3.2 Object
tracking
When comparing the performance of the OT architectures, we found that
the three architectures detected and subsequently tracked more than 120
of the 169 individual fish that swam through the rockwall passage (Table
2). Average precision values for all architectures were above 80% with
Seq-NMS being the most precise (93%) at detecting and tracking the
yellowfin bream. Recall among architectures were very similar at around
73%. The architecture with the highest overall success at detecting and
tracking bream was Seq-NMS (F1=84%) (Table 2).
TABLE 2 Confusion matrix for the three object tracking
architectures (MOSSE, Seq-NMS and SiamMask) are shown as counts of
individual fish, where the true positive means a bream was detected and
tracked correctly for ≥ 50% of the time when it appeared on a video
frame, otherwise, it was false negative. False positives were
misidentified objects (i.e. algae or other fish) that were detected and
tracked.