Direct Generation of Upright Panoramic Image From A Pair of Fisheye Images

Abstract

A 360 degrees panoramic image can be acquired by multiple cameras which cover the surrounding scenes as a camera cluster. The mostly popular and convenient approach is to use a pair of fisheye cameras which point opposite directions with a wider field of view more than hemisphere, respectively. To synthesize a 360 degrees panoramic image, traditional methods need to calibrate the intrinsic parameters of each fisheye camera and estimate the relative pose (extrinsic parameters) of the pair of fisheye cameras. To acquire upright panoramic image, the inclination correction is further needed. Until now, camera parameters and inclination correction are coped with differently. In this paper, we propose a novel deep learning-based method of generating a upright panoramic image directly from a pair of fisheye cameras. That is, we develop a neural network to generate a 360 degrees panoramic image from a pair of inclined fisheye images. First, we collected a comprehensive dataset specifically designed for this task. Then, we constructed a neural network to achieve this task, directly generating upright panoramic images from a pair of inclined fisheye images. The experimental results shown the effectiveness of the proposed method, also.