Applied Mathematical Sciences, Vol. 7, 2013, no.

137, 6803 - 6824

HIKARI Ltd, www.m-hikari.com
http://dx.doi.org/10.12988/ams.2013.311644

Intra Compression Efficiency in VP9 and HEVC

Maxim P. Sharabayko
Postgraduate at Tomsk Polytechnic University
Junior Research Fellow at Tomsk State University of Control Systems and Radioelectronics
634050 Tomsk, Russia
[email protected]
Oleg G. Ponomarev
Assistant professor at Tomsk State University
Senior Research Fellow at Tomsk State University of Control Systems and Radioelectronics
634050 Tomsk, Russia
[email protected]
Roman I. Chernyak
Postgraduate at Tomsk State University
Junior Research Fellow at Tomsk State University of Control Systems and Radioelectronics
634050 Tomsk, Russia
[email protected]

Copyright
2013 Maxim P. Sharabayko, Oleg G. Ponomarev and Roman I. Chernyak.
This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.

Abstract
The amount of video data stored on local devices or transmitted
over the networks is permanently increased. The emerging of a more
efficient next generation video coding standard is of a high demand
at the moment. There seem to be two main contenders for the position of the next state-of-the-art video compression standard: JCT-VC
H.265/HEVC and Google VP9. The announced aim of HEVC is to
achieve twice more efficient compression compared to H.264/AVC, and
VP9 was developed to get half the bit-rate of VP8 with royalty-free video

6804

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

codec. Intra compression is one of the main features that determines
the compression efficiency of the whole codec.
In this paper we get into detailed overview of intra compression
data-flow in HEVC and VP9. We describe common and unique stages
of both standards. Then we carry out experiments with JCT-VC HM
and WebM VP9 encoders on intra compression efficiency. We also turn
some of the HEVC features off to get its dataflow as close to VP9 as
possible. Finally we get into discussion of the efficiency of both codecs,
the corresponding standards and their intra compression algorithms.

Mathematics Subject Classification: 94A08

Keywords: video compression, intra-prediction, Planar prediction, True
Motion, H.265/HEVC, VP9

Introduction

The current industrial video compression standard H.264/AVC was adopted

in 2003. It provides a superior video compression efficiency compared to other
existing and widely spread standards such as MPEG2 or VP8. However the
amount of video data stored on local devices or transmitted over the networks
is permanently increased. According to Cisco [2] mobile video traffic was 51
percent of the entire global Internet traffic by the end of 2012 and it is expected to be 66 percent by 2017. The ability to get better compression rates
will eventually decrease network bandwidth load.
The emerging of a more efficient next generation video coding standard is
of a high demand at the moment. There seem to be two main contenders for
the position of the next state-of-the-art video compression standard: JCT-VC
H.265/HEVC and Google VP9. HEVC is being developed by JCT-VC group the creators of AVC. It is an evolution of AVC concepts with some innovations.
On the other hand, VP9 is a Google initiative to get a royalty-free compression
standard with efficiency superior to AVC. It expands techniques used in AVC
and VP8 and is very likely to replace AVC at least in the YouTube video
service.
The basis of any video compression standard is intra-frame coding that
determines the resulting compression efficiency. In this paper we present a detailed overview of intra-frame compression techniques used in VP9 and HEVC.
Then we analyze their unique parts and carry out experiments on intra-frame
compression efficiency of HM and WebM VP9 encoders.

Intra compression efficiency in VP9 and HEVC

6805

General compression dataflow

Both HEVC and VP9 video compression standards are hybrid block-based
codecs relying on spatial transformations. General compression dataflow of
hybrid block-based encoders is illustrated in Fig.1. The input video frame is
initially partitioned into blocks of the same size called macroblocks. The compression and decoding process works within each macroblock. A macroblock is
subpartitioned into smaller blocks to perform prediction. There are two basic
types of prediction: intra and inter. Intra-prediction works within a current
video frame and is based upon the compressed and decoded data available for
the block being predicted. Inter-prediction is used for motion compensation:
a similar region on previously coded frames close to the current block is used
for prediction. The aim of the prediction process is to reduce data redundancy
and, therefore, not store excessive information in coded bitstream.

