Spherical Superpixels: Benchmark and Evaluation.

Liang Wan,Xiaorui Xu,Qiang Zhao,Wei Feng
DOI: https://doi.org/10.1007/978-3-030-20876-9_44
2019-01-01
Abstract:Although a variety of superpixel algorithms have been developed and adopted as elementary tools in low-level computer vision and multimedia applications, most of them are designed for planar images. The quick growth of spherical panoramic images raises the urgent need of spherical superpixel algorithms and also a unifying benchmark of spherical image segmentation for the quantitative evaluation. In this paper, we present a general framework to establish spherical superpixel algorithms by extending planar counterparts, under which two spherical superpixel algorithms are developed. Furthermore, we propose the first segmentation benchmark of real-captured spherical images, which are manually annotated via a three-stage process. 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.
What problem does this paper attempt to address?