2018/11/20 by Yeonkun Lee, Lee, Yeonkun, Jaeseok Jeong +7 · 4 citations
Computer Science · Environmental Science · #Advanced Vision and Imaging #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Remote Sensing and LiDAR Applications #Video Surveillance and Tracking Methods
paper · pdf · doi:10.48550/arxiv.1811.08196
openalex publication_date 2018/11/20 · openalex created_date 2022/08/01 · openalex updated_date 2026/07/28
Omni-directional cameras have many advantages overconventional cameras in\nthat they have a much wider field-of-view (FOV). Accordingly, several\napproaches have beenproposed recently to apply convolutional neural\nnetworks(CNNs) to omni-directional images for various visual tasks.However,\nmost of them use image representations defined inthe Euclidean space after\ntransforming the omni-directionalviews originally formed in the non-Euclidean\nspace. Thistransformation leads to shape distortion due to nonuniformspatial\nresolving power and the loss of continuity. Theseeffects make existing\nconvolution kernels experience diffi-culties in extracting meaningful\ninformation.This paper presents a novel method to resolve such prob-lems of\napplying CNNs to omni-directional images. Theproposed method utilizes a\nspherical polyhedron to rep-resent omni-directional views. This method\nminimizes thevariance of the spatial resolving power on the sphere sur-face,\nand includes new convolution and pooling methodsfor the proposed\nrepresentation. The proposed method canalso be adopted by any existing\nCNN-based methods. Thefeasibility of the proposed method is demonstrated\nthroughclassification, detection, and semantic segmentation taskswith synthetic\nand real datasets.\n