Scalable Video Genre Classification and Event Detection - Multimedia Database Retrieval: Technology and Applications - page 267

Table 9.5 Average
categorization results (%) of
23-sports data with codebook
size 800 and 1,600

Measurement ED EMD KL-div

cb BU = 800 61.54 75.80 78.59

cb SK = 800 68.31 75.33 73.49

cb BU = 1 , 600 65.68 78.94 82.16

600 65.39 64.28 75.75
BU: codebook generated using bottom-up
structure. SK: codebook generated using
single K-means structure

cb SK = 1 ,

codebook size 800
codebook size 1600

100

90
80
70
60
50
40
30

20
10
0

Fig. 9.7 Genre categorization for the 23-sports dataset with codebook sizes of 800 and 1,600

top layer general codebook. The best performance in two-layer and single-layer
structures are 83

2 %, respectively [ 265 ]. In their work, speeded up
robust features (SURF)-based method is adopted. Similar to SIFT, SURF is also
a scale and rotation-invariant interesting point feature extraction algorithm, which
focuses on the computational efficiency [ 303 ]. Although SURF and SIFT adopt
different key points detection techniques, these two descriptors are comparable in
characterizing local features of sampled frames from a video sequence. Therefore,
such a comparison is valid in genre categorization performances, regardless of the
feature extraction difference. Considering the increment of data in scale about 27 %
(145 h vs. 114

. 83 % and 81 .

2 h), and in variety about 64 % (23 genres vs. 14 genres), using
the bottom-up structure with a codebook size of 1,600 and KL-div measurement,
our experimentation provides comparable results of 82.16 %, with a degradation of
1.67 %.
Although the performance is maintained on average, we also observed that
the individual performance has been fluctuating. This fluctuation is mainly due to
the nature of the adopted k-NN classifier, where distance-based measurement can be
overruled by a strong representation in a large and sparse dataset. We acknowledge
that k-NN may not be the most robust algorithm towards the very large-scale dataset.
However, the k-NN is an efficient method in batch processing. It can be used as

.