Scalable Video Genre Classification and Event Detection - Multimedia Database Retrieval: Technology and Applications - page 268

Table 9.6 Genre categorization accuracy between various video clips with
uniform sampling-based and key-frame/shot-based methods
Three minutes clip Ten seconds clip

Uniform sampling Key-frame/shot Uniform sampling Key-frame/shot

83.83 % 79.41 % 71.90 % 63.10 %

100
90

80
70
60

50
40

30

20
10

0

soccer basketball volleyball tennis table tennis snooker F1 motorcycle swimming diving boxing judo figure sakting speed skating

SVM(14 sports) PLSA(14 sports) SVM (small scale) PLSA (small scale) SVM(small scale w/ soccer codebook) PLSA(small scale w/ soccer codebook)

Fig. 9.8 View type classification using supervised SVM and unsupervised PLSA. First two
columns are with codebook size 800 for 14 sports

a coarse and preliminary execution to quickly prune off the large portion of the
irrelevant data.
From a different perspective, generic properties of the proposed method are
assessed using various video clip lengths and frame sampling methods. As detailed
in Table 9.6 , better performance is acquired using longer lengths of video clips,
while a generic and automatic uniform sampling method outperforms the key-
frame sampling. It is because the proposed method is based on local key-point
descriptors. Therefore, a longer video clip with denser sampling frames provides
more key-points and consequently builds a better distribution than a shorter clip
with less sampled key-frames/shots. Such experimentation demonstrates the merit
of our proposed generic method towards a truly large-scale dataset.

9.4.2 Middle-Level View Classification Using Supervised

SVM and Unsupervised PLSA

Experiments in this section focus on middle-level view classification by utilizing
extracted low-level histogram-based representations. A subset of 14 sports of all
23 sports was used as test data. Figure 9.8 compares both supervised SVM and
unsupervised PLSA results as the 1st and 2nd columns, respectively. On average,
supervised SVM has a classification accuracy of 82

. 86 %, and unsupervised PLSA