Figure 1: Hybrid block-based codec dataflow

Once the prediction is done, it is subtracted from the original data to
get residuals that should be compressed. Residuals are subject to Forward
Discrete-Fourier Transform (DFT). DFT translates spatial residual information into frequency domain. Thus the remaining spatial redundancy of this
information is partly reduced. Quantization is applied to the transformed
matrix to lose insufficient information. The insufficiency threshold is predetermined by encoder configuration. The remaining data and the steps applied
are subject to entropy coding, which makes it possible to get compressed bitstream.
For inter- and intra-prediction purposes the compressed data should be
restored in the encoder. The only data loss takes place after integer DFT
and quantization. Dequantization and inverse DFT are performed to restore
residuals. Then the restored residuals and the predicted values are summed
up to get restored pixel values, identical to those achieved in the decoder.
These restored values are used for intra-prediction within current video frame.
An additional frame post-processing stage is optionally applied to eliminate
image blockiness introduced by DFT and quantization. The final restored and

6806

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

post-processed video frame is stored in Decoded Picture Buffer (DPB) for inter
prediction of further frames.
VP9 and HEVC both utilize the described general compression dataflow,
but differ in details.

3
3.1

Intra compression in HEVC

Macroblock concept

The concept of macroblock in HEVC [5] is represented by the Coding Tree

Unit (CTU). CTU size can be 1616, 3232 or 6464, while AVC macroblock
size is 1616. Larger CTU size aims to improve the efficiency of block partitioning on high resolution video sequences (bitrate savings are about 16% [6]).
Larger blocks provoke the introduction of quad-tree partitioning (Fig. 2) of a
CTU into smaller coding units (CUs). A coding unit is a bottom-level quadtree syntax element of CTU splitting. The CU contains a prediction unit (PU)
and a transform unit (TU).

a)

b)

Figure 2: CTU splitting example with solid lines for CU split: a) with PU
splitting depicted as dotted lines; b) with TU splitting depicted as dotted
lines
The TU is a syntax element responsible for storing transform data. The
CU can be split in TUs in a quad-tree structure down to the smallest TU
size available. Allowed TU sizes are 3232, 1616, 88 and 44 with the
respectively sized DFT matrix.
The PU is a syntax element to store prediction data like the intra-prediction
angle or inter-prediction motion vector. The CU can contain up to four prediction units. CU splitting on PUs can be 2N2N, 2NN, N2N, NN, 2NnU,
2NnD, nL2N and nR2N (Fig. 3) where 2N is a size of a CU being split.
In the intra-prediction mode only 2N2N PU splitting is allowed. An NN
PU split is also possible for a bottom level CU that cannot be further split
into sub CUs.

Intra compression efficiency in VP9 and HEVC

6807

Figure 3: PU splitting

3.2

Intra-prediction modes

Figure 4: Prediction unit

To describe intra-prediction modes in HEVC let us assume the block being
predicted is a pixel matrix P = {p(x, y)}, where x = 0, . . . , w 1 and y =
0, . . . , h 1, with the size w h (Fig. 4). Intra-prediction in HEVC is
always performed on a square-sized pixel matrix, therefore let w = h = s.
intra-prediction of PU pixels may involve below-left (set E = {p(1, y)} when
y = s, . . . , 2 s 1), left (set D = {p(1, y)} when y = 1, . . . , s 1) above-left
(set A = {p(1, 1)}), above (set B = {p(x, 1)} when x = 1, . . . , s 1) and
above-right (set C = {p(x, 1)}, x = s, . . . , 2 s 1) neighboring pixels. The
availability of those pixels is determined by PU positioning.
To perform HEVC intra-prediction first the intra-prediction pattern R =
{r(i)} with i = 2 s, . . . , 2 s should be formed in the following way:
(
p(1, 1 i) if i < 0
r(i) =
p(i 1, 1)
if i > 0
In some cases (see Table 1) for luma blocks a filtered intra-prediction pattern R0 = {r0 (i)} with i = 2 s, . . . , 2 s is used, where

if i = 2 s
2 (r(i) + r(i + 1))
0
1
r (i) = 2 (r(i 1) + r(i))
if i = 2 s

1
(r(i 1) + 2 r(i) + r(i + 1)) otherwise
4

6808

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

In the following description of the intra-prediction process we address the

intra-prediction pattern just as r(i) considering that r0 (i) should be used in
some cases, determined in Table 1.

Mode

4
8
16
32

4
8
16
32

Table 1: Intra-prediction pattern filtering

planar DC 2-8 9 10 11 12-17 18
yes
no no no no no
no
no
yes
no no no no no
no
yes
yes
no yes no no no
yes yes
yes
no yes yes no yes yes yes

cases
19-24 25
no
no
no
no
yes
no
yes yes

26 27 28-34
no no
no
no no
no
no no
yes
no yes yes

Figure 5: HEVC angular intra-prediction modes

There is a total of 35 intra-prediction modes in HEVC: planar (mode 0),
DC (mode 1) and 33 angular modes (modes 2-34 on Fig.5).
DC intra-prediction is the most simple mode in HEVC. All PU pixels are
set equal to the mean value of all available neighboring pixels from sets B and
D. Thus the predicted pixel value is:
!
1
s
X
X
1
p(x, y) = DC =

r(i) +
r(i)
2s
i=s
i=1
The left column p(0, y) and the top row p(x, 0) pixel values of PU are
subjected to additional linear filtering:
1
p(0, y) = (r(y 1) + 3 DC + 2)
4
and

