Video Scene Analysis: A Machine Learning Perspective - Video Segmentation and Its Applications - page 109

Two video groups consisting of four matches are utilized, which are defined as
(a) Dataset A: using two NBA games for training, and using another two Olympic
Games for testing; (b) Database B: using one NBA game for training, and using
another NBA game for testing. Frame-based views from the PLSA model and the
SVM model are applied to Dataset A and B. Therefore, four combinations of view
labels and datasets are defined as PLSA B.

Each video clip used in both training and testing is automatically decimated and
consists of 500 uniformly sampled frames. We use a window size N

+ A, PLSA + B, SVM + A, and SVM +

= 20, with a

window N sliding every 10 frames.
The number of approximated events detected after the first stage is given in
Tab le 4.2 . The precision and recall of the coarse stage basketball score detection are
92.03% and 86.19% respectively. In the second stage, the proposed HCRF-based
model and state of the art HMM and CRF models are evaluated and compared. The
advantage of HCRF over HMM is its relaxation on the Markov property that the
current state St can be inferred from both current observation as well as surrounding
observations. As shown in Table 4.3 , the HCRF has better performance than the CRF
for the same ω values, while both models outperform the HMM baseline. When us-

ing different ω values for both CRF and HCRF, ω = 1 provides better results than

ω = 0, in which neighboring information assists in a better decision making. How-

ever, when ω = 2 is used for HCRF, the performance has been dropped for all cases

comparing with 1. This can be viewed as an over-fitting issue, in which adding

more surrounding information limits the structured prediction ability. In summary,
the proposed HCRF based model with parameter

ω =

1 outperforms both CRF and
HMM models. The best results are obtained at 93.08% and 92.31% by taking SVM
and PLSA based input labels, respectively.

ω =

Tabl e 4. 2 Precision and recall results of basketball score events detection at the first
(coarse) stage
Correctly detected score Detected score Correct total score Precision Recall

(True positive) (Obtained result) (Obtained result) (%) (%)

231 251 268 92.03 86.19

Tabl e 4. 3 Performance comparison on score event detection in basketball.
Dataset A: NBA matches as training, Olympic matches as testing. Dataset B:
NBA matches for both training and testing
Accuracy
Dataset A (NBA/Olympics) Dataset B (NBA/NBA)

SVM + A(%) PLSA + A(%) SVM + B(%) PLAS + B(%)

HMM ω = 0 78.28 75.29 87.50 85.94

CRF ω = 0 78.16 74.57 87.43 86.52

CRF ω = 1 79.52 76.82 88.52 87.89

HCRF ω = 0 80.93 75.53 90.00 90.77

HCRF ω = 1 83.26 80.24 93.08 92.31

HCRF ω = 2 82.09 77.88 91.46 91.77