Markerless Individual Re-Identification of Visually Homogeneous Zebrafish for Aquaculture Monitoring Using Hybrid Spatial–Visual Tracking and Angular-Margin Deep Embeddings

Persistent markerless identification of visually similar fish is difficult when occlusion interrupts tracking. This study evaluates a hybrid spatial–visual pipeline combining YOLO26n detection, ByteTrack association and a Swin-T re-identification (Re-ID) model. Experiments were carried out using two Aalborg University (AAU) zebrafish recordings containing six known individuals. The detector and embedding model were retrained on consecutive temporal blocks separated by ten omitted frames; 426 frames were held out from retraining and checkpoint selection. A 400 px vertical mask simulated occlusion, and predictions were evaluated against ground-truth AAU Object IDs. The added association stage restored the prior identity in three of eighteen eligible episodes after full emergence (16.7%), compared with none for ByteTrack. All three successes bridged changed ByteTrack IDs, with successes in both recordings; one assignment was incorrect and fourteen episodes remained unresolved. Under a separate partial-reappearance definition, Re-ID recovered six of twenty-four episodes (25.0%), versus none for ByteTrack, with one incorrect identity, one different identifier for the same fish and sixteen unresolved episodes. These verified recoveries demonstrate that the added association stage can restore identities after tracker interruption. Limited assignment coverage and testing on previously seen individuals constrain conclusions about reliability and transfer to independent recordings.