SuperAnimal pretrained pose estimation models for behavioral analysis

Area of Science:

• Computational biology
• Ethology
• Machine learning

Background:

• Accurate quantification of animal behavior is crucial for neuroscience, veterinary medicine, and conservation.
• Current pose estimation methods rely on extensive manual labeling and domain expertise, limiting scalability.

Purpose of the Study:

• To develop a unified foundation model for animal pose estimation applicable to over 45 species.
• To reduce the need for manual labeling and improve data efficiency in behavioral analysis.

Main Methods:

• Developed SuperAnimal, a foundation model for multi-species animal pose estimation.
• Utilized unsupervised video adaptation for performance enhancement and jitter reduction.
• Demonstrated fine-tuning capabilities on diverse labeled datasets.

Main Results:

• Achieved excellent performance across six pose estimation benchmarks.
• Showcased 10-100x greater data efficiency compared to prior transfer learning methods.
• Validated utility in behavioral classification and kinematic analysis.

Conclusions:

• SuperAnimal provides a data-efficient, unified solution for animal pose estimation.
• The method significantly lowers the barrier to entry for behavioral analysis across numerous species.
• Enables advanced applications in behavioral classification and kinematics without extensive manual annotation.