Database Reference
In-Depth Information
Table 11.3
Gesture database
# Instances per gesture class
Total
Dataset
Teacher
Student
instances
Gesture Set I:
G
12
,
G
13
,
G
14
,
G
15
,
G
16
,
G
23
,
G
24
,
G
25
,
G
26
,
G
34
,
G
35
,
G
36
,
G
45
,
G
46
,
G
56
10
10
300
Gesture Set II:
G
21
,
G
31
,
G
32
,
G
41
,
G
42
,
G
43
,
G
51
,
G
52
,
G
53
,
G
54
,
G
61
,
G
62
,
G
63
,
G
64
,
G
65
10
10
300
Table 11.4
Gesture
recognition results averaged
over the 15 gestures defined
in the upper triangle in
Table
11.2
Average recognition accuracy
Testing data
L1
L2
HI
Teacher
PO
96.7
98.0
96.7
PSC
79.3
84.0
79.3
PT
98.7
97.3
98.7
PTSC
87.3
92.7
87.3
Student
PO
94.0
100
94.0
PSC
77.3
85.3
77.3
PT
94.7
99.3
94.7
PTSC
86.0
92.0
86.0
This system employed the SSOM configuration C2a, and was trained according
to the joint position feature. In the C2a configuration, the map has the following
setting: The icosahedron level is 2, the number of map nodes is 162, the number of
neighborhoods is 4, and the number of epochs is 100.
Table
11.4
shows the performance of the proposed system for recognition of
ballet dance performed by two persons, Teacher and Student. Here, the template
matching was performed by three similarity metrics: L1 norm, L2 norm, and
histogram intersection (HI). The system can attain more than 98 % recognition
rate averaged over 15 classes for recognition of the Teacher dataset by using the
PT template and HI for similarity matching. The PO template also gave similar
recognition performance to the PT template. Moreover, the system can recognize
dance from the Student dataset with 100 % accuracy by using the PO template and
L2 norm for similarity matching.
Next, the two sets of gestures, Set I and Set II described in Table
11.3
were
used for the experiment. This database contains 30 gestures, where each gesture
G
ij
has its corresponding reversal
G
ji
.Gesture
G
12
is described by the movement
from the 1st position to the 2nd position, whereas
G
21
represents the movement
from the 2nd position to the 1st position. In this case, the POs of
G
12
and
G
21
may
be similar, and thus, they may be incapable for discriminating the two gestures for
recognition. The PTs, on the other hand, may preserve the direction of the movement
within the gestures, and they can be employed for discrimination of the reversals.
This is confirmed by the results shown in Table
11.5
. It can be observed from the
result that the gesture template obtained by PT outperforms other indexing methods
discussed. The recognition rate averaged over 30 gesture classes of the Teacher
Search WWH ::
Custom Search