关键词: WMD, 谱分割, 聚类, Ncut, NystrÖ, m估计

Abstract: To obtain a better segmentation result, this paper used Weighted Mahalanobis Distance (WMD) Gaussian kernel for Nystrm-Ncut segmentation. It proves that weighted Mahalanobis distance Gaussian kernel is more appropriate for spectral graph theoretic methods than Mahalanobis distance, because weighted Mahalanobis distance can compute the similarity between two pixels more accurately.

Key words: weighted mahalanobis distance, spectral graph partition, clustering, normalized cuts, nystrÖ, m approximation