1
p(x, 0) = (r(x + 1) + 3 DC + 2).
4

Intra compression efficiency in VP9 and HEVC

6809

This operation is performed only for PU sizes 44, 88 or 1616 and aims to
smooth the transition from neighboring pixels to the DC value.
Planar intra-prediction is the most computationally expensive. It is a twodimensional linear interpolation. Fig.6 illustrates an example of planar prediction process. Each predicted pixel value p(x, y) is determined like:
p(x, y) =

1
(h(x, y) + v(x, y) + s)
2s

where
h(x, y) = (s x 1) r0 (y 1) + (x + 1) r0 (s)
is a horizontal interpolation of pixel value and
v(x, y) = (s y 1) r0 (x + 1) + (y + 1) r0 (s)
is a vertical interpolation of pixel value.

Figure 6: Illustration of planar intra-prediction interpolation process

Angular intra-prediction modes 2-34 are linear interpolations of pixel values
in the corresponding directions (Fig. 5). Vertical intra-prediction (modes 1834) is an up-down interpolation of neighboring pixel values. Let be an angle
between the vertical y-axis and the interpolation direction (Fig. 7). Therefore
it would have positive values for clockwise directions and negative values for
counterclockwise directions. If the (x, y) position is between two reference
pixels, then the predicted p(x, y) value is linearly interpolated with a 1/32 pixel
accuracy. The vertically predicted pixel value is determined like:


(32 ) r00 (x + 1 + i) + r00 (x + 2 + i) + 16
,
p(x, y) =
32
for x = 0, . . . , s 1 and y = 0, . . . , s 1, where i = 32(y+1)tg()
, =
32
b32 (y + 1) tg()c 32 i, and bc is an integer floor of a value. The reference

6810

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

sample r00 (i) is formed the following way:

(
r(x + 1)
x > 1

 1
r00 (x) =
r( 256 (256 x ctg()) 1) x < 1 and < 0

Figure 7: Vertical intra-prediction interpolation

Horizontal intra-prediction (modes 2-17) is basically the same as vertical,
except the prediction direction is left-to-right and angle is an angle between
the horizontal x-axis and the interpolation direction, taking positive values in
counterclockwise directions. The predicted pixel value p(x, y) is determined
by:


(32 ) r00 (y + 1 + i) + r00 (y + 2 + i) + 16
p(x, y) =
,
32
, =
for x = 0, . . . , s 1 and y = 0, . . . , s 1, where i = 32(x+1)tg()
32
00
b32 (x + 1) tg()c 32 i. The reference sample r (i) is formed the following
way:
(
r(x 1)
y > 1
 1

r00 (x) =
r( 256 (256 y ctg()) + 1) y < 1 and < 0
It is worth mentioning, that intra-prediction modes 2, 10, 18, 26 and 34
have = 0 and the prediction takes only one reference pixel for one predicted
pixel. Also prediction modes 10 and 26 do not use a filtered intra-prediction
pattern, but perform additional border filtering in cases of luma prediction of
sizes 44, 88 or 1616. For horizontal mode 10 the top row of predicted
pixels is:
1
p(x, 0) = r(1) + (r(x + 1) r(0))
2
and for vertical mode 26:
1
p(0, y) = r(1) + (r(y 1) r(0))
2

Intra compression efficiency in VP9 and HEVC

3.3

6811

Transform and quantization

The residuals after subtraction of predicted pixel values are subjected to

DFT and quantization. HEVC has an integer approximation of discrete cosine
transform and discrete sine transform (DCT II and DST VI by classification
in [1] respectively). A transform operation is specified for the 3232, 1616,
88 and 44 pixel matrix. DST is used for 44 intra coded blocks. DCT
transform matrices utilize symmetry properties that allow for using parts of the
butterfly structures used for fast FFT-like algorithms to reduce the number
of multiply/adds [3].
The quantization is inherited from AVC. The modifications are due to the
introduction of additional transform sizes.

3.4

Entropy coding

Entropy coding in HEVC is an evolution of context-adaptive binary arithmetic coding (CABAC) in AVC. The HEVC arithmetic coder is called syntaxbased context-adaptive binary arithmetic coder (SBAC) and has 27 context
models for syntax elements coding. Each context is adaptively changed based
on the data already processed. The principal modifications in SBAC compared to CABAC are related to the provision of parallel coding and decoding
possibility.

