Semantic and Geometric Priors for Monocular SLAM
MCML Authors
Abstract
Abstract
This thesis explores the use of different types of prior information for monocular SLAM systems. We propose a probabilistic map point model that includes predicted semantic information about object classes as a prior into the estimation of the depth and inlier ratio of each map point. We also show that using geometric priors in the form of predicted normal maps can provide valuable information to improve the rotation estimation in man-made environments with regular but low-textured geometries.
BibTeXKey: Bra25