Paper: DB-VIO, A Dual-Branch Framework for Visual Inertial Odometry

rsasaki0109 · x · 2026-08-11

This research introduces DB-VIO, a dual-branch Visual Inertial Odometry (VIO) framework with enhanced visual-inertial representation.

Existing learning-based VIO methods typically rely on unified visual-inertial representations and a single temporal model for full-pose estimation, struggling to capture the heterogeneous dynamics of rotation and translation. DB-VIO addresses this by incorporating depth cues to improve monocular visual features and explicitly extracting underlying rotational kinematics, thereby enhancing rotation-related cues in IMU features and improving 6-DoF motion estimation accuracy for mobile robotic systems.

Original post →

More from Embodied

Embodied channel →