3.5

Post-processing

HEVC has two adaptive post-processing stages: deblock filtering (DBF)

and Sample-Adaptive Offset (SAO). Both stages are optional and can be
turned off.
The aim of the DBF is to reduce blockiness artifacts introduced after quantization of transformed coefficients. The filtering is applied at TU boundaries
with the absolute position multiple to 8. The vertical boundaries are processed
first, and the horizontal boundaries - second. A filtering algorithm is adaptive to fluctuations of pixel values near the boundaries. If the fluctuation is
low, then the strong filter is applied affecting 3 pixels on both sides of the
boundary. If the fluctuations are high, then the weak filtering is applied
affecting 2 pixels from both sides. When the boundary pixel value variance
within boundary segment is higher than the threshold, no filtering is applied.
The second post-processing stage is a nonlinear SAO transformation. It
was introduced in HEVC to compensate general losses in the decoded CTU.
The reconstructed CTU pixel values are modified by adding predetermined
offset values. These values are determined by the encoder and transmitted
in the bitstream. The affected pixels are chosen by their intensities which
determines the nonlinearity.

6812

4
4.1

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

Intra compression in VP9

Macroblock concept

The concept of macroblock in VP9 is represented by the superblock (Fig.

8). The size of a superblock is 6464 pixels and its subdivision is pretty much
like in HEVC: the partitioning also forms a quad-tree structure. Superblocks
can be subdivided down to 44 blocks. Each sub-block may be further split on
prediction blocks (Fig. 8, a) and transform blocks (Fig. 8, b). Unlike HEVC,
any sub-block can be split on prediction blocks in intra mode. Furthermore
rectangular intra-prediction blocks are possible. Intra-prediction in VP9 is still
performed on square regions thus rectangular prediction blocks represent two
square prediction blocks with the same prediction mode. Giving an analogy to
HEVC, prediction splitting 2N2N, NN, 2NN or N2N is available (Fig.
8, a), where 2N2N is the size of the block being split. It is worth mentioning
that 44 prediction blocks are determined within corresponding 88 block as
a group, unlike other prediction sizes when prediction data is stored per each
prediction block.
Like in HEVC, a sub-block can be split into transform blocks in a quad-tree
structure down to the smallest 44 block. The allowed DFT matrix sizes are
3232, 1616, 88 and 44 (Fig. 8, b).

a)

b)

Figure 8: Superblock splitting example with solid lines for block split: a)
with prediction splitting depicted as dotted lines; b) with transform splitting
depicted as dotted lines

Intra compression efficiency in VP9 and HEVC

4.2

6813

Intra-prediction modes

Figure 9: VP9 intra-prediction block

VP9 intra-prediction may involve neighboring pixels from left (set D =
{p(1, y)} when y = 1, . . . , s 1), above (set B = {p(x, 1)} when x =
1, . . . , s 1), above left (set A = {p(1, 1)}) and above right (set C =
{p(x, 1)}, x = s, . . . , 2 s 1) neighboring pixels (Fig. 9). The availability
of those pixels is also determined by block positioning. This is pretty much
like in AVC.

Figure 10: VP9 angular intra-prediction modes

There is a total of ten intra-prediction modes in VP9: DC, True Motion
(TM), and eight angular modes (H, V, D207, D153, D135, D117, D63 and D45
on Fig. 10).
To describe intra-prediction modes in VP9 let us do similar assumptions as
for HEVC where appropriate. Thus the block being predicted is a pixel matrix
P = {p(x, y)}, where x = 0, . . . , w 1 and y = 0, . . . , h 1, with the size
w h (Fig. 9) and s = w = h. An intra-prediction neighbors matrix R = r(i)

6814

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

with i = s, . . . , 2 s is smaller then for HEVC and is formed like:

(
i < 0 : p(1, 1 i)
r(i) =
i > 0 : p(i i, 1)
DC prediction in VP9 is much like in HEVC or AVC: each predicted pixel
is set equal to the mean value of neighboring pixels:
1
p(x, y) =

2s

1
X
i=s

r(i) +

s
X

!
r(i)

i=1

Unlike HEVC, no additional border filtering is performed in DC prediction

or any intra-prediction mode.
The True Motion intra-prediction mode in VP9 is an alternative to planar
mode in HEVC, but it is far easier. Each pixel value is calculated as follows:
p(x, y) = r(y 1) + r(x + 1) r(0)
Each directional intra-prediction mode (H, V, D207, D153, D135, D117,
D63 and D45) is a linear interpolation of pixel values in corresponding direction. For example, mode D45 is a 45-degree interpolation of pixel values,
mode D207 is a 207-degree interpolation of pixel values. All directional intraprediction modes can be described with the help of two functions:
(
r(i)
if i = s
b(i) = 1
(r(i 1) + r(i)) otherwise
2

if i = s
4 (3 r(i) + r(i + 1))
1
t(i) = 4 (3 r(i) + r(i 1))
if i = 2 s

1
(r(i 1) + 2 r(i) + r(i + 1)) otherwise
4
Bearing in mind the size of R is 3 s + 1, all 8 directional intra-prediction
modes can be determined by 2 s + 1 known r(i) values and 6 s precalcuated
b(i) and t(i) values, which may come out to be one of the intra-prediction
optimization techniques for VP9 encoder.
With the above equations, the D45 predicted pixel value can be described
as:
(
r(2 s)
if x = s 1 and y = s 1
p(x, y) =
t(x + y + 2) otherwise
and the example prediction matrix for an 88 block would be:

