Discussion
Where the pipeline stands
AquaMind is still far from production-ready. The tracker holds identity well when a fish reappears from behind a plant, but crossings remain its weak point, measured against a human-reviewed ground truth in Stage 4. The chasing classifier of Stage 7 is the working behaviour model. The feeding-strike classifier of Stage 8 is a benchmarked negative result: at 60 frames per second the cues that define a strike are not in the data. Every model so far was also trained and validated within the same few videos, so none of them has yet been tested on a video it has never seen.
What does carry over is the method. Its steps, from versioned datasets to registered models, can be reused in any computer-vision project, whatever the animal or the question.
What comes next
In the longer run, the same approach could cover more species and the experimental setups common in behavioural ecology and pharmacology, perhaps one day in labs that still score behaviour by eye. Other aquatic model species, such as tadpoles, are filmed in similar tanks and scored in similar ways. What would carry over is the method, not the models: each new species would need its own labelled frames, a retrained detector and, for identity and behaviour, retuned settings. Before that, the pipeline needs a held-out test video for every model, and a simple deployment so that it can be run by someone other than its author.
Was it worth it
Setting up a pipeline like this is a real investment: labelling, training and tuning take months. It pays off through reuse. Once built for one species and one setup, every new experiment with them is scored at almost no extra cost, the same way every time, by an observer that never tires. For a lab that runs the same assay again and again, that is worth it; for a one-off study, scoring by eye may still be cheaper.