DataAugmentation
ComposedProjectiveTransform implements efficient composition of
ProjectiveTransforms.
"""
ComposedProjectiveTransform(tfms...)
Wrap multiple projective `tfms` and apply them efficiently.
The projections are fused into a single projection and only
points inside the final crop are evaluated.
"""
struct
ComposedProjectiveTransform
{
T
<:
Tuple
}
<:
ProjectiveTransform
tfms
::
T
end
ComposedProjectiveTransform
(
tfms
...
)
=
ComposedProjectiveTransform
(
tfms
)
Composing
ProjectiveTransforms should create a
ComposedProjectiveTransform:
compose
(
tfm1
::
ProjectiveTransform
,
tfm2
::
ProjectiveTransform
)
=
ComposedProjectiveTransform
(
tfm1
,
tfm2
)
compose
(
composed
::
ComposedProjectiveTransform
,
tfm
::
ProjectiveTransform
)
=
ComposedProjectiveTransform
(
composed
.
tfms
...
,
tfm
)
compose
(
tfm
::
ProjectiveTransform
,
composed
::
ComposedProjectiveTransform
)
=
ComposedProjectiveTransform
(
tfm
,
composed
.
tfms
...
)
The random state is collected from the transformations that make up the
ComposedProjectiveTransform:
getrandstate
(
tfm
::
ComposedProjectiveTransform
)
=
getrandstate
.
(
tfm
.
tfms
)
To combine the concrete projections
P, we use
ImageTransformations.compose (aliased to
∘). The bounds are also transformed at each step.
function
getprojection
(
composed
::
ComposedProjectiveTransform
,
bounds
;
randstate
=
getrandstate
(
composed
)
)
P
=
CoordinateTransformations
.
IdentityTransformation
(
)
for
(
tfm
,
r
)
in
zip
(
composed
.
tfms
,
randstate
)
P_tfm
=
getprojection
(
tfm
,
bounds
;
randstate
=
r
)
bounds
=
projectionbounds
(
tfm
,
P_tfm
,
bounds
;
randstate
=
r
)
P
=
P_tfm
∘
P
end
return
P
end
function
projectionbounds
(
composed
::
ComposedProjectiveTransform
,
P
,
bounds
;
randstate
=
getrandstate
(
composed
)
)
return
transformbounds
(
bounds
,
P
)
end