Information Technology Reference
In-Depth Information
Tabl e 2 Accuracy achieved by the SVM classifier for the Original and Uniform
LBP-TOP
Method
Accuracy
(SVM)
Descriptor
length
Computational
time (s)
LBP-TOP
8
,
8
,
8
,
2
,
2
,
2
86.25 %
768
0.0139
Uniform LBP-TOP
8
,
8
,
8
,
2
,
2
,
2
81.78 %
177
0.0243
Due to this considerations, we finally choose to use LBP-TOP
8
,
8
,
8
,
2
,
2
,
2
for the following experiments as it is computationally more ecient and the
accuracy is among the highest.
The time calculated in the following tables is measured on a computer
equipped with a 3 Ghz Pentium 4 CPU and 3 Gb RAM. As dimensionality
reduction technique, we used Principal Component Analysis (PCA) and set
the final dimension to
.
In Table 3, the Extended LBP-TOP is evaluated and different number of
slices is taken into account. As we can see, the Extended LBP-TOP descriptor
performs better than the original one, since more information is taken into
consideration at different times in XY planes and at different locations in the
XT and YT planes. Although best result is obtained with 6 slices on each
axis, the computational time is almost double than the Extended LBP-TOP
version with 3 slices; because of this issue, in the following we are computing
the Extended version on only 3 slices for each axis.
In Figure 10 the three slices on XY plane from the cuboid of Figure 8 are
shown.
100
Fig. 10 Extended LBP-TOP: 3 slices in XY plane
In the following tests, the results are showed for different codebook's size,
in order to find the best dimension of the codebook, and therefore the feature
lenght of the signature of each video, as explained in previous section.
As dimensionality reduction techniques, we applied PCA and OLPP [5, 17],
reducing the dimensions to 100. Dealing with a shorter descriptor vector is
Search WWH ::
Custom Search