N
2
N
3
N
4
X
...
...
...
...
...
Figure 2: Relative positions of four nearest encoded neigh-
bours of the current MB.
The MB skipping strategies discussed up to this
point have been integrated within the Baseline pro-
file of the Reference Software JM11.0, for testing
purposes on the HD sequences Car, Shields, and
Park Run. The Car sequence presents a dominant
static component, the sky in the background, which
is mainly encoded as SKIP, whereas the part of the
frame showing a highway with passing cars is mainly
encoded as INTER (i.e. predicted). The proposed
schemes try to increase the number of skipped MBs
within this region. The Shields sequence is some-
how more “complicated”, because it is very rich in
details, so that increasing the number of SKIP MBs
could determine an unacceptable loss in quality. Fi-
nally, the Park Run sequence is a good representa-
tive of sequences rich in motion, as it features a mov-
ing subject, together with a movement of the camera
(panning) tracking it. Sample frames of the three se-
quences are shown in Figure 3.
4 PERFORMANCE EVALUATION
OF THE MB SKIPPING
ALGORITHMS ON HD
SEQUENCES
In order to test the performances provided by the MB
skipping algorithms previously discussed, at first they
have all been applied to encode each sequence in
MACROBLOCK SKIPPING ALGORITHMS FOR HIGH DEFINITION H.264/AVC VIDEO CODING IN THE
BASELINE PROFILE
63
a)
b)
c)
Figure 3: Sample frames of the: a) Car, b) Shields, and c)
Park Run sequences.
the Baseline profile, then, the one showing the best
behavior has also been tested with the Rate Distor-
tion Optimization (RDO) option activated, and with
a frame rate constrained at 15 fps. The experimental
results are reported in the following.
4.1 Tests on the Car Sequence
Table 3 shows the variation of the total encoding time,
and the quality loss, with respect to the Reference
Software, in the case of the Car sequence, for each
algorithm and some variants, i.e. different threshold
values for Alg
1
, different amounts of neighbor MBs
for Alg
2
, different values for the constant C in Alg
3
,
and combinations of Alg
1
and Alg
2
.
The joint adoption of Alg
1
(1) and Alg
2
(2) provides
the best trade-off between quality loss and coding
time reduction; in the case of HD sequences, a quality
loss of up to 0.2 dB is considered acceptable. For this
Table 3: Experimental results on the Car sequence.
Algorithm Coding PSNR Y
Time (%) (dB)
Alg
1
(1) -21.1 -0.79
Alg
1
(2) -30.3 -2.94
Alg
2
(1) -50.1 -6.97
Alg
2
(2) -29.3 -3.4
Alg
2
(3) -31.5 -3.46
Alg
3
(C = 0.5) -11.4 -0.33
Alg
3
(C = 0.85) -12.6 -0.64
Alg
1
(1)Alg
2
(2) -19.2 -0.16
Alg
1
(1)Alg
2
(2) -53.5 -7.92
case, a comparison between the Reference Software
and the Alg
1
(1)Alg
2
(2) option is provided in Fig-
ure 4, showing the average PSNR values of the Luma
component over each frame.
Reference Software
Alg
1
(1) Alg
2
(2)
35,5
36
36,5
37
37,5
38
38,5
39
0 50 100 150
frame #
Avg PSNR Y (dB)
Figure 4: Frame-by-frame PSNR Y comparison between
Reference Software and Alg
1
(1)Alg
2
(2) option, for the
Car sequence.
Fluctuations in the PSNR Y values, due to the
Alg
1
(1)Alg
2
(2) option, with respect to the Refer-
ence Software, can be explained by considering that
some MBs, corresponding to the moving cars, are
skipped anyway, because the majority of their neigh-
bours, belonging to the static highway in the back-
ground, gets skipped. This phenomenon causes a de-
crease in the average PSNR Y values, that becomes
quite high (over 1 dB) in some frames (30, 155).
