Information Technology Reference
In-Depth Information
(a) Original images
(b) Gradient images
(c) Gabor images
Fig. 15 Different enhancement of XY, XT and YT images before the computation
of LBP operator
Tabl e 4 Computational time and accuracy for different (CS)LBP-TOP implemen-
tations, k=1000 visual words
Descriptor
length
Computational
time (s)
Accuracy
(SVM)
Method
LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
768
0.0139
86.25 %
CSLBP-TOP 10 , 10 , 10 , 2 , 2 , 2
96
0.0115
85.00 %
Extended LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
2304
0.0314
88.19 %
Extended LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
+
100
0.0319
87.87 %
PCA
Gradient LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
768
0.0788
90.07 %
Extended Gradient
LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
2304
0.0992
90.72 %
Extended Gradient
LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
100
0.1004
91.25 %
+PCA
Extended Gradient
CSLBP-TOP 10 , 10 , 10 , 2 , 2 , 2
288
0.0926
91.09 %
Gabor LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
768
1.037
88.73 %
Extended Gabor
LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
2304
1.320
92.22 %
Extended Gabor
LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
100
1.331
91.71 %
+PCA
As a comparison, we evaluate Laptev's method [11] with the same frame-
work as illustrated in Section 2 with a codebook size of k = 1000
visual words.
We use Laptev's code publicly available on his website and recently being up-
dated with the latest settings used in [11]. The combination of Laptev's 3D
Harris corner detector and Laptev's HOG-HOF descriptor make us reach an
accuracy of 89.88%. In the paper by Laptev [11], 91.8% of accuracy is ob-
tained on KTH database, but different channel combinations for HOG and
HOF are being used. Using the terminology of the article, our descriptors
are computing in a 1×1 grid. The time for the HOG-HOF descriptor in Ta-
ble 3 is referred to both detection and description parts, as the description
part cannot be computed regardless of the detection part in the provided
executable. Therefore, we expect the description part to be about half the
time shown in the table. The computational time for HOG-HOF is affected
 
Search WWH ::




Custom Search