Figure 3
Cumulative density plot (‘precision curve’) of the tracking error. Precision is the fraction of all datasets for which tracking was better than a certain tracking error with respect to manual selection (OP1). The curves in (a) show the accumulated tracking errors of all 66 datasets for the automatic algorithms and the alternative manual selection (OP2). With a precision of 73.5%, KCF performs better than the RealTITracker (66.8%) and closest to the manual operator (OP2: 82.3%), while the other OpenCV trackers perform substantially worse. (b) shows only the KCF precision curves for all 66 datasets separately. The 5-mm threshold used for evaluation and comparison of tracking precision is indicated by the dashed line. Legend entries are sorted according these precision values.