6815

Intra compression efficiency in VP9 and HEVC

p(x, y) =

t(2) t(3) t(4) t(5)

t(3) t(4) t(5) t(6)
t(4) t(5) t(6) t(7)
t(5) t(6) t(7) t(8)
t(6) t(7) t(8) t(9)
t(7) t(8) t(9) t(10)
t(8) t(9) t(10) t(11)
t(9) t(10) t(11) t(12)

t(6)
t(7)
t(8)
t(9)
t(10)
t(11)
t(12)
t(13)

t(7)
t(8)
t(9)
t(10)
t(11)
t(12)
t(13)
t(14)

t(8)
t(9)
t(10)
t(11)
t(12)
t(13)
t(14)
t(15)

t(9)
t(10)
t(11)
t(12)
t(13)
t(14)
t(15)
r(16)

The D63 intra-prediction is conducted the following way:

(


t 12 (y + 1) + x + 1 f or odd y


p(x, y) =
b 12 y + x + 2
f or even y

and the example prediction matrix for an 88 block would be:

p(x, y) =

b(2)
t(2)
b(3)
t(3)
b(4)
t(4)
b(5)
t(5)

b(3)
t(3)
b(4)
t(4)
b(5)
t(5)
b(6)
t(6)

b(4)
t(4)
b(5)
t(5)
b(6)
t(6)
b(7)
t(7)

b(5)
t(5)
b(6)
t(6)
b(7)
t(7)
b(8)
t(8)

b(6) b(7) b(8)

t(6) t(7) t(8)
b(7) b(8) b(9)
t(7) t(8) t(9)
b(8) b(9) b(10)
t(8) t(9) t(10)
b(9) b(10) b(11)
t(9) t(10) t(11)

b(9)
t(9)
b(10)
t(10)
b(11)
t(11)
b(12)
t(12)

The V intra-prediction mode stands for vertical interpolation direction,

when the predicted pixel has the following value:
p(x, y) = r(x + 1)
and the example prediction matrix for an 88 block would be:

r(1) r(2) r(3) r(4) r(5) r(6) r(7) r(8)

r(1) r(2) r(3) r(4) r(5) r(6) r(7) r(8)

r(1) r(2) r(3) r(4) r(5) r(6) r(7) r(8)

r(1) r(2) r(3) r(4) r(5) r(6) r(7) r(8)

p(x, y) =
r(1) r(2) r(3) r(4) r(5) r(6) r(7) r(8)

r(1) r(2) r(3) r(4) r(5) r(6) r(7) r(8)

r(1) r(2) r(3) r(4) r(5) r(6) r(7) r(8)

r(1) r(2) r(3) r(4) r(5) r(6) r(7) r(8)
The D117 intra-prediction is conducted the following way:

6816

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

1

f or even y and x > 12 y

b x 2 y


p(x, y) = t x 12 (y + 1) f or odd y and x > 12 (y + 1)

t (2 x y)
otherwise
and the example prediction matrix for an 88 block:

p(x, y) =

b(0)
t(1)
t(2)
t(3)
t(4)
t(5)
t(6)
t(7)

b(1)
b(2)
b(3)
t(0)
t(1)
t(2)
b(0)
b(1)
b(2)
t(1) t(0)
t(1)
t(2) b(0)
b(1)
t(3) t(1) t(0)
t(4) t(2) b(0)
t(5) t(3) t(1)

b(4)
t(3)
b(3)
t(2)
b(2)
t(1)
b(1)
t(0)

b(5)
t(4)
b(4)
t(3)
b(3)
t(2)
b(2)
t(1)

b(6)
t(5)
b(5)
t(4)
b(4)
t(3)
b(3)
t(2)

b(7)
t(6)
b(6)
t(5)
b(5)
t(4)
b(4)
t(3)