The same coding option, Alg
1
(1)Alg
2
(2), has
been tested with RDO activated and a target frame
rate of 15 fps. The results obtained are reported in
Table 4. The adoption of the proposed skipping al-
gorithm provides remarkable improvements, with re-
spect to the Reference Software, by reducing the re-
quired coding time. This is a valuable issue when
dealing with time-constrained systems.
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
64
Table 4: Performance comparison between
Alg
1
(1)Alg
2
(2) and Reference Software, with RDO
activated, and a target frame rate of 15 fps (Car).
Option Coding PSNR Y
Time (%) (dB)
RDO -14.08 -0.16
15 fps -9.83 -0.16
4.2 Tests on the Shields Sequence
Following the same approach used for testing the MB
skipping algorithms on the Car sequence, Table 5 re-
ports the variation of the total encoding time, and the
quality loss, with respect to the Reference Software,
in the case of the Shields sequence.
Table 5: Experimental results on the Shields sequence.
Algorithm Coding PSNR Y
Time (%) (dB)
Alg
1
(1) -12.5 -0.48
Alg
1
(2) -13.2 -1.27
Alg
2
(1) -22.2 -0.17
Alg
2
(2) -13.8 0
Alg
2
(3) -13.9 0
Alg
3
(C = 0.5) -13.3 -0.28
Alg
3
(C = 0.85) -18.3 -1.27
Alg
1
(1)Alg
2
(2) -9.6 0
Alg
1
(1)Alg
2
(2) -19.4 -0.83
The best behaviour, in this case, is obtained by ap-
plying Alg
2
(1): the very low motion of the camera
permits to skip a good amount of MBs in some parts
of the frame that are quite homogeneous. This way, it
is possible to perform an efficient Mode Decision on
the basis of the neighbouring MBs. Figure 5 shows
the frame-by-frame trend of the average PSNR of the
Luma component, computed by applying Alg
2
(1),
with respect to the Reference Software.
There is a good match between the average PSNR
Y values determined by the algorithm, and those pro-
vided by the Reference Software, thanks to the fact
that the sequence under test is a low motion one, and
it does not present sudden variations of the scene. The
Alg
2
(1) coding option has been tested with RDO ac-
tivated and a target frame rate of 15 fps. The corre-
sponding performances are reported in Table 6; they
show not a remarkable improvement, with respect to
the Reference Software, meaning that the sequence
is handled similarly by the two encoder implementa-
tions.
4.3 Tests on the Park Run Sequence
The last series of results obtained by testing the pro-
posed algorithms are related to the Park Run HD se-
34,6
34,8
35
35,2
35,4
35,6
35,8
36
36,2
0 50 100 150
frame #
Avg PSNR Y (dB)
Alg
2
(1)
Reference Software
Figure 5: Frame-by-frame PSNR Y comparison between
Reference Software and Alg
2
(1) option, for the Shields se-
quence.
Table 6: Performance comparison between Alg
2
(1) and
Reference Software, with RDO activated, and a target frame
rate of 15 fps (Shields).
Option Coding PSNR Y
Time (%) (dB)
RDO -1.23 0
15 fps -2.97 0
quence. The variations of the total encoding time, and
the quality loss, with respect to the Reference Soft-
ware, for this last sequence under test, are reported in
Table 7.
Table 7: Experimental results on the Park Run sequence.
Algorithm Coding PSNR Y
Time (%) (dB)
Alg
1
(1) -9.8 -5.49
Alg
1
(2) -13.2 -10.17
Alg
2
(1) -9.5 0
Alg
2
(2) -9.8 0
Alg
2
(3) -9.8 0
Alg
3
(C = 0.5) -8.2 -0.22
Alg
3
(C = 0.85) -16.6 -3.48
Alg
1
(1)Alg
2
(2) -6.3 0
Alg
1
(1)Alg
2
(2) -7.5 -5.6
The Alg
2
(3) coding option provides a limited reduc-
tion of the total encodig time, without affecting the
final quality of the sequence. Many of the other al-
gorithms tested, though ensuring a stronger reduction
of the coding time, cause an unacceptable penalty of
the final quality; we could generalize this result to
high motion sequences, being Park Run a good rep-
resentative of them. For the algorithm providing the
best performances, a frame-by-frame comparison of
the average PSNR Y values is proposed in Figure 6.
