Unifying Geometric- and Learning-Based Visual–Inertial Odometry: A Comprehensive Evaluation and Comparative Study

Visual–Inertial Odometry (VIO) is a core component of state estimation in robotics and autonomous systems. Over the past decade, diverse VIO approaches have emerged, including classical geometric and optimization-based methods, hybrid learning-augmented systems, and fully learning-driven models. However, objective comparison remains difficult due to heterogeneous evaluation protocols, datasets, and metrics, complicating algorithm selection for practical deployment. This paper presents a unified, decision-oriented benchmark of representative VIO algorithms across multiple methodological paradigms. The methods are evaluated on EuRoC MAV, TUM-VI, KITTI Odometry, and the challenging multi-platform Fusion-PortableV2 dataset under a consistent evaluation protocol. Performance is assessed in terms of trajectory accuracy, drift behavior, scale consistency, and computational efficiency. Beyond accuracy, we analyze front-end and back-end computational costs to reveal trade-offs between robustness, precision, and resource consumption. Based on extensive cross-dataset experiments, we provide practical guidance for algorithm selection, identify common limitations and failure modes, and discuss empirically grounded open challenges and future research directions. This work serves as a practical reference for researchers and engineers in the VIO community.