Abstract
A superpixel can be characterized as a vector in a color space or a covariance matrix on a manifold, by which two graph layers can be modeled on the common vertex sets. In this paper, we propose a novel approach for clustering such kind of multi-layer bipartite graphs. By Laplacian eigenmaps, each layer of the bipartite graph can be represented as a subspace of Rn so that the layers are regarded as points on Grassmann manifolds. With a properly defined distance metric, we fuse the graph layers into one and partition the final layer by spectral clustering. The experiments on the Berkeley Segmentation Datasets show that our new algorithm gives better or competitive segmentation compared with other bipartite graph related approaches.