| Abstract: |
This thesis introduces directional mean shift (DMS) as an extension to the standard mean shift. Mean shift is a well-known adaptive step-size mode seeking algorithm for kernel density function, which has been widely adapted to various vision and pattern analysis problems. The proposed DMS extends the mean shift for handling directional statistics toward analyzing directional data which occurs commonly in vision problems. Defining a metric for measuring directional distances, we derive the DMS algorithm as a convergent mode seeker for a kernel density function defined over circular domains. As our further contributions, we provide the proof of convergence and demonstrate adaptations of DMS to two different applications: 1) image segmentation in the HSV color space and 2) 3D medical structure topology classification. For the former, DMS is used to perform image segmentation with the circular hue component. For the latter, DMS is used to solve a clustering problem with a 2D image unfolded from a 3D spherical data. In both applications, our experiments demonstrate the effectiveness of DMS in contrast to the original mean shift defined in Euclidean domain.
|