The D135 intra-prediction mode has the following function system (not
finished):
p(x, y) = t(x y)
and the example prediction matrix for an 88 block:

p(x, y) =

t(0)
t(1)
t(2)
t(3)
t(4)
t(5)
t(6)
t(7)

t(1)
t(0)
t(1)
t(2)
t(3)
t(4)
t(5)
t(6)

t(2)
t(1)
t(0)
t(1)
t(2)
t(3)
t(4)
t(5)

t(3)
t(4)
t(5)
t(6)
t(2)
t(3)
t(4)
t(5)
t(1)
t(2)
t(3)
t(4)
t(0)
t(1)
t(2)
t(3)
t(1) t(0)
t(1)
t(2)
t(2) t(1) t(0)
t(1)
t(3) t(2) t(1) t(0)
t(4) t(3) t(2) t(1)

t(7)
t(6)
t(5)
t(4)
t(3)
t(2)
t(1)
t(0)

The D153 directional prediction:

1

f or even x and y > 21 x

b 2 x y


p(x, y) = t 21 (x 1) y f or odd x and y > 12 (x 1)

t (x 2 y 1)
otherwise
with the example 88 prediction matrix:

6817

Intra compression efficiency in VP9 and HEVC

p(x, y) =

b(0)
b(1)
b(2)
b(3)
b(4)
b(5)
b(6)
b(7)

t(0)
t(1)
t(2)
t(3)
t(4)
t(5)
t(6)
t(7)

t(1)
b(0)
b(1)
b(2)
b(3)
b(4)
b(5)
b(6)

t(2)
t(0)
t(1)
t(2)
t(3)
t(4)
t(5)
t(6)

t(3)
t(1)
b(0)
b(1)
b(2)
b(3)
b(4)
b(5)

t(4)
t(2)
t(0)
t(1)
t(2)
t(3)
t(4)
t(5)

t(5)
t(3)
t(1)
b(0)
b(1)
b(2)
b(3)
b(4)

t(6)
t(4)
t(2)
t(0)
t(1)
t(2)
t(3)
t(4)

The horizontal intra-prediction mode, depicted as H, is just a row spread

of neighboring value:
p(x, y) = r(1 y)
with the example 88 prediction matrix:

r(1) r(1) r(1) r(1)

r(2) r(2) r(2) r(2)

r(3) r(3) r(3) r(3)

r(4) r(4) r(4) r(4)

p(x, y) =
r(5) r(5) r(5) r(5)

r(6) r(6) r(6) r(6)

r(7) r(7) r(7) r(7)

r(8) r(8) r(8) r(8)

r(1)
r(2)
r(3)
r(4)
r(5)
r(6)
r(7)
r(8)

r(1)
r(2)
r(3)
r(4)
r(5)
r(6)
r(7)
r(8)

r(1)
r(2)
r(3)
r(4)
r(5)
r(6)
r(7)
r(8)

r(1)
r(2)
r(3)
r(4)
r(5)
r(6)
r(7)
r(8)

t(5)
t(6)
t(7)
t(8)
r(8)
r(8)
r(8)
r(8)

The D207 intra-prediction mode has the following equation:

if x > 2 (s y) 1
r(s)
1
p(x, y) = b(y 2 x 1)
f or even x

1
t(y 2 (x + 1) 1) f or odd x
with the example 88 prediction matrix:

b(1) t(2) b(2) t(3)

b(2) t(3) b(3) t(4)

b(3) t(4) b(4) t(5)

b(4) t(5) b(5) t(6)

p(x, y) =
b(5) t(6) b(6) t(7)

b(6) t(7) b(7) t(8)

b(7) t(8) b(8) r(8)

b(8) r(8) r(8) r(8)

b(3)
b(4)
b(5)
b(6)
b(7)
b(8)
r(8)
r(8)

t(4)
t(5)
t(6)
t(7)
t(8)
r(8)
r(8)
r(8)

b(4)
b(5)
b(6)
b(7)
b(8)
r(8)
r(8)
r(8)

6818

4.3

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

Transform and quantization

The residuals after subtraction of predicted pixel values are subjected to

DFT and quantization. Transform blocks can be 3232, 1616, 88 or 44
pixels. VP9 uses integer approximation of DCT II and DST II for all transform sizes except for 44 transform where DST VI is used by classification in
[1]. In addition, VP9 introduces support for a new transform type, the Asymmetric Discrete Sine Transform (ADST), which can be used in combination
with specific intra-prediction modes. Intra-prediction modes that predict from
a left edge can use the 1-D ADST in the horizontal direction, combined with
a 1-D DCT in the vertical direction. Similarly, the residual signal resulting
from intra-prediction modes that predict from the top edge can employ a vertical 1-D ADST transform combined with a horizontal 1-D DCT transform.
Intra-prediction modes that predict from both edges such as the True Motion
mode and some diagonal intra-prediction modes, use the 1-D ADST in both
horizontal and vertical directions [4].

4.4

Entropy coding

VP9 uses 8-bit arithmetic coding engine from VP8 known as bool-coder.
Unlike AVC or HEVC, the probabilities of VP9 bool-coder do not change
adaptively within a frame. VP9 makes use of forward context updates through
the use of flags in the frame header that signal modifications of the coding
contexts at the start of each frame. These probabilities are stored in what is
known as a frame context. The decoder maintains four of these contexts, and
each frame specifies which one to use in bitstream.

