Database Reference
In-Depth Information
Fig. 9.9
HCRF input shown
in Eq. (
9.7
), by sliding
window average result on
view types of decoded image
sequence
Fig. 9.10
Current state
influenced by surrounding
observed states
Table 9.7
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
Correct result
Obtained result
%
%
231
251
268
92.03
86.19
The number of approximated events detected after the first stage is given in
Table
9.7
. The precision and recall of the coarse-stage basketball score detection are
92
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
S
t
can be inferred from both current observations, as well
as surrounding observations, as illustrated in Fig.
9.10
. In the experiment, the
circumferential range number is selected at
.
03 % and 86
.
2. As shown in Table
9.8
,the
HCRF has better performance than the CRF for the same
ˉ
=
0
,
1
,
ˉ
values, while both
models outperform the HMM baseline. When using different
ˉ
values for both
CRF and HCRF,
0, in which neighboring
information assists in better decision-making. However, when
ˉ
=
1 provides better results than
ˉ
=
2isusedfor
HCRF, the performance has been dropped for all cases compared with
ˉ
=
1. This
performance degradation can be viewed as an overfitting issue, in which adding
more surrounding information limits the structured prediction ability. A similar
overfitting problem is also observed in gesture recognition research using HCRF
[
293
]. In summary, the proposed HCRF-based model with parameter
ˉ
=
ˉ
=
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.
On the other hand, by comparing the proposed PLSA with SVM benchmark,
performance discrepancy of the event detection has been shortened, despite the
input view classification (as shown in Fig.
9.8
) has PLSA (70
.
.
14 %) outperformed
by SVM (82
86 %. For Dataset A, the average difference shows
that SVM outperforms PLSA by 3
.
00 %) with 11
.
.
65 %, while in Dataset B, such a difference is
only 0
.
47 %. This tolerable difference demonstrates the robustness and resilience
Search WWH ::
Custom Search