Learned Monocular Depth Priors in Visual-Inertial Initialization

AbstractVisual-inertial odometry (VIO) is the pose estimation backbone for most AR/VR and autonomous robotic systems today, in both academia industry. However, these are highly sensitive to initialization of key parameters such as sensor biases, gravity direction, metric scale. In practical scenarios where high-parallax or variable acceleration assumptions rarely met (e.g. hovering aerial robot, smartphone AR user not gesticulating with phone), classical visual-inertial formulations often become ill-conditioned and/or fail meaningfully converge. this paper we target specifically low-excitation critical in-the-wild usage. We propose circumvent limitations structure-from-motion (SfM) by incorporating a new learning-based measurement higher-level input. leverage learned monocular depth images (mono-depth) constrain relative features, upgrade mono-depths scale jointly optimizing their scales shifts. Our experiments show significant improvement problem conditioning compared formulation initialization, demonstrate accuracy robustness improvements state-of-the-art on public benchmarks, particularly under scenarios. further extend implementation within an existing system illustrate impact our improved method resulting tracking trajectories.KeywordsVisual-inertial initializationMonocular depthVisual-inertial structure from motion