Estimating Pose of Omnidirectional Camera by Convolutional Neural Network

Abstract

Estimating camera pose is a basic task for computer/robot vision. In this paper we investigated the estimation of the pose of an omnidirectional camera by applying the PoseNet [1] to omnidirectional images. An omnidirectional image is represented by an equirectangular image, perspective mosaic image representation generated from a cubemap, and multi-channel perspective image representation. The comparative experiments are carried out for these three representation of omnidirectional images and the effect of the field of view of an omnidirectional image on the accuracy of camera pose estimation. The improvement of the pose accuracy with the increase of the field of view is confirmed.