Noisy Labels Make Sense: Data-Driven Smartphone Inertial Tracking without Tedious Annotations

Empowered by deep learning, data-driven smartphone inertial tracking approaches have attracted much attention due to high accuracy and robustness. However, training deep models requires significant amounts of IMU data with high-precision labels which incurs high annotation costs. In this work, we propose a practical data-driven IMU tracking solution without tedious annotations in a systematic way. We firstly design a simple but efficient method to annotate IMU data with phone labels based on the location service provided by smartphones. However, the phone labels are too noisy to train an accurate deep model. In-depth experimental studies reveal the noises are dependent along an IMU sequence; the noisy phone labels have knowledge useful for learning and make sense. The unique characteristics of the phone labels make existing learning-with-noisy-label models (LNLs) fail to work. We then propose EasyTrack, a noise-resistant data-driven IMU tracking system, which adopts a dual-model framework to enable LNLs. To handle the noisy labels, we design a series of effectively noise-resistant techniques. Extensive experiment results demonstrate without tedious annotations, EasyTrack achieves high accuracy by learning with noisy phone labels, outperforming existing LNLs and data-driven IMU tracking approaches.