<s>
Small	B-Algorithm
object	I-Algorithm
detection	I-Algorithm
is	O
a	O
particular	O
case	O
of	O
object	B-General_Concept
detection	I-General_Concept
where	O
various	O
techniques	O
are	O
employed	O
to	O
detect	O
small	O
objects	O
in	O
digital	O
images	O
and	O
videos	O
.	O
</s>
<s>
In	O
areas	O
such	O
as	O
aerial	O
imagery	O
,	O
state-of-the-art	O
object	B-General_Concept
detection	I-General_Concept
techniques	O
under	O
performed	O
because	O
of	O
small	O
objects	O
.	O
</s>
<s>
Small	B-Algorithm
object	I-Algorithm
detection	I-Algorithm
has	O
applications	O
in	O
various	O
fields	O
such	O
as	O
Video	O
surveillance	O
(	O
Traffic	O
video	O
Surveillance	O
,	O
Small	B-Application
object	I-Application
retrieval	I-Application
,	O
Anomaly	B-Algorithm
detection	I-Algorithm
,	O
Maritime	O
surveillance	O
,	O
Drone	O
surveying	O
,	O
Traffic	B-Operating_System
flow	I-Operating_System
analysis	I-Operating_System
,	O
and	O
Object	B-Operating_System
tracking	I-Operating_System
.	O
</s>
<s>
Modern-day	O
object	B-General_Concept
detection	I-General_Concept
algorithms	O
such	O
as	O
You	O
Only	O
Look	O
Once(YOLO )	O
heavily	O
uses	O
convolution	O
layers	O
to	O
learn	O
features	B-Algorithm
.	O
</s>
<s>
Also	O
,	O
modern-day	O
detectors	O
,	O
such	O
as	O
YOLO	B-General_Concept
,	O
rely	O
on	O
anchors	O
.	O
</s>
<s>
Latest	O
versions	O
of	O
YOLO	B-General_Concept
(	O
starting	O
from	O
YOLOv5	O
)	O
uses	O
an	O
auto-anchor	O
algorithm	O
to	O
find	O
good	O
anchors	O
based	O
on	O
the	O
nature	O
of	O
object	O
sizes	O
in	O
the	O
data	O
set	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
models	O
have	O
billions	O
of	O
neurons	O
that	O
settle	O
down	O
to	O
some	O
weights	O
after	O
training	O
.	O
</s>
<s>
Data	B-General_Concept
augmentation	I-General_Concept
is	O
useful	O
technique	O
to	O
generate	O
more	O
diverse	O
data	O
from	O
an	O
existing	O
data	O
set	O
.	O
</s>
<s>
These	O
help	O
to	O
get	O
more	O
features	B-Algorithm
from	O
objects	O
and	O
eventually	O
learn	O
the	O
best	O
from	O
them	O
.	O
</s>
<s>
For	O
example	O
,	O
a	O
bike	O
object	O
in	O
the	O
1280	O
X	O
1280	O
resolution	B-Algorithm
image	O
has	O
more	O
features	B-Algorithm
than	O
the	O
640	O
X	O
640	O
resolution	B-Algorithm
.	O
</s>
<s>
Selecting	O
anchor	O
size	O
plays	O
a	O
vital	O
role	O
in	O
small	B-Algorithm
object	I-Algorithm
detection	I-Algorithm
.	O
</s>
<s>
YOLOv5	O
uses	O
a	O
K-means	B-Algorithm
algorithm	I-Algorithm
to	O
define	O
anchor	O
size	O
.	O
</s>
<s>
State-of-the-art	O
object	O
detectors	O
allow	O
only	O
the	O
fixed	O
size	O
of	O
image	O
and	O
change	O
the	O
input	O
image	B-Algorithm
size	I-Algorithm
according	O
to	O
it	O
.	O
</s>
<s>
The	O
tiling	O
approach	O
helps	O
when	O
an	O
image	O
has	O
a	O
high	B-Algorithm
resolution	I-Algorithm
than	O
the	O
model	O
's	O
fixed	O
input	O
size	O
;	O
instead	O
of	O
scaling	O
it	O
down	O
,	O
the	O
image	O
is	O
broken	O
down	O
into	O
tiles	O
and	O
then	O
used	O
in	O
training	O
.	O
</s>
<s>
Use	O
a	O
feature	O
pyramid	B-Algorithm
network	O
to	O
learn	O
features	B-Algorithm
at	O
a	O
multi-scale	O
:	O
e.g.	O
,	O
Twin	O
Feature	O
Pyramid	B-Algorithm
Networks	O
(	O
TFPN	O
)	O
,	O
Extended	O
Feature	O
Pyramid	B-Algorithm
Network	O
(	O
EFPN	O
)	O
.	O
</s>
<s>
FPN	O
helps	O
to	O
sustain	O
features	B-Algorithm
of	O
small	O
objects	O
against	O
convolution	O
layers	O
.	O
</s>
<s>
Hyper-parameters	B-General_Concept
define	O
their	O
dimensions	O
.	O
</s>
<s>
Various	O
deep	B-Algorithm
learning	I-Algorithm
techniques	O
are	O
available	O
that	O
focus	O
on	O
such	O
object	B-General_Concept
detection	I-General_Concept
problems	O
:	O
e.g.	O
,	O
Feature-Fused	O
SSD	O
,	O
YOLO-Z	O
.	O
</s>
<s>
Such	O
methods	O
work	O
on	O
"	O
How	O
to	O
sustain	O
features	B-Algorithm
of	O
small	O
objects	O
while	O
they	O
pass	O
through	O
convolution	O
networks.	O
"	O
</s>