The same MB skipping algorithm has been tested
also with RDO activated, and a target frame rate of 15
fps. The corresponding performances are reported in
Table 8. The application of the proposed skipping al-
MACROBLOCK SKIPPING ALGORITHMS FOR HIGH DEFINITION H.264/AVC VIDEO CODING IN THE
BASELINE PROFILE
65
Table 8: Performance comparison between Alg
2
(3) and
Reference Software, with RDO activated, and a target frame
rate of 15 fps (Park Run).
Option Coding PSNR Y
Time (%) (dB)
RDO -0.32 0
15 fps -2.07 0
gorithm provides slight improvements in constrained
systems, with a small reduction of the total encoding
time.
Reference Software
Alg
2
(3)
32,5
33
33,5
34
34,5
0 50 100 150
frame #
Avg PSNR Y (dB)
Figure 6: Frame-by-frame PSNR Y comparison between
Reference Software and Alg
2
(3) option, for the Park Run
sequence.
The last comparison presented in Figure 7 shows
the reduction of the coding time provided by some
of the proposed algorithms, for each sequence under
test. As a general remark, MB skipping algorithms
have a stronger effect on sequences presenting sud-
den variations of the scene, like Car, whereas their
performances get similar to those of the Reference
Software, when dealing with static or low motion se-
quences, like Shields.
5 CONCLUSION
This paper discussed different MB skipping algo-
rithms to increase the efficiency of the Mode Decision
process, when applying the Baseline profile of the
H.264/AVC standard for the real time coding of HD
sequences. The algorithms rely on three basic strate-
gies and their combinations. The results presented in
the paper show that different skipping schemes should
be applied, according with the main foreseen features
of each sequence, to get the best trade off between
coding time reduction and quality loss. The most re-
markable improvements are usually obtained for se-
4000
5000
6000
7000
8000
REFERENCE VICINI 3 SAD c=0,5 SAE E VICINI
r.u.
CAR PARK RUN SHIELD
8
7
6
5
4
r.u.
Ref. Soft.
Alg
2
(3) Alg
3
(C=0.5)
Alg
1
(1)
Alg
2
(2)
Figure 7: Coding time variations due to some of the pro-
posed MB skipping algorithms, for each HD sequence un-
der test.
quences presenting sudden variations of the scene, or
a high motion degree.
REFERENCES
International Telecommunication Union. (2005). Advanced
Video Coding for generic audiovisual services. In
Telecom. Standardization Sector of ITU. ITU Ed.
C. Kannangara, I. Richardson, M. Bystrom, J. Solera, Y.
Zhao, A. MacLennan and R. Cooney. (2006). Low-
Complexity Skip Prediction for H.264 Through La-
grangian Cost Estimation. In IEEE Trans. Circ. Sys.
Video Techn. Vol. 16, No. 2, pp. 202-208, February
2006. IEEE. 2006.
Y. Ivanov and C. J. Bleakley. (2006). Skip predic-
tion and early termination for fast mode decision in
H.264/AVC. In ICDT 2006, 2006 International Con-
ference on Digital Telecommunications. IEEE. 2006.
I. Richardson and Y. Zhao. (2004). Video Encoder Com-
plexity Reduction by Estimating Skip Mode Disortion.
In ICIP04, 2004 International Conference on Image
Processing. IEEE. 2004.
S. Beesley, A. Armstrong, C. Grecos, and D. Parish. (2005).
Active Macroblock Skipping in the H.264 Video Cod-
ing Standard. In 5th IASTED International Confer-
ence Visulization, Imaging, and Image Processing.
IASTED. 2005.
A. Yu, and G. Martin. (2004). Advanced Block Size
Selection Algorithm for Inter Frame Coding in
H.264/MPEG-4 AVC. In ICIP04, 2004 International
Conference on Image Processing. IEEE. 2004.
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
66