Graphics Reference
In-Depth Information
1.4
Orientation
In order to estimate the main direction of the key point, we mainly select pairs with
symmetric receptive fields with respect to the center, and sum the estimated local
gradients. As it shown in Figure 3, we select 33 points to compute the local gradients,
Let G be the set of all the pairs used to compute the local gradients:
r
1
r
2
1
PP
r
1
r
1
ʟʟ
ʟ=
((I(P
)
(I(P
)))
( 2 )
ʟ
ʟ
M
PP
r
1
r
2
PG
ʟʟ
ʟ
where M is the number of pairs in G and P ri
ʟ
is the 2D vector of the spatial coordi-
nates of the center of receptive field.
Fig 3. Illustration of the pairs selected to computed the orientation
2
Performance Evaluation
2.1
Dataset and Evaluation Criterion
To test the CS-FREAK descriptors performance, we build this new descriptor based on
the OpenCV FREAK codes, and conduct a comparative experiment with SIFT and
FREAK descriptors. We used the dataset introduced by Mikolajczyk and Schmid [7] as
the test image data sets, to eva-luate the rotation, change of scale, change of viewpoint,
blur (Gaussian), and brightness change. All algorithms are running on Intel (R) Core
(TM) i5-2.5 GHz, 2 G RAM with Window 7 + VS2010 + Opencv2.4.4. SIFT and
FREAK algorithm codes are provided by OpenCV2.4.4 library. And we present our
tests using the FAST detector also provided by OpenCV.
Descriptors are evaluated by means of precision-recall graphs as proposed in [13].
This criterion is based on the number of correct matches and the number of false
matches obtained for an image pair. A match is correct if the overlap error < 0.5.
#
correct matches
#
false matches
recall
=
1
precision
=
( 3 )
#
correspondences
#
all matches
where # correspondences is the ground truth number of matches.
 
Search WWH ::




Custom Search