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.

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

 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

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) +

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

∑

i, j



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

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:

residue

∑

m,n,t

− B

m,n,t−1

(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

m,n,t

and B

m,n,t−1

, 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

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

, the algorithm discussed in

3.2 will be referred to as Alg

, ﬁnally the algorithm

discussed in 3.3 will be referred to as Alg

The implementation of Alg

, 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

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

, 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

, S

) (7)

being S

the residue of the i-th neighbour.

...

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

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

, different amounts of neighbor MBs

for Alg

, different values for the constant C in Alg

and combinations of Alg

and Alg

The joint adoption of Alg

(1) and Alg

(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) -21.1 -0.79

Alg

(2) -30.3 -2.94

Alg

(1) -50.1 -6.97

Alg

(2) -29.3 -3.4

Alg

(3) -31.5 -3.46

Alg

(C = 0.5) -11.4 -0.33

Alg

(C = 0.85) -12.6 -0.64

Alg

(1)∧Alg

(2) -19.2 -0.16

Alg

(1)∨Alg

(2) -53.5 -7.92

case, a comparison between the Reference Software

and the Alg

(1)∧Alg

(2) option is provided in Fig-

ure 4, showing the average PSNR values of the Luma

component over each frame.

Reference Software

Alg

(1) ∧ Alg

(2)

35,5

36,5

37,5

38,5

0 50 100 150

frame #

Avg PSNR Y (dB)

Figure 4: Frame-by-frame PSNR Y comparison between

Reference Software and Alg

(1)∧Alg

(2) option, for the

Car sequence.

Fluctuations in the PSNR Y values, due to the

Alg

(1)∧Alg

(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)∧Alg

(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

Table 4: Performance comparison between

Alg

(1)∧Alg

(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) -12.5 -0.48

Alg

(2) -13.2 -1.27

Alg

(1) -22.2 -0.17

Alg

(2) -13.8 ≈ 0

Alg

(3) -13.9 ≈ 0

Alg

(C = 0.5) -13.3 -0.28

Alg

(C = 0.85) -18.3 -1.27

Alg

(1)∧Alg

(2) -9.6 ≈ 0

Alg

(1)∨Alg

(2) -19.4 -0.83

The best behaviour, in this case, is obtained by ap-

plying Alg

(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

(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

(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

35,2

35,4

35,6

35,8

36,2

0 50 100 150

frame #

Avg PSNR Y (dB)

Alg

(1)

Reference Software

Figure 5: Frame-by-frame PSNR Y comparison between

Reference Software and Alg

(1) option, for the Shields se-

quence.

Table 6: Performance comparison between Alg

(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) -9.8 -5.49

Alg

(2) -13.2 -10.17

Alg

(1) -9.5 ≈ 0

Alg

(2) -9.8 ≈ 0

Alg

(3) -9.8 ≈ 0

Alg

(C = 0.5) -8.2 -0.22

Alg

(C = 0.85) -16.6 -3.48

Alg

(1)∧Alg

(2) -6.3 ≈ 0

Alg

(1)∨Alg

(2) -7.5 -5.6

The Alg

(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

Table 8: Performance comparison between Alg

(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

(3)

32,5

33,5

34,5

0 50 100 150

frame #

Avg PSNR Y (dB)

Figure 6: Frame-by-frame PSNR Y comparison between

Reference Software and Alg

(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

r.u.

Ref. Soft.

Alg

(3) Alg

(C=0.5)

Alg

(1)

∧

Alg

(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