sample frame of these results. From Fig. 6 we can make two
observations.
The first observation is that SVMC outperforms the other
methods (i.e. highest NMI and lowest St overall) followed
by IDPMM and CF. SVMC can effectively cluster features
and maintain the best results over time. The transformation
to a higher dimensional space allows SVMC to capture the
intrinsic similarities among feature vectors of the same flow
type although they appear interwoven to an observer (Fig. 7h).
IDPMM is the second-best method as it can accurately cluster
the circular patterns, but it is unable to separate the bidirectional
flow on the top-left (Fig. 7e) because the algorithm
cannot temporally propagate the similarity among features.
Also, because the similarity among features is low, several
feature vectors are deemed to be outliers. CF achieves the
third-best results as it can accurately cluster flows with evident
opposite directions (top-left Fig. 7g). Recall that CF produces
clusters based on motion vector similarities in the input