Object Detection of Equirectangular Images by Swapping the Equator and Poles
Zihao Zhao, Shigang Li · 2024
An equirectangular image contains a large distortion as an object approaching the poles. Can we use a neural network which is pretrained based on perspective images to detect objects in an equirectangular image? In this paper we propose an object detection method of equirectangular images by swapping the equator and poles. First, for an input equirectangular image, we map it onto a sphere. After rotating the spherical image by 90 degrees around a horizontal axis, we map the rotated spherical image back onto an equirectangular image. Then, the original input equirectangular image and the swapped one are processed by a neural network pretrained by perspective images. The effectiveness of the proposed method is shown by the preliminary experimental results.