Information Technology Reference
In-Depth Information
In the following figure, the LBP-TOP method is visually explained on real
data.
Fig. 9 LBP-TOP methodology
The original LBP-TOP and the Extended LBP-TOP are computed on the
original cuboid or on the gradient cuboid. In Table 1 the accuracy results for
different values of
R
and
P
are shown. The notation of parameters is as illus-
trated in Equation 5. As it is possible to notice, better classification accuracy
has been obtained with the parameter
P
greater than 6 and radius
R
equal to
2. The performance is generally slightly decreasing as the radius
R
is getting
bigger, while it is increasing as the number of neighbors
P
is increased. This
could be explained as more neighbors permit to take more information into
account, but further distance from the central point means loosing important
local information. However, the drawback is a higher computational cost and
a higher dimensionality of the feature vector.
LBP − TOP
8
,
8
,
8
,
2
,
2
,
2
produces
a
768
vector
length,
while
LBP − TOP
10
,
10
,
10
,
2
,
2
,
2
has a descriptor dimension of
3072
. The final de-
scriptor of
LBP − TOP
12
,
12
,
12
,
2
,
2
,
2
will be
12288
vector lengths.
Tabl e 1 Accuracy achieved by the SVM classifier for different parameter values of
LBP-TOP
P
4 6 8 10
2 71.81 % 85.65 % 86.25 % 86.32 %
3 84.54 % 85.18 % 85.12 % 86.69 %
4 81.34 % 85.12 % 85.46 % 83.82 %
We have noticed that the use of uniform LBP operator decreases the per-
formance results compared with the original operator, since less information
is kept into account (see Table 2). Multiresolution LBP operator has also been
tested, but the gain in performances is not considerable with the increase of
the descriptor length and computational cost.
Search WWH ::
Custom Search