Offloading Visual-Inertial Odometry for Low Power Extended Reality

Qinjun Jiang, Muhammad Huzaifa, William Sentosa, Jeffrey Zhang, Steven Gao, Yihan Pang, Brighten Godfrey, Sarita V. Adve

Abstract: Power consumption is a constraint in designing all-day wear extended reality (XR) devices. To reduce power consumption, we present a novel end-to-end XR system with visual-inertial odometry (VIO) offloaded to a remote server. Even with long network latency, the system produces an end-to-end XR experience comparable to a system without offloading while providing significant power savings. This latency and bandwidth-resilient nature of head tracking is a new and surprising result.