Instant Visual Odometry Initialization for Mobile AR.

Area of Science:

• Computer Vision
• Robotics
• Augmented Reality

Background:

• Standard visual odometry (VO) and SLAM require motion parallax for initialization, causing delays in mobile AR.
• Delayed initialization hinders instant display of world-locked effects in AR applications.

Purpose of the Study:

• Develop a 6-DoF monocular visual odometry system with instant initialization.
• Eliminate the need for motion parallax during the initialization phase.
• Improve the accuracy and consistency of pose estimation for AR applications.

Main Methods:

• A novel pose estimator decoupling 5-DoF relative rotation/translation direction from 1-DoF translation magnitude.
• Leveraging user motion (rotation then translation) to naturally transition scale estimation.
• Implementing a new residual for the relative pose problem using Jacobians and Levenberg-Marquardt optimization on a 5-DoF manifold.

Main Results:

• The proposed pose estimator outperforms classical methods in low-parallax configurations.
• The relative pose estimator surpasses state-of-the-art approaches when initial guesses are available.
• Validation performed on public and synthetic datasets, with a new real-world dataset released.

Conclusions:

• The developed visual odometry enables instant, parallax-free initialization for mobile AR.
• The system effectively handles scale estimation by transitioning between rotation-only and translational motion.
• This solution can be integrated as a full odometry or pre-SLAM component for platforms like ARKit and ARCore.