Graphics Reference
In-Depth Information
0.2
0.2
0.15
0.15
0.1
0.1
0.05
0.05
0
0
20
40
60
80
20
40
60
80
σ
σ
Figure 4.4.
Selecting the characteristic scale of a feature using the normalized Laplacian. Top
row: original images with manually-selected center locations (white dots). Bottom row: the nor-
malized Laplacian as a function of scale. The characteristic scale
σ
that maximizes the normalized
Laplacian is used as the radius of the corresponding circle in the top row. The ratio between the
two characteristic scales is 2.64, which is almost the same as the actual underlying scale factor
of 2.59 relating the images.
of the circle in the top row). We will discuss the normalized Laplacian further in the
next section.
Mikolajczyk and Schmid [
325
] adopted this approach to compute what they called
Harris-Laplace features
. We use the same two steps as previously shown to detect
Harris corners at each scale, and add the additional step
6
3. For each detected feature (say at scale
k
n
0
), retain it only if its normalized
Laplacian is above a certain threshold, and it forms a local maximum in the
scale dimension, that is:
σ
x
,
y
,
k
n
x
,
y
,
k
n
−
1
x
,
y
,
k
n
x
,
y
,
k
n
+
1
NL
(
σ
0
)>
NL
(
σ
0
)
and
NL
(
σ
0
)>
NL
(
σ
0
)
(4.22)
6
A slight modification that gives higher localization accuracy but has more computational cost was
described in [
327
].
