A Case Study on Visual-Inertial Odometry using Supervised, Semi-Supervised and Unsupervised Learning Methods

This paper presents a pilot study comparing three different learning-based visual-inertial odometry (VIO) approaches: supervised, semi-supervised, and unsupervised. Localization and navigation have been the ancient bur important topic in both research area and industry. Many well-developed algorithms have been established regarding this research task using a single sensor or multiple sensors. VIO, that uses images and inertial measurements to estimate the motion, is considered as one of the key technologies to virtual reality and argument reality. With the rapid development of artificial intelligence technology, people have started to explore new methods for VIO instead of traditional feature-based methods. The advantages of using learning-based method can be found in eliminating the calibration and enhance the robustness and accuracy. However, most of the popular learning-based VIO systems require ground truth during training. The lack of training dataset limits the power of neural networks. In this study, we proposed both semi-supervised and unsupervised methods and compared the performances between the supervised model and them. The neural networks were trained and tested on two well-known datasets: KITTI Dataset and EuRoC MAV Dataset.