Abstract:
Kernel density estimation and mode finding techniques play an active role in solving contemporary computer vision problems, like edge preserving smoothing, segmentation, registration, motion estimation and tracking. The mean shift algorithm is a popular approach to locate density modes. Recently we proposed the multiscale mode filter, a generalization of the mean shift filter, which is able to avoid spurious modes while minimizing outlier sensitivity. In this paper we evaluate the effectiveness of the multiscale mode filter in edge preserving smoothing and image segmentation.