4.5

Post-processing

There is only one possible post-processing stage in VP9: deblock filter. It

aims to reduce blockiness artifacts on superblocks filtering vertical edges first,
and horizontal edges second. VP9 has 16-, 8-, 4- and 2-pixels wide filters with
half filter size on each side of a boundary. VP9 also incorporates a flatness
detector in the loop filter that detects flat regions and varies the filter strength
and size accordingly.

Experimental results

The experimental comparison was carried out on the JCT-VC video sequences, listed in Table 2. They have different resolutions and frame-rates,
covering the most usecases possible. Bitrate-PSNR plots are illustrated in Fig.
1115. The summarized results are present in Table 3.

Intra compression efficiency in VP9 and HEVC

6819

Table 2: Test video sequence set

Sequence
Resolution, pixels Frame-rate, Hz Number of frames
BlowingBubbles
416240
50
500
BasketballDrill
832480
50
500
FourPeople
1280720
60
600
Kimono
19201080
24
240
PeopleOnStreet
25601600
30
150

For comparison purposes open-source implementations of the reviewed codecs

were used. HEVC compression efficiency was measured with the HM Test
Model. Verification of coding parameters was done with Elecard HEVC Analyzer. The HM encoder is configured with constant quantization mode when
all compressed frames has the same quantizer. This configuration eliminates
an influence of the rate-control efficiency on the encoder.
Evaluation of VP9 and VP8 compression performance was carried out with
the VPX encoder from The WebM Project as it is the only implementation
of this standard. The CodecVisa Analyzer was used for verification of compression parameters. The VPX encoder was configured with the constrained
quality mode and limited quantization parameter to emulate the constant
quantizer mode.
For AVC evaluation the JM reference encoder and Elecard Stream Analyzer
were used. It was also configured to work in the constant quantizer mode.
As stated in the overview, HEVC has more intra-prediction modes than
VP9. Some angular intra prediction modes in both standards have similar interpolation directions. To estimate a profit of using more prediction directions,
we modified the HM encoder implementation to use only 10 intra prediction
modes which alternatives are present in VP9. The HM encoder was allowed to
use planar (mode 0), DC (mode 1) and 8 angular prediction modes 5, 10, 15,
18, 21, 26, 31 and 34. The alternative VP9 modes are DC, True Motion, D207,
H, D153, D135, D117, V, D63 and D45 respectively. This HM modification is
depicted as HM less angles on Fig. 1115 and in Tables 34.
To bring HEVC even closer to VP9, in addition to the previous modification, we also turned SAO post-processing stage off because it is not present in
VP9. This HM configuration is depicted as HM less angles without SAO on
Fig. 1115 and in Tables 34.

6820

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

Figure 11: Bitrate-PSNR plot for intra-frame coding on BlowingBubbles sequence

On the BlowingBubbles sequence (Fig. 11) the HM encoder is 20.7% more
efficient than the VP9 encoder. Limiting HM intra prediction modes to 10
decreases HM compression efficiency by 11.8%. When we turned SAO off, the
HM compression efficiency was further reduced by 2.4%, but still the modified
HM encoder is 6.5% more efficient that the VP9 encoder.

Figure 12: Bitrate-PSNR plot for intra-frame coding on BasketballDrill sequence

On the BasketballDrill sequence (Fig. 12) the HM encoder is 13.9% more
efficient than the VP9 encoder. Again limiting HM intra prediction modes,

Intra compression efficiency in VP9 and HEVC

6821

the bitrate increases by 7%. Turning SAO off make the modified HM encoder
19.6% less efficient that the genuine version, but stil 6.6% more efficient than
the VP9 encoder.

Figure 13: Bitrate-PSNR plot for intra-frame coding on FourPeople sequence

On the FourPeople sequence (Fig. 13) the HM encoder is 16.8% more efficient than the VP9 encoder. Without 25 intra prediction modes the HM
encoder is 7.5% more efficient, and without SAO post-processing it still provides 6% lower bitrate compared to VP9.

Figure 14: Bitrate-PSNR plot for intra-frame coding on Kimono sequence

A bit different results were achieved on Kimono video sequence. The HM

6822

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

encoder is 12% more efficient than the VP9 encoder. However HM intra prediction modes limit does not have as much effect as on the other video sequences:
bitrate increase is only 2.4%. This is due to the structure of visual data in
this sequence. The dominant HM intra prediction modes for this sequence are
Planar, DC and vertical mode 26. All this modes are present in the modified
HM encoder therefore the impact of intra prediction restrictions is rather low.
If SAO is additionally turned off, the modified HM encoder is still 8.8% more
efficient compared to VP9.

