Image Enhancement using GANs for Monocular Visual Odometry

Abstract

Drones, mobile robots, and autonomous vehicles use Visual Odometry (VO) to move around complex environments. ORB-SLAM or deep learning-based approaches like DF-VO are two of the state-of-the-art technics for monocular VO. Those two technics perform correctly in outdoor scenarios but show some limitations in indoor environments. The extreme lighting conditions, non-Lambertian surfaces, or occlusion of indoor environments can disturb the visual information, and so the odometry information. Generative Adversarial Network (GAN) architectures recently proposed in the literature can help to overcome image low-light and blurring limitations. This research study aims to assess image enhancement's impact using GANS on the Visual Odometry algorithm DF-VO. Since DF-VO is also based on visual geometric information, the paper first considers the effect of two different GAN architectures in the camera's calibration. Then, the impact in the odometry information computed by DF-VO is evaluated. The preliminary results show that the reprojection error and the uncertainty of the calibration of a pin-hole-based camera do not increase significantly, and DF-VO's performance is improved.