Wang Sijie ; Kang Qiyu ; She Rui ; Tay Wee Peng ; Hartmannsgruber Andreas ; Navarro Diego Navarro

Abstract

Camera relocalization has various applications in autonomous driving.Previous camera pose regression models consider only ideal scenarios wherethere is little environmental perturbation. To deal with challenging drivingenvironments that may have changing seasons, weather, illumination, and thepresence of unstable objects, we propose RobustLoc, which derives itsrobustness against perturbations from neural differential equations. Our modeluses a convolutional neural network to extract feature maps from multi-viewimages, a robust neural differential equation diffusion block module to diffuseinformation interactively, and a branched pose decoder with multi-layertraining to estimate the vehicle poses. Experiments demonstrate that RobustLocsurpasses current state-of-the-art camera pose regression models and achievesrobust performance in various environments. Our code is released at:https://github.com/sijieaaa/RobustLoc