Deep learning-based human pose estimation from a small number of body-worn inertial measurement units (IMUs) is crucial for practical AR/VR and healthcare applications. However, most existing datasets are collected under laboratory conditions with high-end hardware and carefully calibrated, perfectly placed sensors. As a result, they implicitly assume near-ideal sensor behavior and mounting, which does not reflect the noisy, misaligned, and drift-prone data typically produced by affordable commodity IMU devices in real-world use.
This project focuses on recording a new motion dataset that explicitly captures these real-world imperfections. The goals include:
- Collecting synchronized IMU and reference motion capture data across a diverse set of activities
- Running a user study to collect high-quality, systematically recorded IMU data from non-specialized IMU devices across a diverse group of participants
- Preprocessing and packaging the recordings for seamless integration with an existing benchmarking platform
