PoseGuru: Landmarks for Explainable Pose Correction Using Exemplar-Guided Algorithmic Recourse

Abstract

Human pose correction is crucial in fields such as fitness, rehabilitation, and sports. Despite recent advancements, existing approaches often lack explainable, personalized, and fine-grained corrections. We present PoseGuru, an explainable optimization based approach, leveraging algorithmic recourse and counterfactuals to iteratively refine pose landmarks by minimizing classification loss, enforcing user-specific anatomical constraints, and precisely aligning with the target pose. Our method is simple, interpretable, and adaptable, enabling easy incorporation of application-specific constraints. For robust evaluation, we introduce two new datasets, YogaHPC and Pilates32+P, generated by biomechanically perturbing landmarks of correct poses. PoseGuru consistently outperformed existing methods on both datasets, as assessed using metrics such as MPIJAD and PCIK. Furthermore, a comprehensive user study involving Yoga and Pilates experts confirmed PoseGuru's effectiveness, highlighting its capacity to facilitate user-driven pose correction across diverse pose types. Overall, PoseGuru provides an explainable and personalized solution suitable for critical applications in fitness and rehabilitation.