Graphics Reference
In-Depth Information
votes .The FREAK model is simplified to 5-layer with 25 sampling points, which can
simplify the calculation and at the same time form a sufficient amount of the votes, as
it shown in Figure 1,the detailed description of our matching strategy based on two-
step voting is introduced as the following.
Extract the key point and simplify the FREAK descriptor. The simplified descrip-
tor is generated according to the normal FREAK descriptor generation method, with
the identification model proposed in [12], which is more reasonably discerning to
quantify the differences between the test sequences.
exp(
−
M (a))
2
D
=
,..0 a N,
s t
≤ <
arg maxR(a, b)
0
(
1
)
≤< ≠
b
; a
b
aN
T
(P )
a
,
where
M
(a)
=
0
≤<
−
0.5
R
(a, b)
is the correlation between test se-
N
quence
a
and
b
.
Sort the sampling pairs by
D
, and select the first N points binary test
results (we take N=64 in the experiment) as the simplified binary descriptors. Use the
hamming distance as similarity measure method, looking for their
m
nearest key
points from images to be matched. This process obtained the crude matching results.
(a)FREAK sampling modal (b)our simpling modal
Fig. 1.
The sampling modal
1.3
Center Symmetry FREAK and Voting Strategy
Utilize the neighborhood information of the sampling points as the vote data. We draw the
ideas from [14] to encode the intensity information. In the 5-layer simplified model; there
are 15 point pairs which are symmetry to the key point. As it shown in Figure 2,
P(P, P)
cs
=
is a symmetry pair, regard
P as the center, and connect
P and
x
to get
12
line , connect
P and
x
to get
line , connect
P and
x
to get
line , connect
P and
1
2
3
1
x
to get
line ,then define a point set
ˁ =
{(x ); x
∈
circle
x
∈
line }, j
=
1, 2, 3, 4
,
4
j
j
i
m
j
j
ˁ =
2
{(x ); x
'
'
∈
circle
line },
j
x
'
∈
similarly
Phas
, define,
vote_data =
j
=
1,2,3,4
n
j
j
j
j
Search WWH ::
Custom Search