Robust Subspace Segmentation via Sparse Relation Representation

Spectral clustering based algorithms are powerful tools for solving subspace segmentation problems. The existing spectral clustering based subspace segmentation algorithms use original data matrices to produce the affinity graphs. In real applications, data samples are usually corrupted by different kinds of noise, hence the obtained affinity graphs may not reveal the intrinsic subspace structures of data sets. In this paper, we present the conception of relation representation, which means a point’s neighborhood relation could be represented by the rest points’ neighborhood relations. Based on this conception, we propose a kind of sparse relation representation (SRR) for subspace segmentation. The experimental results obtained on several benchmark databases show that SRR outperforms some existing related methods.