Rejecting Outliers Based on Correspondence Manifold
基于对应流形的错误匹配剔除研究(英文)

Finding reliable corresponding points between two images is a fundamental problem in computer vision, especially with the development of L∞ vision framework. This paper introduce the correspondence manifold and propose a novel scheme to reject outliers by learning upward views of the manifold. The proposed scheme is independent of the parametric model to be estimated and overcomes the following limitations of the available methods in the published works: efficiency sharply goes down with the increase of outlier percentage and the number of the estimated model parameters; outlier rejecting is coupled with model selection and model estimation. Experiments on real image pairs show the excellent performance of our proposed scheme.