<s>
Video	B-Algorithm
Multimethod	I-Algorithm
Assessment	I-Algorithm
Fusion	I-Algorithm
(	O
VMAF	B-Algorithm
)	O
is	O
an	O
objective	O
full-reference	O
video	B-Device
quality	I-Device
metric	O
developed	O
by	O
Netflix	O
in	O
cooperation	O
with	O
the	O
University	O
of	O
Southern	O
California	O
,	O
The	O
IPI/LS2N	O
lab	O
Nantes	O
Université	O
,	O
and	O
the	O
Laboratory	O
for	O
Image	O
and	O
Video	O
Engineering	O
(	O
LIVE	O
)	O
at	O
The	O
University	O
of	O
Texas	O
at	O
Austin	O
.	O
</s>
<s>
It	O
predicts	O
subjective	B-Device
video	I-Device
quality	I-Device
based	O
on	O
a	O
reference	O
and	O
distorted	O
video	O
sequence	O
.	O
</s>
<s>
The	O
metric	O
can	O
be	O
used	O
to	O
evaluate	O
the	O
quality	O
of	O
different	O
video	B-Application
codecs	I-Application
,	O
encoders	O
,	O
encoding	O
settings	O
,	O
or	O
transmission	O
variants	O
.	O
</s>
<s>
The	O
metric	O
is	O
based	O
on	O
initial	O
work	O
from	O
the	O
group	O
of	O
Professor	O
C.-C	O
.	O
Jay	O
Kuo	O
at	O
the	O
University	O
of	O
Southern	O
California	O
.	O
</s>
<s>
Here	O
,	O
the	O
applicability	O
of	O
fusion	O
of	O
different	O
video	B-Device
quality	I-Device
metrics	O
using	O
support	B-Algorithm
vector	I-Algorithm
machines	I-Algorithm
(	O
SVM	B-Algorithm
)	O
has	O
been	O
investigated	O
,	O
leading	O
to	O
a	O
"	O
FVQA	O
(	O
Fusion-based	O
Video	B-Device
Quality	I-Device
Assessment	O
)	O
Index	O
"	O
that	O
has	O
been	O
shown	O
to	O
outperform	O
existing	O
image	O
quality	O
metrics	O
on	O
a	O
subjective	B-Device
video	I-Device
quality	I-Device
database	O
.	O
</s>
<s>
Subsequently	O
renamed	O
"	O
Video	B-Algorithm
Multimethod	I-Algorithm
Assessment	I-Algorithm
Fusion	I-Algorithm
"	O
,	O
it	O
was	O
announced	O
on	O
the	O
Netflix	O
TechBlog	O
in	O
June	O
2016	O
and	O
version	O
0.3.1	O
of	O
the	O
reference	O
implementation	O
was	O
made	O
available	O
under	O
a	O
permissive	O
open-source	O
license	O
.	O
</s>
<s>
In	O
2018	O
,	O
a	O
model	O
that	O
predicts	O
the	O
quality	O
of	O
up	O
to	O
4K	B-Architecture
resolution	I-Architecture
content	O
was	O
released	O
.	O
</s>
<s>
VMAF	B-Algorithm
uses	O
existing	O
image	O
quality	O
metrics	O
and	O
other	O
features	O
to	O
predict	O
video	B-Device
quality	I-Device
:	O
</s>
<s>
The	O
above	O
features	O
are	O
fused	O
using	O
a	O
SVM-based	O
regression	O
to	O
provide	O
a	O
single	O
output	O
score	O
in	O
the	O
range	O
of	O
0	O
–	O
100	O
per	O
video	O
frame	O
,	O
with	O
100	O
being	O
quality	O
identical	O
to	O
the	O
reference	O
video	O
.	O
</s>
<s>
Due	O
to	O
the	O
public	O
availability	O
of	O
the	O
training	O
source	O
code	O
(	O
"	O
VMAF	B-Algorithm
Development	O
Kit	O
"	O
,	O
VDK	O
)	O
,	O
the	O
fusion	O
method	O
can	O
be	O
re-trained	O
and	O
evaluated	O
based	O
on	O
different	O
video	O
datasets	O
and	O
features	O
.	O
</s>
<s>
Anti-noise	O
signal-to-noise	O
ratio	O
(	O
AN-SNR	O
)	O
was	O
used	O
in	O
earlier	O
versions	O
of	O
VMAF	B-Algorithm
as	O
a	O
quality	O
metric	O
but	O
was	O
subsequently	O
abandoned	O
.	O
</s>
<s>
An	O
early	O
version	O
of	O
VMAF	B-Algorithm
has	O
been	O
shown	O
to	O
outperform	O
other	O
image	O
and	O
video	B-Device
quality	I-Device
metrics	O
such	O
as	O
SSIM	B-Algorithm
,	O
PSNR-HVS	O
and	O
VQM-VFD	O
on	O
three	O
of	O
four	O
datasets	O
in	O
terms	O
of	O
prediction	O
accuracy	O
,	O
when	O
compared	O
to	O
subjective	B-Device
ratings	I-Device
.	O
</s>
<s>
Its	O
performance	O
has	O
also	O
been	O
analyzed	O
in	O
another	O
paper	O
,	O
which	O
found	O
that	O
VMAF	B-Algorithm
did	O
not	O
perform	O
better	O
than	O
SSIM	B-Algorithm
and	O
MS-SSIM	O
on	O
a	O
video	O
dataset	O
.	O
</s>
<s>
In	O
,	O
where	O
its	O
various	O
versions	O
(	O
including	O
VMAF	B-Algorithm
NEG	O
)	O
were	O
tested	O
,	O
VMAF	B-Algorithm
outperformed	O
all	O
other	O
metrics	O
on	O
all	O
compression	O
standards	O
(	O
H.265	O
,	O
VP9	O
,	O
AV1	O
,	O
VVC	O
)	O
.	O
</s>
<s>
VMAF	B-Algorithm
scores	O
can	O
be	O
artificially	O
increased	O
without	O
improving	O
perceived	O
quality	O
by	O
applying	O
various	O
operations	O
before	O
or	O
after	O
distorting	O
the	O
video	O
,	O
sometimes	O
without	O
impacting	O
the	O
popular	O
PSNR	O
metric	O
.	O
</s>
<s>
A	O
reference	O
implementation	O
written	O
in	O
C	B-Language
and	O
Python	B-Language
(	O
"	O
VMAF	B-Algorithm
Development	O
Kit	O
,	O
VDK	O
"	O
)	O
is	O
published	O
as	O
free	B-Application
software	I-Application
under	O
the	O
terms	O
of	O
BSD+Patent	O
license	O
.	O
</s>
<s>
Its	O
source	O
code	O
and	O
additional	O
material	O
are	O
available	O
on	O
GitHub	B-Application
.	O
</s>
