Pose estimation analysis and fine-tuning on the REHAB24-6 rehabilitation dataset

Human motion analysis is a key enabler for remote healthcare applications, particularly in physical rehabilitation. In this context, mobile devices equipped with RGB cameras seem to be a promising technology for monitoring patients during home-based exercises and providing real-time feedback. This relies on pose estimation algorithms that extract spatio-temporal features of human motion from video data. While state-of-the-art models can estimate body pose from mobile video streams, their effectiveness in rehabilitation scenarios remains underexplored. To address this, we introduce the REHAB24-6 dataset, which includes untrimmed RGB videos, 2D and 3D skeletal ground truth annotations, and temporal segmentation for six common rehabilitation exercises. We also propose an evaluation protocol for assessing different aspects of quality of pose estimation methods, dealing with challenges that arise when different skeleton formats are compared. Additionally, we show how fine-tuning of existing models on our dataset leads to improved quality. Our experimental results compare several state-of-the-art approaches and highlight their key limitations – particularly in depth estimation – offering practical insights for selecting and improving pose estimation systems for rehabilitation monitoring.