IMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks
IMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks
IMU-Bench is a benchmark of ten tasks from eight datasets that evaluates the multi-task performance and cross-task transfer of wrist-worn inertial measurement units (IMUs). The study compares six self-supervised methods, finding that the masked autoencoder InrtlWrist achieves the best or joint-best performance on nine of ten tasks, with fine-tuning further improving results on key tasks. The research underscores the trade-offs between task-specific supervision and cross-task generalization.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source:arXiv Robotics — research abstracts · arxiv.org