Figure 15: Bitrate-PSNR plot for intra-frame coding on PeopleOnStreet sequence

The results for PeopleOnStreet video sequence are almost similar. The
HM encoder provides 14.5% bitrate savings compared to the VP9 encoder.
Turning several intra modes off makes it only 6.7% more efficient than VP9.
And turning SAO off makes it 4.9% more efficient than VP9.
Table 3: The HM
XXX
Codec
XXX
Sequence
XXX
BlowingBubbles
BasketballDrill
FourPeople
Kimono
PeopleOnStreet
Average

XXX

encoder bitrate compared to the VP9 encoder

HM
79,3%
86,1%
83,2%
88,0%
85,5%
84,4%

HM less angles HM less angles, SAO off

91,1%
93,1%
92,5%
90,4%
93,3%
92,1%

93,5%
93,4%
94,0%
91,2%
95,1%
93,4%

Intra compression efficiency in VP9 and HEVC

6823

The comparison of the HM encoder bitrate with the VP9 encoder results
at the same PSNR quality level is given in Table 3. The main HEVC bitrate
savings compared to VP9 are provided by additional intra prediction modes:
they comprise 7.7% on average. SAO post-processing stage provides only 1.3%
bitrate savings.
Table 4: HM and VP9 encoders bitrate compared to the JM encoder
XX
XXX
XX Codec
HM
VP9
Sequence XXXXX
BlowingBubbles
78.16% 92.19%
BasketballDrill
66.83% 85.19%
FourPeople
73.25% 89.02%
Kimono
72.31% 81.94%
PeopleOnStreet
68.65% 81.97%
Average
71.84% 86.06%
Both HM and VP9 are more efficienct compared to the JM encoder (Table
4) in intra frame compression. But while VP9 provides 14% bitrate decrease for
intra compression compared to the JM encoder, the HM encoder provides the
whole 28% bitrate decrease at the same visual quality on average. Considering
the aim of VP9 was to increase compression efficiency twice compared to VP8,
VP9 is about 18% more efficient than VP8. It is also worth mentioning that
VP9 provides about 28% bitrate savings compared to the VP8 encoder on the
Kimono video sequence. Also it seems to be no significant dependency on
resulution of the intra coded video sequence. Compression efficiency is mainly
determined by the structure and content of video frame.

Conclusion

Both VP9 and HEVC compression standards provide higher compressoin

efficiency compared to the current industrial video compression standard AVC.
HEVC provides better compression rates than VP9, but VP9 is patent-free and
can be used without licensing expenses.
In our experiments we tried to figure out the main reasons for HEVC to
be more efficient in intra coding compared to VP9. We showed that more
angular intra prediction modes provide the most significant influence on intra
compression efficiency (about 7.7% bitrate savings). The HEVC SAO postprocessing stage has less impact (about 1.3% bitrate savings). Finally the
modified HM encoder with 10 intra prediction modes and without SAO postprocessing is still about 7% more efficient compared to the VP9 encoder. We
assume this is due to the adaptive HEVC entropy coding. Syntax-adaptive

6824

M.P. Sharabayko, O.G. Ponomarev and R.I. Chernyak

BAC should be more efficient than frame-adaptive bool-coder, and this topic
is a subject for further research.
Further experiments should also be carried out on inter-compression efficiency of both standards.
Acknowledgements. The results were obtained at Tomsk State University of Control Systems and Radioelectronics as part of the complex project
Provision of multimedia broadcasting services in Internet public networks,
based on peer-to-peer network technology and adaptive data streaming with
the financial support of the Ministry of Education and Science of the Russian
Federation.

References
[1] V. Britanak, P.C. Yip and K.R. Rao, Discrete Cosine and Sine Transforms. General Properties, Fast Algorithms and Integer Approximation,
Academic Press, London, 2006.
[2] Cisco. 2013. Cisco Visual Networking Index: Global Data Traffic Forecast Update, 2012-2017, White Paper, February 2013.
[3] A. Fuldseth, G. Bjontegaard, M. Sadafale, M. Budagavi, Transform design
for HEVC with 16 bit intermediate data representation, Doc. JCTVCE243, Geneva, CH, 16-23 March, 2011.
[4] A. Grange, H. Alvestrand, A VP9 Bitstream Overview (Internet-Draft),
Google, August 2013.
[5] Recommendation H.265: High efficiency video coding, ITU-T, April 2013.
[6] O.G. Ponomarev, M.P. Sharabayko, A.A. Pozdnyakov, Research on video
compression methods and algorithms efficiency of H.265/HEVC standard,
Elektrosvyazj, 3 (2013), 29 - 33.
Received: November 1, 2013