FIGURE 2 Interaction between the object detection model and
tracking architectures. The object detection model activates all three
tracking architectures. For MOSSE and SiamMask the tracker continues for
4 frames after the initial detection. For Seq-NMS, the movement was
determined by calculating the vector direction between two detections.
For all architectures a check was made to determine if the tracker
continued, stopped, or a new tracker started. For MOSSE and SiamMask the
check was made after 4 tracking frames from the first detection. For
SeqNMS the check was made for all frames after the first detection. The
interaction between detections and tracker occurred through the whole
length of a video where the object detection model detected a yellowfin
bream and was carried for all frames, videos and cameras. All trackers
provided a direction of movement for each frame where the interaction
between the detection and tracking occurred successfully.
3.
Results
3.1 Object
detection
When using the Mask R-CNN framework for detecting yellowfin bream we
obtained an 81% mAP50 value and an F1 score of 91% (Table 1). The OD
model missed 21 fish (false negatives) and misidentified 8 objects (i.e.
algae or other fish) as bream (false positives) out of the 169 fish
(ground-truth) that were observed.
TABLE 1 Object detection map50 and the evaluation results of
the Mask R-CNN yellowfin bream model. The confusion matrix is shown as
counts of individual fish, where the true positives were the correct
detection of yellowfin bream. Yellowfin bream not detected were false
negatives and misidentified objects were false positives.