Self‐Supervised Panoramic Image Generation of Spherical Cameras

Spherical cameras with 360° field of view are widely used in virtual reality and robotics. A typical spherical camera consists of two fisheye cameras facing opposite directions, each covering more than a hemisphere. In such a setup, the distance between the viewpoints of the two fisheye cameras is very small. Consequently, the parameters necessary for synthesizing panoramic views from the two fisheye cameras can be reduced to the intrinsic parameters of each fisheye camera and the relative rotation between them. In this paper, we propose a self‐supervised method for panoramic image generation using such dual‐fisheye 360° cameras. Our network employs a shared encoder with dedicated output heads to estimate intrinsic parameters (focal length and distortion coefficients) and a pose network to estimate the relative rotation between the two fisheye cameras. A differentiable projection module then synthesizes a panoramic image from the fisheye pair. This end‐to‐end framework requires no ground truth or manual annotations, leveraging the overlapping regions between fisheye images on the sphere. Experimental results demonstrate that our method outperforms existing state‐of‐the‐art approaches. © 2026 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.