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

Area of Science:

• Behavioral ecology
• Machine learning applications
• Animal identification

Background:

• Accurate identification of unmarked animals is crucial for ecological and behavioral research.
• Representation learning models show high performance in single-session animal identification tasks.
• The stability and reliability of these models across multiple sessions are not well understood.

Purpose of the Study:

• To evaluate the stability and accuracy of representation learning for animal identification across different sessions.
• To assess how visual changes between sessions affect identification performance.
• To determine if generalization training improves cross-session identification consistency.

Main Methods:

• Utilized synthetic and real-world datasets for cross-session experiments.
• Assessed identity consistency, accuracy metrics, and the nature of learned representations.
• Compared performance under varying degrees of visual changes between sessions.

Main Results:

• Identification accuracy significantly degrades with minimal visual changes between sessions.
• While multiple sessions improve generalization, substantial performance degradation persists.
• Learned representations show instability across sessions, impacting reliable identity tracking.

Conclusions:

• Current representation learning methods, while accurate in single sessions, lack the stability required for reliable cross-session animal identification.
• Reported single-session accuracy is insufficient to guarantee consistent animal identities over time in ecological research.
• Further research is needed to develop robust methods for stable, long-term animal identification in dynamic environments.