Hyperparameter Sensitivity Analysis and Optimization for General Mammal Pose Estimation with KITPose

General mammals pose estimation remains challenging due to high inter-species variability and limited labeled data. While the KITPose framework achieves state-of-the-art performance on benchmarks like AP-10K and AnimalPose, its hyperparameter configuration was empirically fixed and not systematically studied. This paper presents the first comprehensive hyperparameter sensitivity analysis and optimization for KITPose. A two-stage experimental design is adopted: Stage 1 conducts single-variable controlled experiments across six key hyperparameterstransformer depth, attention head count, learning rate, optimizer, scheduler, and augmentation scaleon both AP-10K and AnimalPose. Results reveal that depth, head count, and learning rate are the most influential factors. Stage 2 applies Bayesian optimization via Optuna to jointly optimize these three parameters within empirically grounded ranges. Automated tuning confirms that shallow architectures (depth 2) with multi-head attention ( 2 heads) and moderate learning rates (1e-44e-4) consistently outperform deeper or under-parameterized variants. Notably, increasing depth to 4 leads to sharp performance degradation (>10% mAP drop), validating the original KITPose design.