Summary
Wrist-worn sensors that measure movement can be used for many tasks like recognizing activities, estimating body pose, and assessing clinical motion. The authors created IMU-Bench, a platform to test how well different methods work across ten tasks and eight datasets, including both real and simulated motions. They found that their approach, a masked autoencoder called InrtlWrist, generally performed best at predicting these tasks and transferring learned knowledge. However, training for one task could sometimes reduce performance on others, showing the challenges in using shared sensor data across multiple uses. Their work provides tools and comparisons to help developers build better wrist-IMU based applications.
What this means in practice
- •For wearable device developers: Improve wrist sensor algorithms by selecting pretrained models that transfer effectively across recognition, pose, and clinical assessment tasks.
- •For mobile health software teams: Develop multi-purpose motion analysis apps exploiting shared IMU data representations to enhance motor function tracking and clinical assessments.
Authors
Yiquan Li, Aidan Bradshaw, Chenfeng Gao, Karan Ahuja
Abstract
Wrist-worn inertial measurement units (IMUs) provide a common sensing modality for tasks ranging from activity and gesture recognition to pose estimation, inertial odometry and clinical motor assessment. Existing wearable benchmarks evaluate performance across multiple tasks, but leave open how supervision for one task affects transfer to others and whether these tasks benefit from a shared representation. We present IMU-Bench, a benchmark of ten tasks from eight datasets for studying IMU multi-task performance and cross-task transfer. We compare six self-supervised methods pretrained on unlabeled recorded and simulated wrist motion, evaluating classification, pose and velocity estimation, and IMU-language retrieval. We find that our masked autoencoder, InrtlWrist, achieves the best or joint-best frozen-probe performance on nine of ten tasks, and fine-tuning improves on training from scratch on nine of ten tasks, including full-body pose, odometry and retrieval. Across transfer and joint-training experiments, supervision that benefits one target can reduce performance on others, with supervised sources transferring within kinematic estimation while, under frozen probing, none approaches the pretrained encoder on language retrieval, and joint-training gains that come at a cost to the paired task. These findings distinguish broad downstream usefulness from complementary task supervision. Code, task cards and a leaderboard are publicly available.