where qt
min produces a single large cluster enclosing all feature
vectors [16] and qt = qt
max leads to a number of clusters equal
to the number of feature vectors. Figure 2 shows an example
where the clustering is performed with qt = qt
min and with
larger values of qt, which increase the number of clusters.
The best value of qt can be determined by iteratively
analyzing the clustering performance. In particular, because
feature vectors are unlabeled, we design a measure to perform
an unsupervised self-assessment of the clustering performance
namely Dissimilarity Score St that quantifies the homogeneity
of the feature vectors belonging to each specific cluster.