Semi-Supervised Discriminative Mutual Subspace Method

Xianhua Zeng, Jingjing Zhong

Abstract: Subspace recognition has recently attracted more attention for vector set or image set matching in machine learning and computer vision. In this paper, we firstly give a more simple proof of Procrustes metric (Theorem 2) than literature [1,7]. Then, a novel Semi-Supervised Discriminative Mutual Subspace Method (SS-DMSM) is proposed based on Procrustes metric. For finding a better discriminative subspace, our SS-DMSM algorithm sufficiently considers the intrinsic geometric information on Grassmann manifold that is the set of all subspaces, and effectively uses the label information of those training subspaces. Experimental results on Cambridge gesture database and ETH-80 database show that our SS-DMSM algorithm outperforms the classical MSM and CMSM algorithms.