Spectral Embedding of Feature Hypergraphs
In this paper we investigate how to establish a hypergraph model for characterizing object structures and how to embed this model into a low-dimensional pattern space. Each hyperedge of the hypergraph model is derived from a seed feature point of the object and embodies those neighbouring feature points that satisfy a similarity constraint. We show how to construct the Laplacian matrix of the hypergraph. We adopt the spectral method to construct pattern vectors from the hypergraph Laplacian. We apply principal component analysis (PCA) to the pattern vectors to embed them into a low-dimensional space. Experimental results show that the proposed scheme yields good clusters of distinct objects viewed from different directions.