Information Technology Reference
In-Depth Information
directions and combined into the gradient cuboid permits to have a better
performance for LBP-TOP in the description of actions. The gabor filtering
applied to the original cuboid helps in increasing the final accuracy for LBP-
TOP, although the performance is lower than using the gradient image. This
could be explained as the gradient images encode more relevant information
for describing the motion inside each video patch than the Gabor images.
Moreover, the gradient image better defines the borders of the movement,
while gabor image better highlights the area of motion, as shown in Figure
15. A further improvement in the performances can be achieved by applying
the Extended LBP-TOP on the gradient or gabor cuboids. The Extended
LBP-TOP applied on gabor cuboids is giving very close performance with
the Extended LBP-TOP applied on gradient cuboids.
In Figure 13, a comparison between LBP-TOP and CSLBP-TOP is shown,
keeping fixed the number of visual words k=1000. As we can notice, the
performance of CSLBP-TOP operator is close to that of LBP-TOP, as well
as Extended CSLBP-TOP is very similar with Extended LBP-TOP. A higher
number of neighbors is needed for CSLBP-TOP to reach better classification
accuracy; as the plots show, a number of neighbors equal to 10 or 12 permits
to reach performance similar, and even slightly better, to the original LBP-
TOP. However, given a fixed number of neighbors, CSLBP-TOP's descriptor
is 16 times shorter than that of LBP-TOP. For a number of neighbors equal to
8, the descriptor is 48 length, while LBP-TOP's descriptor is 768 dimensions
length.
Figure 14 highlights the best results for LBP-TOP and CSLBP-TOP. The
best results have been achieved with the original LBP-TOP implementation.
CSLBP operator applied to Gradient or Gabor images gives worse accuracy
results.
In general, CSLBP-TOP performs similar with the original LBP-TOP in
the field of human action recognition. If the number of neighbors are increased
(i.e.
P
=12
), Extended CSLBP-TOP is slightly outperforming the Extended
LBP-TOP (
P
=8
), as more spatial information is taken into account during
the computation of CS-LBP operator. The Extended Gradient CSLBP-TOP
version is performing best among the descriptors based on the CS-LBP oper-
ator, reaching almost the performance of Extended Gradient LBP-TOP using
1-NN classifier.
Best performances have been achieved by using the Extended Gradient
LBP-TOP. The classification accuracy has been of 92.69% and 92.57% if
PCA is applied using the 1-NN classifier with
χ
2
distance and setting the
codebook's size equal to 1250.
Table 4 shows the computational time for describing one small video patch
and classification accuracy. As can be seen, the time is increasing if a higher
number of slices is taken into account and if the gradient or gabor cuboid is
computed. CSLBP-TOP implementation is slightly faster, as less comparisons
have to be computed and the final histogram is shorter. The Extended Gabor
LBP-TOP requires more time among all LBP-TOP methods.
Search WWH ::
Custom Search