Motion Database Retrieval with Application to Gesture Recognition in a Virtual Reality Dance Training System - Multimedia Database Retrieval: Technology and Applications - page 326

Table 11.8 Averaged recognition rate (%) over 40 classes, obtained
by the template matching methods (PO, PSC, PT, and PTSC),
compared to the HMM
Method Similarity function Average recognition accuracy (%)

PO L1 41.8

L2 41.0

HI 41.8

PSC L1 19.3

L2 23.5

HI 19.0

PT L1 42.4

L2 40.8

HI 42.4

PTSC L1 25.7

L2 28.4

HI 26.3

HMM 77.3

The SSOM was trained by the C1 configuration, using the following procedural
parameters: icosahedron level = 1, map nodes = 42, neighborhood = 3, and epochs

= 200. In order to assess the performance of the gesture template definition (i.e.,

PO, PSC, PT, and PTSC discussed in Sect. 11.5 ) and matching criteria, the system
was trained using 50 % of all samples. From the full set of gesture instances,
50 % of each class were randomly selected and used to form gesture templates,
while all 100 % were compared with these templates. The results are displayed in
Table 11.8 .
From the result, we observe that recognition performance was quite low for all
scenarios. This is partly due to the complexity of the gesture movements. The sparse
codes of postures appear to give lowest performance, since they considered only the
existence of the postures for constructing gesture templates (each of which was a
binary sparse code vector of 42 dimensions resulting from a 42-node SSOM). The
PT gave a better performance than the PO method. The approach based on PT takes
into account temporal information about the gesture (which is lacking in the PO
vector), so it would seem reasonable that its accuracy in classification should be
better.
The proposed system outlined in Fig. 11.8 was implemented. The SSOM was
constructed in the same way as the previous experiment, using the C1 configuration.
However, the SSOM for each gesture class was trained separately, resulting in
multiple codebooks. For each class, the number of codewords in the sub-codebook
was 42. For HMMs, the number of states N S was set to 42, and the number of
symbols N O was 50. We let the number of states correspond roughly to the number
of postures within the gesture. We restrict each gesture model to having the same
number of states; this implies that the models will work best when they represent
gestures with the same number of postures.