Discrete Spherical Image Representation for CNN‐Based Spherical Camera Pose Estimation

Abstract

A spherical image has a full field of view. Using spherical images can estimate camera pose more accurately than conventional perspective images. It is known that a sphere can be quasi‐uniform sampled by dividing a geodesic dome repeatedly; it results in a representation with rotational invariance approximately. Based on this fact, in this paper we proposed a method of estimating a spherical camera pose based on the geodesic‐dome‐division‐based discrete spherical image representation by modifying the conventional perspective‐image‐oriented convolutional neural network, called PoseNet. First, a spherical image is represented as a geodesic‐dome‐division‐based discrete spherical image. Then, to cope with the six‐connectivity of the pixels of a geodesic‐dome‐division‐based discrete spherical image using the conventional convolutional neural network framework, the discrete spherical image is flattened as a connected five parallelograms and resampled as an eight‐connectivity image. Consequently, via the comparative experiments, it is found that the geodesic‐dome‐division‐based discrete spherical image representation approach can achieve a better result of camera pose estimation than equirectangular image and CubeMap representations. © 2021 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.