VSMP-IMU: Video-Grounded Semantic Motion Programs for Sensor-Aware Synthetic IMU Generation

Summary

Wearable human activity recognition is chronically limited by the scarcity of labeled sensor data, and the problem is worst exactly where it matters: low-resource settings, class-imbalanced long-tail activities, and generalization to unseen subjects. Synthetic IMU generation is the standard remedy, but existing approaches sit on a trade-off. Video-driven methods are visually grounded in how activities are really performed, yet inherit pose-estimation errors.