Framework And Evaluation For Spherical Superpixels

Heaps of superpixel algorithms have been developed and adopted as elementary tools in low-level computer vision and multimedia applications.Nowadays there are many kinds of superpixel algorithms such as watershed-based algorithms,density-based algorithms,graph-based algorithms and clustering-based algorithms,however,most of them are designed for planar images.In recent years,the quick growth of spherical panoramic images has raised the urgent need of spherical superpixel algorithms but until now only two spherical ones exist.In addition,a unifying benchmark of spherical image segmentation for the quantitative evaluation is urgently required due to the fact that there have been no high-quality spherical segmentaion benchmarks so far.Addressing the problem that only two spherical segmentation algorithms exist,in this paper,we present a general framework to establish spherical superpixel algorithms by extending planar counterparts.This framework mainly takes three factors: spatial distance measure,neighboring range and boundaries problem into consideration,under which two spherical superpixel algorithms are developed.In order to tackle the problem of lacking spherical segmentaion dataset,we propose the first segmentation benchmark of real-captured spherical images,which are manually annotated via a three-stage process.Besides,we use this benchmark to evaluate eight algorithms,including four spherical ones and the four corresponding planar ones,and discuss the results with respect to quantitative segmentation quality,runtime as well as visual quality.