Introduction
Automatic monitoring of farmed fish can lead to improvements on feeding management and a better understanding of stress events. We therefore present an algorithmic pipeline to leverage the SAM model to cut out fish segmentations without the need for costly fish annotations. We also realized the downstream tasks of calculating orientation angles and identifying fish that do not swim in similar orientations as their close conspecifics.
Materials and methods
A top camera with a video resolution of 1080p @ 30fps was installed above each octagon-like tank (6 tanks, volume of 3.3m3 each). A total of 1000 Atlantic salmon with a weight of 40 grams ± 10 grams were placed in each tank. Using segmentation masks from SAM1 we calculated the orientation of detected fish in video frames. Then we determine if those fish are swimming coherently with its peers by filtering out fish in each k-mean cluster (which used the spatial coordinates of fish) that fell out of the interval with the median absolute distance around the circular median angle of the cluster, respectively. Aggregating the results gave rise to abnormalities scores.
Results and Discussion
We were able to show that both feeding and panic avoidance could be measured with these three metrics. We also showed that panic avoidance has a stronger spike/fall, meaning the intensity of reaction and the increase/decrease in our metrics was proportional. Further work is needed to differentiate between different behavioral patterns that could occlude results. Additionally, the detection count could be increased in the future to reflect the tank behavior more reliably. Overall, the presented methodology tries to push towards easily accessible and interpretable readings for facility personnel.
Acknowledgments
This work was funded by the Norwegian Research Council, Project number 320717 and 323300.