DM-VIO: Delayed Marginalization Visual-Inertial Odometry

We present DM-VIO, a monocular visual-inertial odometry system based on two novel techniques called delayed marginalization and pose graph bundle adjustment. DM-VIO performs photometric adjustment with dynamic weight for visual residuals. adopt marginalization, which is popular strategy to keep the update time constrained, but it cannot easily be reversed, linearization points of connected variables have fixed. To overcome this we propose marginalization: The idea maintain second factor graph, where delayed. This allows us later readvance yielding an updated prior new consistent points. In addition, enables inject IMU information into already marginalized states. foundation proposed adjustment, use initialization. contrast works initialization, able capture full uncertainty, improving scale estimation. order cope initially unobservable scale, continue optimize gravity direction in main after initialization complete. evaluate our EuRoC, TUM-VI, 4Seasons datasets, comprise flying drone, large-scale handheld, automotive scenarios. Thanks exceeds state art odometry, even outperforming stereo-inertial methods while using only single camera IMU. code will published at http://vision.in.tum.de/dm-vio