Evaluating Visual-Selective Visual-Inertial Odometry: An End-to-End Multi-Modal Pose Estimation Framework for Underwater Environments

This paper presents an evaluation of the performance of Visual-Selective Visual-Inertial Odometry (VS-VIO), a hybrid learning-based multi-modal pose estimation framework, in the challenging underwater domain. The assessment is based on Root Mean Square (RMSE) scores for translation and rotation vectors, compared to their reference values. The underwater environment, characterized by low lighting and high turbidity due to suspended particles, poses significant challenges for pose estimation. Understanding how hybrid learning-based multi-modal frameworks perform in such conditions is crucial for improving underwater navigation and exploration. In this study, we thoroughly analyze the performance of VS-VIO and its baseline model at the sub-sequence level, focusing on pose error. Additionally, we assess various technical aspects during the inference phase, including inference speed, power consumption, GPU utilization, GPU memory usage, and temperature. All evaluations are conducted using the AQUALOC dataset. Our findings reveal that the policy network within VS-VIO exhibits the ability to dynamically reduce the utilization of the visual modality while maintaining pose estimation accuracy. However, our analysis shows no statistically significant reduction in the percentage of visual modality usage when altering the penalty factor. These insights provide valuable guidelines for enhancing the performance of hybrid learning-based multimodal pose estimation frameworks in challenging underwater environments, contributing to advancements in underwater navigation and exploration technologies.