FishProx: Proximate conspecific interaction of Atlantic salmon (Salmo salar) for behavioural analysis through instance segmentation masks

Abstract

In modern fish production and experimental facilities, camera installations are increasingly common for monitoring animal welfare. This opens up a new alley of autonomous tracking systems levering streamed data. However, the substantial amount of data necessitates condensation for human operators. In this paper, we present the out-of-the-box pipeline called FishProx that starts with the imagery of fish in tanks and ends in the calculation of welfare impacting factors through anomaly scores. The utilisation of the Segment Anything Model (SAM) is the basis for the identification of feeding behaviour and also (ab)normal behaviour through analysing snapshots of the fish tanks to infer metrics such as fish cohesion and cluster alignment, alongside the detection count. Our method employs automatic and robust algorithms for camera systems, making additional labelled data unnecessary, thereby saving energy and costs. Furthermore, automatic fish observation coupled with integrated analytical capabilities can lead to a sensible autonomous decision-making process. Lastly, by providing objective metrics, our approach might guide the enabling of automatic quantification of fish states, making comparative studies possible.