<s>
In	O
computer	B-Application
vision	I-Application
,	O
object	B-Algorithm
co-segmentation	I-Algorithm
is	O
a	O
special	O
case	O
of	O
image	B-Algorithm
segmentation	I-Algorithm
,	O
which	O
is	O
defined	O
as	O
jointly	O
segmenting	O
semantically	O
similar	O
objects	O
in	O
multiple	O
images	O
or	O
video	O
frames	O
.	O
</s>
<s>
It	O
is	O
often	O
challenging	O
to	O
extract	O
segmentation	B-Algorithm
masks	O
of	O
a	O
target/object	O
from	O
a	O
noisy	O
collection	O
of	O
images	O
or	O
video	O
frames	O
,	O
which	O
involves	O
object	B-General_Concept
discovery	I-General_Concept
coupled	O
with	O
segmentation	B-Algorithm
.	O
</s>
<s>
Early	O
methods	O
typically	O
involve	O
mid-level	O
representations	O
such	O
as	O
object	B-General_Concept
proposals	I-General_Concept
.	O
</s>
<s>
A	O
joint	O
object	O
discover	O
and	O
co-segmentation	O
method	O
based	O
on	O
coupled	O
dynamic	O
Markov	O
networks	O
has	O
been	O
proposed	O
recently	O
,	O
which	O
claims	O
significant	O
improvements	O
in	O
robustness	O
against	O
irrelevant/noisy	O
video	O
frames	O
.	O
</s>
<s>
Unlike	O
previous	O
efforts	O
which	O
conveniently	O
assumes	O
the	O
consistent	O
presence	O
of	O
the	O
target	O
objects	O
throughout	O
the	O
input	O
video	O
,	O
this	O
coupled	O
dual	O
dynamic	O
Markov	O
network	O
based	O
algorithm	O
simultaneously	O
carries	O
out	O
both	O
the	O
detection	O
and	O
segmentation	B-Algorithm
tasks	O
with	O
two	O
respective	O
Markov	O
networks	O
jointly	O
updated	O
via	O
belief	O
propagation	O
.	O
</s>
<s>
Specifically	O
,	O
the	O
Markov	O
network	O
responsible	O
for	O
segmentation	B-Algorithm
is	O
initialized	O
with	O
superpixels	O
and	O
provides	O
information	O
for	O
its	O
Markov	O
counterpart	O
responsible	O
for	O
the	O
object	B-General_Concept
detection	I-General_Concept
task	O
.	O
</s>
<s>
Conversely	O
,	O
the	O
Markov	O
network	O
responsible	O
for	O
detection	O
builds	O
the	O
object	O
proposal	O
graph	O
with	O
inputs	O
including	O
the	O
spatio-temporal	O
segmentation	B-Algorithm
tubes	O
.	O
</s>
<s>
Graph	B-Algorithm
cut	I-Algorithm
optimization	O
is	O
a	O
popular	O
tool	O
in	O
computer	B-Application
vision	I-Application
,	O
especially	O
in	O
earlier	O
image	B-Algorithm
segmentation	I-Algorithm
applications	O
.	O
</s>
<s>
In	O
addition	O
,	O
as	O
a	O
core	O
advantage	O
over	O
co-occurrence	O
based	O
approach	O
,	O
hypergraph	O
implicitly	O
retains	O
more	O
complex	O
correspondences	O
among	O
its	O
vertices	O
,	O
with	O
the	O
hyperedge	O
weights	O
conveniently	O
computed	O
by	O
eigenvalue	O
decomposition	O
of	O
Laplacian	B-Algorithm
matrices	I-Algorithm
.	O
</s>
<s>
In	O
action	B-Application
localization	I-Application
applications	O
,	O
object	B-Algorithm
co-segmentation	I-Algorithm
is	O
also	O
implemented	O
as	O
the	O
segment-tube	O
spatio-temporal	O
detector	O
.	O
</s>
<s>
Inspired	O
by	O
the	O
recent	O
spatio-temporal	O
action	B-Application
localization	I-Application
efforts	O
with	O
tubelets	O
(	O
sequences	O
of	O
bounding	O
boxes	O
)	O
,	O
Le	O
et	O
al	O
.	O
</s>
<s>
present	O
a	O
new	O
spatio-temporal	O
action	B-Application
localization	I-Application
detector	O
Segment-tube	O
,	O
which	O
consists	O
of	O
sequences	O
of	O
per-frame	O
segmentation	B-Algorithm
masks	O
.	O
</s>
<s>
Simultaneously	O
,	O
the	O
Segment-tube	O
detector	O
produces	O
per-frame	O
segmentation	B-Algorithm
masks	O
instead	O
of	O
bounding	O
boxes	O
,	O
offering	O
superior	O
spatial	O
accuracy	O
to	O
tubelets	O
.	O
</s>
<s>
This	O
is	O
achieved	O
by	O
alternating	O
iterative	O
optimization	O
between	O
temporal	O
action	B-Application
localization	I-Application
and	O
spatial	O
action	O
segmentation	B-Algorithm
.	O
</s>
<s>
Initialized	O
with	O
saliency	O
based	O
image	B-Algorithm
segmentation	I-Algorithm
on	O
individual	O
frames	O
,	O
this	O
method	O
first	O
performs	O
temporal	O
action	B-Application
localization	I-Application
step	O
with	O
a	O
cascaded	O
3D	O
CNN	B-Architecture
and	O
LSTM	B-Algorithm
,	O
and	O
pinpoints	O
the	O
starting	O
frame	O
and	O
the	O
ending	O
frame	O
of	O
a	O
target	O
action	O
with	O
a	O
coarse-to-fine	O
strategy	O
.	O
</s>
<s>
Subsequently	O
,	O
the	O
segment-tube	O
detector	O
refines	O
per-frame	O
spatial	O
segmentation	B-Algorithm
with	O
graph	B-Algorithm
cut	I-Algorithm
by	O
focusing	O
on	O
relevant	O
frames	O
identified	O
by	O
the	O
temporal	O
action	B-Application
localization	I-Application
step	O
.	O
</s>
<s>
The	O
optimization	O
alternates	O
between	O
the	O
temporal	O
action	B-Application
localization	I-Application
and	O
spatial	O
action	O
segmentation	B-Algorithm
in	O
an	O
iterative	O
manner	O
.	O
</s>
<s>
Upon	O
practical	O
convergence	O
,	O
the	O
final	O
spatio-temporal	O
action	B-Application
localization	I-Application
results	O
are	O
obtained	O
in	O
the	O
format	O
of	O
a	O
sequence	O
of	O
per-frame	O
segmentation	B-Algorithm
masks	O
(	O
bottom	O
row	O
in	O
the	O
flowchart	O
)	O
with	O
precise	O
starting/ending	O
frames	O
.	O
</s>
