<s>
PhyCV	B-Application
is	O
the	O
first	O
computer	O
vision	O
library	O
which	O
utilizes	O
algorithms	O
directly	O
derived	O
from	O
the	O
equations	O
of	O
physics	O
governing	O
physical	O
phenomena	O
.	O
</s>
<s>
Currently	O
PhyCV	B-Application
has	O
three	O
algorithms	O
,	O
Phase-Stretch	O
Transform	O
(	O
PST	O
)	O
and	O
Phase-Stretch	O
Adaptive	O
Gradient-Field	O
Extractor	O
(	O
PAGE	O
)	O
,	O
and	O
Vision	O
Enhancement	O
via	O
Virtual	O
diffraction	O
and	O
coherent	O
Detection	O
(	O
VEViD	O
)	O
.	O
</s>
<s>
PhyCV	B-Application
is	O
now	O
available	O
on	O
and	O
can	O
be	O
installed	O
from	O
.	O
</s>
<s>
Algorithms	O
in	O
PhyCV	B-Application
are	O
inspired	O
by	O
the	O
physics	O
of	O
the	O
photonic	O
time	O
stretch	O
(	O
a	O
hardware	O
technique	O
for	O
ultrafast	O
and	O
single-shot	O
data	O
acquisition	O
)	O
.	O
</s>
<s>
PST	O
is	O
an	O
edge	B-Algorithm
detection	I-Algorithm
algorithm	O
that	O
was	O
open-sourced	O
in	O
2016	O
and	O
has	O
800+	O
stars	O
and	O
200+	O
forks	O
on	O
GitHub	O
.	O
</s>
<s>
PAGE	O
is	O
a	O
directional	O
edge	B-Algorithm
detection	I-Algorithm
algorithm	O
that	O
was	O
open-sourced	O
in	O
February	O
,	O
2022	O
.	O
</s>
<s>
PhyCV	B-Application
was	O
originally	O
developed	O
and	O
open-sourced	O
by	O
@	O
UCLA	O
in	O
May	O
2022	O
.	O
</s>
<s>
In	O
the	O
initial	O
release	O
of	O
PhyCV	B-Application
,	O
the	O
original	O
open-sourced	O
code	O
of	O
PST	O
and	O
PAGE	O
is	O
significantly	O
refactored	O
and	O
improved	O
to	O
be	O
modular	O
,	O
more	O
efficient	O
,	O
GPU-accelerated	O
and	O
object-oriented	O
.	O
</s>
<s>
VEViD	O
is	O
a	O
low-light	O
and	O
color	O
enhancement	O
algorithm	O
that	O
was	O
added	O
to	O
PhyCV	B-Application
in	O
November	O
2022	O
.	O
</s>
<s>
The	O
code	O
in	O
PhyCV	B-Application
has	O
a	O
modular	O
design	O
which	O
faithfully	O
follows	O
the	O
physical	O
process	O
from	O
which	O
the	O
algorithm	O
was	O
originated	O
.	O
</s>
<s>
Both	O
PST	O
and	O
PAGE	O
modules	O
in	O
the	O
PhyCV	B-Application
library	O
emulate	O
the	O
propagation	O
of	O
the	O
input	O
signal	O
(	O
original	O
digital	O
image	O
)	O
through	O
a	O
device	O
with	O
engineered	O
diffractive	O
property	O
followed	O
by	O
coherent	O
(	O
phase	O
)	O
detection	O
.	O
</s>
<s>
In	O
the	O
implementation	O
of	O
PhyCV	B-Application
,	O
each	O
algorithm	O
is	O
represented	O
as	O
a	O
class	O
in	O
Python	B-Language
and	O
each	O
class	O
has	O
methods	B-Language
that	O
simulate	O
the	O
steps	O
described	O
above	O
.	O
</s>
<s>
PhyCV	B-Application
supports	O
GPU	O
acceleration	O
.	O
</s>
<s>
The	O
GPU	O
versions	O
of	O
PST	O
and	O
PAGE	O
are	O
built	O
on	O
PyTorch	B-Algorithm
accelerated	O
by	O
the	O
CUDA	B-Architecture
toolkit	O
.	O
</s>
<s>
The	O
running	O
time	O
per	O
frame	O
of	O
PhyCV	B-Application
algorithms	O
on	O
CPU	O
(	O
Intel	O
i9-9900K	O
)	O
and	O
GPU	O
(	O
NVIDIA	O
TITAN	O
RTX	O
)	O
for	O
videos	O
at	O
different	O
resolutions	O
are	O
shown	O
below	O
.	O
</s>
<s>
Note	O
that	O
the	O
PhyCV	B-Application
low-light	O
enhancement	O
operates	O
in	O
the	O
HSV	O
color	O
space	O
,	O
so	O
the	O
running	O
time	O
also	O
includes	O
RGB	O
to	O
HSV	O
conversion	O
.	O
</s>
<s>
When	O
dealing	O
with	O
real-time	O
video	O
streams	O
from	O
cameras	O
,	O
the	O
frames	O
are	O
captured	O
and	O
buffered	O
in	O
CPU	O
and	O
have	O
to	O
be	O
moved	O
to	O
GPU	O
to	O
run	O
the	O
GPU-accelerated	O
PhyCV	B-Application
algorithms	O
.	O
</s>
<s>
Currently	O
the	O
parameters	O
of	O
PhyCV	B-Application
algorithms	O
have	O
to	O
be	O
manually	O
tuned	O
for	O
differently	O
images	O
.	O
</s>
