Generalized Shortest Path-based Superpixels for Accurate Segmentation of Spherical Images

Most of existing superpixel methods are designed to segment standard planar\nimages as pre-processing for computer vision pipelines. Nevertheless, the\nincreasing number of applications based on wide angle capture devices, mainly\ngenerating 360{\\deg} spherical images, have enforced the need for dedicated\nsuperpixel approaches. In this paper, we introduce a new superpixel method for\nspherical images called SphSPS (for Spherical Shortest Path-based Superpixels).\nOur approach respects the spherical geometry and generalizes the notion of\nshortest path between a pixel and a superpixel center on the 3D spherical\nacquisition space. We show that the feature information on such path can be\nefficiently integrated into our clustering framework and jointly improves the\nrespect of object contours and the shape regularity. To relevantly evaluate\nthis last aspect in the spherical space, we also generalize a planar global\nregularity metric. Finally, the proposed SphSPS method obtains significantly\nbetter performance than both planar and recent spherical superpixel approaches\non the reference 360{\\deg} spherical panorama segmentation dataset.\n