Digital Signal Processing Reference
In-Depth Information
Fig. 11.10
Plots of flattened
metre vector
m
for “
Maid
Of Orleans
”(3/4metre,
top
)
and “
Hit the Road Jack
”(4/4
metre,
bottom
)
1.0
0.5
0.0
0
2
4
6
8
10
12
14
16
18
20
factor
1.0
0.5
0.0
0
2
4
6
8
10
12
14
16
18
20
factor
T
were introduced so far—they form the feature set
F
all
. Given the suitability of
linear SVMs with SMO learning in [
24
,
39
], these have been used for classification
in the work described here. To determine the most relevant features for metre and
ballroom dance style classification from the set
F
all
, a Sequential Floating Forward
Search (SFFS) [
39
] with the target classifier—the SVMs—was carried out once per
task. This led to the feature sub-set
F
metre
for metre classification:
T
ratio
, metre vector
m
elements 4, 6, 8, 16, and the Tatum vector
T
. Further, the ballroom dance style
classification feature sub-set
F
dance
found resembles: metre
M
,
T
ratio
,
T
slope
,
T
peakdist
,
metre vector
m
elements 4-6, 8, 11, 12, 14, 15, 19, and the Tatum vector
T
without
elements 21 and 29.
11.3.2.4 Recognition
Metre and ballroom dance style are classified by a data-learnt approach, namely
SVM. Given the continuous value nature of tempo, one may think of using SVR
for tempo assessment. This was tested on the BRD set, but observed as not able
to identify a few percent relative BPM deviation. Thus, a hybrid classification and
regression approach is considered: The tempo range is divided into few overlapping
tempo ranges. A natural choice in the context of ballroom dance style is to use these
styles as tempo classes as these are usually limited to a specific tempo range. Such a
regulation is officially provided by the International DanceSport Federation's tempo
regulation for competitions. In [
13
,
87
], this fact is used the other way around: Tempo
ranges are used there to assess the ballroom dance style.
