Presented at IROS 2022

The study on “Unsupervised Simultaneous Learning for Camera Relocalization and Depth Estimation," by Shun Taguchi and Noriaki Hirose, was selected for the presentation at The 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).

Monocular camera relocalization is a technology that aims to estimate the position and posture from a single camera image and important for determining geopositions when GPS cannot be used. So far, methods that utilize deep learning to perform supervised learning on camera images and location/posture pairs have been proposed for camera relocalization. In this presentation, we propose a method for unsupervised learning from videos by simultaneously studying camera relocalization and depth estimation. Learning is performed by mutually minimizing the estimated error in images across multiple images based on depth. The proposed method achieves greater precision in relocalization than the leading Visual SLAM, and depth estimation than a representative method. The derived results are expected to contribute to the realization of indoor autonomous robots and services using positioning information.

Title: Unsupervised Simultaneous Learning for Camera Re-localization and Depth Estimation from Video
Authors: Taguchi, S., Hirose, N.
Appears in: IEEE/RSJ International Conference on Intelligent Robots and Systems
Presented: October 25, 2022
https://doi.org/10.1109/IROS47612.2022.9982213