Scalable Video Genre Classification and Event Detection - Multimedia Database Retrieval: Technology and Applications - page 266

Table 9.4 ʴ dev and computation time in codebook

generation using bottom-up (BU) and single K-means (SK) structures
Codebook size

SSE deviation percentage

cb BU = 800 cb SK = 800 cb BU = 1 , 600 cb SK = 1 , 600

Computation 4h 350 h 9h 648 h

ʴ 1.4 % 3.7 %

dev

350 h, while the bottom-up clustering only takes 4 h. When the codebook size is
doubled to 1,600, the computation for single K-means and bottom-up clustering
are 648 and 9 h, respectively. With a truly distributed processing environment
using multiple computers, bottom-up processing time will be further reduced. This
comparison of computational complexity demonstrates that our generic framework
using robust bottom-up clustering for codebook generation can replace the single
K-means in dealing with large-scale and diverse datasets.
For the accuracy performance using k-NN and various dissimilarities, Table 9.5
shows the average genre categorization results for 23 different sports. The proposed
bottom-up codebook generation manifests a better and more robust performance
than single K-means codebook generation in both EMD and KL-div measurements.
By comparing the row-wise's dissimilarities, the bottom-up structure is more
consistent with codebook sizes of 800 and 1,600. On the contrary, the single
K-means codebook generation is unstable for both histogram and mLDA-based
distributions. For instance, the performance at a codebook size of 800 using EMD
has about a 7 % increment from ED dissimilarity (75 31 %), while the

counterpart at a codebook size of 1,600 using EMD has dropped 1

. 33 % vs. 68 .

. 1 % from ED

dissimilarity (64 39 %). One reason is that the single K-means clustering

on over three million input SIFT points hardly reaches the optimal value. As a
summary, KL-div performs the best among three dissimilarity measures. Using
the bottom-up structure, results of the codebook size 1,600 outperform the cases
with size 800 in all measurements with consistency. Oppositely, single K-means
clustering results are not consistent.
Another merit of the bottom-up structure is its preservation of individual
genre characteristics from the 1st-level K-means. On the contrary, single K-means
codebook generation covers all the data; thus, a weakly distinguishable genre is
easily overruled by a strong one. This reasoning explains why with the increase of
codebook size from 800 to 1,600, the bottom-up process has about a 4 % improve-
ment for KL-div, while the single K-means process has only a 2 % increment for
KL-div.
The individual sport genre classification result is illustrated in Fig. 9.7 .On
average, a codebook size of 1,600 gives an average of 3

. 28 % vs. 65 .

6 % higher than the
codebook size of 800, which corresponds with the empirical studies from other
research groups [ 258 , 261 ].
To evaluate the generic and extensive properties of our proposed method,
experimental results on the 23-sports dataset are compared with results in Li et al.'s
work [ 265 ], where a top-down process was adopted using single K-means as its

.