Edited by author 10-11-2001 02:54 AM
I think how to decide symilarity matrix S is a key problem. And choosing S including two aspects, firstly, how to choose the model of S? Secondly, if some good model for S has been chosen, as said in "A Random Walks View of Spectral Segmentation", choose some positive symmetrical matrix, how to decide the parameter of the model? In "learning segmentation by Random Walks", the authors have shown a training method to decide S's parameters for a supervised case, but how about an unsupervised case? Thinking intuitively, S should have great influence on the following steps, so I want to know how to deal with it in real application.

To Hsin-Hao:
I can't agree with your idea about there is no difference between segmentation and clustering. Clutering is just one technique to deal with segmentation problem. In some segmentation technique, for example, just use thresholding and masking, edge/boundary detection and region growing, it has nothing to do with clustering.