IMUPoser: From Recurrence to Attention: Optimizing IMUPoser with Transformers

With the widespread adoption of mobile devices, inertial measurement units (IMUs) have emerged as an appealing solution for human pose estimation. Unlike traditional motion capture systems, IMUs, found in common consumer devices including smartphones, smartwatches, and earbuds, enable cost-effective and on-the-go pose tracking. They work across numerous application domains, including fitness monitoring, mobile human-computer interaction, context-aware virtual assistance, and rehabilitation. Nevertheless, the noise in low-cost IMU data, the variability in sensor placement, and the limitations of current deep learning models all present significant challenges to reliable pose estimation. This study addresses these issues by investigating and proposing a series of model optimizations, as well as augmenting robustness and generalization in real-world deployment scenarios.