Towards Representation Learning for Phenotyping beyond Animal Pose Estimation

Understanding and quantifying behavior is crucial for phenotyping in biological and medical research. While pose estimation methods like DeepLabCut (DLC) provide structured representations of animal movement, they struggle with occlusions and rapid motion. In this study, we integrate TAPIR with DLC to enhance robust pose tracking, improving continuity and reducing missing key points. Furthermore, we apply CEBRA to refined pose sequences to learn behavioral representations, facilitating computational phenotyping. Experimental results show that our method significantly improves tracking performance, providing a foundation for structured and interpretable representation learning of biological dynamics.