Novel semisupervised high-dimensional correspondences learning method

Abstract

Correspondence is one of the big challenges in machine learning and image processing. To match two high-dimensional data sets with a certain number of aligned training examples, a novel semisupervised method is proposed. It is mainly based on two manifold learning approaches: maximum variance unfolding (MVU) and locally linear embedding (LLE). We have modified MVU to a semi-supervised version to solve the correspondence problem. Additionally, the nonuniform warps and folds caused by employing LLE alone and the computational burden of MVU disappear when they are combined. The proposed algorithm outperforms traditional methods in accuracy and efficiency. Three examples are performed to demonstrate the potential of this method.