Superpixels Generation of RGB-D Images Based on Geodesic Distance

Abstract

A novel algorithm for generating superpixels of RGB-D images is presented in this paper. A regular triangular mesh is constructed by the depth and a local geometric features sensitive initialization method is proposed for initializing seeds by a density function. Over-segmentation of the vertices on mesh can be generated by minimizing a new energy function defined by weighted geodesic distance which can be used for measuring the similarity of vertices with color information. At last, superpixels are generated by re-mapping the mesh over-segmentation to 2D image. During energy optimizing, we will check the topology correctness of the superpixels and refine the topology of the superpixels. Experiments on a large RGB-D images database show that the superpixels generated by the new method can adhere to the object boundaries well and outperform the state-of-the-art methods.