Researchers automate detection of fish cleaning interactions

A preprint posted to bioRxiv, titled "Using supervised machine learning to quantify cleaning behaviour," reports a semi-automated pipeline that tracks and classifies cleaning interactions between the cleaner wrasse (Labroides dimidiatus) and the powder blue tang (Acanthurus leucosternon) in a controlled three-dimensional laboratory setting. The authors used DeepLabCut for simultaneous markerless pose estimation of both fish, then built a classifier on the resulting tracking data. The preprint reports the classifier detected cleaning interactions with about 90% accuracy and reduced the footage requiring manual annotation by roughly 75%, and the authors describe it as the first multi-species 3D DeepLabCut model for cleaner-client fish.

Per the preprint, DeepLabCut was trained on labeled frames to estimate poses for both species, after which features from the tracked keypoints fed a supervised classifier that labeled interaction windows. The authors report low tracking error for the pose estimator and, for the downstream classifier, roughly 90% detection accuracy, a false-positive rate near 15% of non-interactions, and identification of about 25% of footage as containing interactions. As a preprint, the manuscript has not yet completed peer review.

Industry-pattern observation: markerless pose-estimation pipelines like DeepLabCut routinely replace labor-intensive video annotation in behavioral ecology and commonly deliver large reductions in human labeling time. A recurring limitation is generalization, because models trained on controlled lab footage often degrade on cluttered field video with variable lighting, occlusion, and multiple animals. A reported false-positive rate near 15% also means automated pre-filtering accelerates work but still requires human review of flagged clips, which teams should factor into annotation resource planning.