Beyond accuracy: Cross-session stability in representation-based animal identification

Identification of unmarked animals is central to behavioral and ecological research. Representation learning currently achieves near-perfect performance in single-session experiments. However, stability across sessions remains poorly characterized. We evaluate identity consistency, accuracy, and learned representation using synthetic and real cross-session experiments. The results show that identification degrades even under minimal visual changes. Although exposure to multiple sessions improves generalization, substantial degradation remains. This indicates that the reported accuracy alone is insufficient to guarantee stable identities across sessions.