Negative Results in IMU-based Skiing Style Recognition: Skill Level Distribution Shift

Human Activity Recognition (HAR) models often struggle with generalization, meaning they perform poorly when deployed in new, unseen environments or with different users than those in the training set. This poor generalization occurs because HAR models often experience variability in sensor data due to differences in individuals (e.g., age, gender, physical characteristics) and environmental factors (e.g., sensor placement, lighting conditions, background noise), which cause distribution shift. All mentioned forms of variability can appear in recreational alpine skiing. However, skiers' skill is a factor that has been less studied and highlighted by scholars. In this study, we explain why a model struggles to generalize across a mixed-skill dataset. We employed an autoencoder-based multi-task learning model, which, despite achieving state-of-the-art on standard HAR datasets and promising results on a skiing dataset, failed to generalize in a real-world alpine skiing setting. We identified and quantified a skill-related distribution shift as the cause; low-skilled and experienced skiers occupy distinct regions in latent feature space with Wasserstein-1 distances increasing from 5.39 to 41.85 for the most basic to the most advanced skiing technique.